46 results about "Parallel generation" patented technology

The invention discloses a water chilling unit fault diagnosis method and system for unbalanced data, and belongs to the field of water chilling unit fault diagnosis, and the method comprises the steps: restoring balance fault data with original data characteristics through employing original data based on a disturbance reconstruction mechanism; performing extraction and modeling on the time sequence characteristics and the space structure of the balance fault data by using a space-time automatic encoder, and reconstructing original data; respectively extracting time sequence features and topological features by utilizing a dynamic topological time sequence feature integration network, and performing adaptive fusion by adopting a dynamic gating fusion mechanism; and classifying the fused features to obtain a fault diagnosis result. According to the method, parallel generation of multiple types of fault samples is achieved, it is guaranteed that each type of generated fault data has similar time-space characteristics with original fault data, the method can adapt to the real-time changing working condition of the water chilling unit, collaborative modeling of time sequence characteristics and dynamic topological characteristics in a complex multi-sensor system of the water chilling unit is achieved, and the modeling efficiency is improved. And high-precision real-time diagnosis can also be realized on edge equipment.

The invention discloses a transformer over-capacitance threshold value dynamic adjustment method based on real-time monitoring. The method comprises the following steps: S1, multi-dimensional total element real-time acquisition; s2, data layering preprocessing and directional transmission; s3, preliminarily judging an edge end basic threshold value and a cluster; s4, calculating insulation, environment and operation three-dimensional coupling threshold values; s5, optimizing a small sample causal interpretable AI threshold value; s6, new energy consumption and cluster cooperation threshold correction; according to the multi-dimensional strategy parallel generation method, by constructing an insulation, environment and operation three-dimensional coupling threshold calculation system, the suitability of super-capacity protection and the actual operation state of the transformer is improved; an AI threshold optimization framework can be explained by means of small sample causality, and transparency and reliability of super-capacity threshold adjustment logic are guaranteed; through the new energy consumption and cluster collaborative threshold correction and multi-strategy fusion optimization mechanism, the new energy consumption efficiency is guaranteed, and unnecessary protection actions are reduced through cluster load linkage.

The invention relates to the technical field of pseudo-random binary sequence design, in particular to a parallel PRBS generation method and system for testing a multi-protocol high-speed serial transceiver, and the method comprises the steps: setting demand parameters of a pseudo-random binary sequence for testing the multi-protocol high-speed serial transceiver, and enabling the demand parameters to comprise an order, a bit number and a feedback polynomial; constructing a generator matrix for describing state transition of the linear feedback shift register according to the feedback polynomial; according to the bit number of the pseudo-random binary sequence, generating a mapping matrix of the pseudo-random binary sequence through the exponentiation of the generated matrix; based on the mapping matrix and the current beat linear feedback shift register state, outputting a pseudo-random binary sequence meeting the requirement parameters for testing the multi-protocol high-speed serial transceiver. According to the method, parallel generation of any PRBS can be supported, the PRBS parallel implementation logic configuration is simple, the hardware implementation complexity is low, the PRBS output throughput rate is high, and the design requirements of multi-protocol high-speed SerDes can be met.

The invention discloses a hydrological electronic achievement archive batch output method and system and a medium, and the method comprises the following steps: data extraction: reading annual calibration data of specified equipment from a database, including calibration time, a standard value, a measured value and an error rate field; template rendering: reading an Excel template by using a POI (Point of Interest), and mapping data to a template placeholder; and PDF generation: creating a Document object by the iTextPDF, setting a page margin, inserting a page header, dynamically rendering a table, and outputting the PDF through HttpServletResponse. And in combination with intelligent excel-to-pdf of the StirlingPDF, a three-level processing normal form of'type shunting-parallel generation-dynamic assembly 'of the equipment detection document is realized in a breakthrough manner.

The invention discloses a dual reasoning acceleration method based on draft prediction and compressed memory, which comprises the following steps of: 1) acquiring an input context of model reasoning and a token sequence generated by a model at each stage of reasoning; 2) according to an input context of model reasoning and a token sequence generated by the model, predicting subsequent candidate tokens in parallel as draft tokens; the method comprises the following steps of (1) performing parallel generation, (2) performing main model verification, (3) organizing draft tokens according to a tree structure, (4) performing main model verification, (5) compressing each reasoning block into compressed tokens with a fixed length s, and (6) performing context reconstruction. According to the method, through the combination of parallel generation and context compression, the problem of serial low efficiency of traditional autoregression is solved, and the redundant overhead of long-sequence reasoning is relieved.

The application relates to an acceleration method of an AES key expansion algorithm and an instruction set processor, wherein the acceleration method is based on a VAESKEYE extension instruction set, a plurality of independent AES key expansion algorithms are accelerated in parallel by pipelining execution of instructions in the VAESKEYE extension instruction set; the VAESKEYE extension instruction set adopts a RISC architecture and is used for generating any one of AES key expansion words in the AES key expansion algorithm according to a source operand; the VAESKEYE extension instruction set adopts a plurality of AES key expansion word parallel generation algorithms, the plurality of AES key expansion word parallel generation algorithms refer to algorithms for taking a key word or a key expansion word and an AES key expansion round number in 8 independent AES key expansion algorithms as inputs and respectively completing generation of 8 independent AES key expansion words in parallel. The application greatly improves the speed of executing the AES key expansion algorithm.

The application discloses a kind of parallel random number generation devices based on pulse driving micro-ring resonator cavity, by adopting pulse mode driving micro-ring resonator cavity, improve the instantaneous power level in cavity, enhance nonlinear effect, to effectively expand the effective bandwidth of chaotic entropy source, improve the complexity of entropy source.Meanwhile, using the characteristics that micro-ring resonator cavity naturally supports multi-wavelength output, and combining with filter demultiplexing structure, the parallel output of multiple chaotic entropy sources is realized, which is beneficial to improve the parallel generation ability of random number.Finally, by combining post-processing methods such as difference, quantization and decorrelation, the overall generation rate of random bits is improved while ensuring randomness quality.The application has relatively simple structure, is suitable for integrated implementation, and can meet the application requirements of high-speed parallel random number generation.

Embodiments of the present disclosure relate to a method, an apparatus, an electronic device and a medium for event extraction. The method comprises: extracting a plurality of named entities from a document as a plurality of event arguments. The method further includes determining an event type and a template corresponding to the event type in the document. The method also includes filling the plurality of event arguments in respective locations in the template to generate a plurality of candidate event records, and filtering the plurality of candidate event records to obtain one or more target event record. In this way, respective candidate event records are generated through iteration while event extraction is performed at a document level, which can avoid performance fluctuations caused by manual selection of an event role generation sequence, and can also avoid under-fitting brought about by parallel generation to thus improve the event extraction accuracy.

The invention provides an aircraft manufacturing process instruction intelligent generation method based on hybrid retrieval enhanced generation, and the method guarantees that a generated instruction has professional sentence patterns with historical experience and compliance indexes with standard specifications through a hybrid retrieval mechanism. A large model is limited in a specific process generation link through the agent workflow, strict placeholder filling logic is matched, the illusion problem that key parameters are pinched by the model is thoroughly eradicated, meanwhile, the parallel generation architecture enables the generation time of a long-flow process instruction to be greatly shortened, and unification of high accuracy, high controllability and high efficiency is achieved.

This invention provides a method, apparatus, electronic device, and storage medium for generating multiple candidate responses, belonging to the field of artificial intelligence technology. The method includes: inputting a target prompt into a large language model; generating multiple candidate responses for the target prompt through multiple parallel generation threads; during the generation of multiple candidate responses word-by-word, obtaining hidden state features corresponding to each parallel generation thread using the network layer of the large language model; performing batch-dimensional attention calculations on each hidden state feature using a cross-thread shared layer to obtain shared state features; and using the output layer of the large language model to predict the output words of each parallel generation thread based on the shared state features. This invention breaks the independence between parallel generation threads by introducing a cross-thread shared layer between the network layer and the output layer, enabling interaction and fusion of feature dimensions among the threads during the generation process. This effectively utilizes the global information of the parallel group, improving the diversity and quality of generated candidate responses.

The invention discloses an auto-regression and mask diffusion double-normal-form fusion language model hardware acceleration system based on an FPGA (Field Programmable Gate Array), and relates to the field of FPGA and machine learning. The invention provides a model core calculation process-oriented hardware acceleration architecture by aiming at a language model fusing double normal forms of an autoregression model and a mask diffusion model and adopting the characteristics of revising an attention mechanism and parallel generation of a KV cache. The method disclosed by the invention is implemented by taking an Eso-LMs (Esolic Language Models) model as an example. The system comprises a layer normalization and adaptive layer normalization modulation module, a QKV projection module, a rotation position coding application module, a KV cache management module, a multi-head attention calculation module, an output projection and residual connection module and a multi-layer perceptron module. By optimizing the calculation sequence and the data flow of each module in the model calculation process and adopting the pipeline parallel and resource reuse technology, the efficient reasoning acceleration of the language model fusing the autoregression and mask diffusion double normal forms on the FPGA platform is realized, and the model reasoning speed is obviously improved. According to the system, the advantage of FPGA customizable hardware acceleration is fully exerted, the utilization efficiency of hardware resources is improved, data pipeline blockage is eliminated, calculation delay and storage overhead are reduced by reconstructing the data flow direction and constructing a whole-process pipeline processing architecture, and the system is suitable for the deployment requirement of an edge calculation scene.

The embodiment of the invention relates to the technical field of artificial intelligence, and provides a parallel generation method of response texts and electronic equipment. The method comprises the following steps: receiving a group of tokens, and obtaining a representation vector corresponding to each token based on at least one historical text block and the token; denoising the representation vectors corresponding to the group of tokens in parallel, and obtaining a prediction text block corresponding to the group of tokens based on the denoised representation vectors; in response to the fact that the reasoning task is not completed, taking the prediction text block as a new historical text block, generating a group of new tokens based on at least one historical text block, and returning to the step of obtaining the representation vector corresponding to each token based on the at least one historical text block and the token by using the coding module; and in response to completion of the reasoning task, obtaining a response text corresponding to the reasoning task based on the at least one historical text block and the predicted text block. According to the method provided by the invention, the token of the response text can be generated in a parallel mode, and the problems of long reasoning delay and low response speed in the prior art are solved.

This application proposes an automatic method and apparatus for generating distillation data in underground coal mines. The method includes: acquiring unlabeled data and combining it with instruction data to form a distillation dataset, and distributing it to multiple parallel generation channels; forming a structured candidate inference result set; fusing and calibrating the candidate inference result set to generate labeled distillation data, performing multi-dimensional qualification checks, and selecting valid labeled results; constructing an unbiased distillation training dataset, and inputting it into the target model for supervised fine-tuning training to obtain a distillation model adapted to underground coal mine scenarios. By automatically generating highly diverse distillation data that conforms to the characteristics of underground scenarios, the problem of high labeling costs and uneven data quality is effectively solved; the model's distillation performance is improved through preference consistency constraints and multi-dimensional quality checks, significantly enhancing the generalization ability and recognition accuracy of the distillation model in complex underground environments.

The invention provides an underground mine live-action three-dimensional modeling system and method based on unmanned aerial vehicle SLAM and multi-sensor fusion. The underground mine live-action three-dimensional modeling system comprises a hardware platform and a software platform. The hardware platform takes an unmanned aerial vehicle as a main body, and integrates a visual sensor, a laser radar, an IMU, a self-adaptive lighting system, a visual laser acquisition control point device, a roof control point identification camera and a wireless communication module; the wireless communication module is used for transmitting the acquired data back to a software system in real time and transmitting remote control instructions among the components; the software platform comprises a front-end SLAM module and a modeling and rendering module; the front-end SLAM module adopts a vision-laser radar-IMU tight coupling SLAM algorithm and fuses control point data acquired by a top plate control point identification camera to realize real-time point cloud registration and map construction of absolute coordinate constraint in a GPS-free environment; the core of the modeling and rendering module is a Mesh and 3DGS double-model parallel generation technology, and the modeling and rendering module has the parallel operation capability of synchronously completing Mesh model generation and 3DGS model training.

The invention discloses a content generation data processing method and system based on multi-agent collaboration, and the method comprises the steps: carrying out the semantic deconstruction of a multi-modal content demand inputted by a user through a task analysis agent, and generating a structured task description vector containing a theme dimension, a style element and an emotional tendency; generating a multi-modal content draft based on the structured task description vector; inputting the multi-modal content draft into a quality evaluation agent network, and generating a quality evaluation matrix containing semantic coherence scores and style matching degrees; and dynamically adjusting a content generation strategy through a decision coordination agent according to the quality evaluation matrix, optimizing an agent cooperation weight by adopting a reinforcement learning mechanism, and outputting a final optimized multi-modal content work. By utilizing the embodiment of the invention, high-quality and high-consistency parallel generation and deep fusion of multi-modal contents such as texts, images and audios can be realized, and the creativity and semantic coherence of final works and the matching degree with user requirements are improved.

The invention discloses an electric vehicle self-charging system and a self-charging method thereof. The system comprises a first motor, a second motor, a first generator, a second generator, a vehicle-mounted battery and a control circuit. In the running process of the vehicle, the system converts kinetic energy of the vehicle into electric energy and stores the electric energy in a vehicle-mounted battery through an energy recovery mode and a parallel power generation mode; when the vehicle is in a static state, through a static power generation mode or a low-power operation mode, self-charging is achieved through a low-power-consumption motor and a high-efficiency generator. The control circuit dynamically adjusts the working mode of each component according to the running state and working condition of the vehicle, and the energy recovery and charging efficiency is optimized.

The present invention discloses a parallel boundary layer mesh generation method and device for computational fluid dynamics. The method includes: first, generating an initial background mesh with boundary layer elements according to the original CAD geometric model; then, obtaining the physical surface boundary of the background mesh and the mesh elements connected thereto based on the initial background mesh and its topological connection relationship to construct a fixed closed region; further, performing regional decomposition on the initial background mesh using a graph partitioning tool; secondly, densifying the surface mesh of the sub-region; then, performing parallel calculations on the normal vectors of the surface mesh points and the layer advancement distance according to the boundary layer elements, and generating triangular prism elements by applying the layer advancement method; finally, performing quality inspection and intersection judgment on all boundary layer elements to obtain a complete boundary layer mesh. The present invention can make up for the regret of the lack of a parallel generation method for boundary layer meshes in large-scale mesh generation, realize the rapid parallel generation of large-scale boundary layer meshes, and overcome the performance bottleneck in terms of generation efficiency.

The invention provides a multi-machine surveying and mapping management method and device, electronic equipment and a storage medium, and relates to the technical field of surveying and mapping. The method comprises the following steps: receiving a task instruction comprising a target ground resolution, a minimum overlapping degree, a task priority and a landform situation; according to the target ground resolution and a preset algorithm, calculating to obtain a navigation height, generating a plurality of navigation routes in combination with the landform situation, the lowest overlapping degree and the preset algorithm, and then generating control instructions corresponding to the unmanned aerial vehicles in combination with task priorities to control the unmanned aerial vehicles to perform surveying and mapping; and during surveying and mapping, surveying and mapping data returned by the unmanned aerial vehicles are spliced in real time. According to the method, the navigation height of the unmanned aerial vehicle is determined through the target ground resolution, and the navigation route is correspondingly generated according to the landform condition by combining the lowest overlapping degree, so that parallel generation of the multi-vehicle navigation route is realized, and surveying and mapping data returned by each unmanned aerial vehicle are spliced in real time; therefore, the reasonability of route planning and task allocation during unmanned aerial vehicle surveying and mapping is improved, and the surveying and mapping accuracy and efficiency are improved.

The present invention relates to a parallel syndrome generation device for a MPEG-2 synchronization byte decoder based on a lookup table, applicable to the digital multi-program system B recommended by ITU-T J.83. The device is characterized in that it mainly consists of a 1504-bit memory, 15 registers R0 to R 14 , 128 two-input AND gates and 128 two-input XOR gates. The syndrome generation device provided by the present invention inputs the received data of 188 bytes in a 16-bit parallel manner and outputs 1-byte syndrome in an 8-bit parallel manner, which can effectively reduce the logical resource requirements while increasing the processing speed by 16 times, and has the advantages of fast running speed and less consumption of logical resources.

The application discloses a text candidate screening method and device. The method comprises the following steps: obtaining initial candidate texts generated in parallel by a language model, and calculating the quality scores of the initial candidate texts; performing an iterative screening step to determine final candidate texts, which firstly takes the text with the highest quality score as the first final candidate text, and then in the repeatedly performed selection operation, for each candidate text, calculates the difference score between the candidate text and the selected final candidate set, and calculates a comprehensive score based on the quality score and the difference score, and determines the candidate text with the highest comprehensive score as the new final candidate text. The application guarantees efficiency through parallel generation, and solves the problem that it is difficult to balance the quality and diversity of generated texts in the prior art through the iterative screening strategy combining the quality score and the dynamic difference score, so that the diversity of the final candidate set can be significantly improved while ensuring high quality.

This invention discloses a cloud-based dialogue response method, device, and medium for AI toys, relating to the field of natural language processing technology. The method includes: performing structured task translation processing on a semantic constraint guidance graph and writing corresponding fields to generate a task-driven script; using a constraint projection hierarchical construction method to perform hierarchical response channel orchestration and parallel generation processing on the task-driven script and the semantic constraint guidance graph to obtain hierarchical candidate response clusters; performing convergence and optimization compensation on the hierarchical candidate response clusters to obtain the target response content, and combining the target response content with the task-driven script to encapsulate an executable response, generating a cloud-based response result package; and performing terminal response delivery and memory storage write-back based on the cloud-based response result package to generate response execution information. This invention, by performing semantic deconstruction and relational constraint construction on the natural language requests of AI toys, helps improve the accuracy, coherence, and terminal execution adaptability of the final response content.

The application discloses a kind of application log fragment parallel generation and multi-level materialized view aggregation processing method, predefine application log general framework using standard log format and carry out encapsulation, at least including application identification, log type and log detailed content;In application, the log generator instance of encapsulated application log general framework is introduced, and it is matched to log recording demand, and the API provided by application log general framework is called to generate log record at the log recording point of the program code of application, and application log general framework establishes mapping relationship with application by application identification;When debugging or running application, application log is generated, and saved as original log file;In response to log aggregation instruction, according to preset aggregation strategy, multi-level materialized view aggregation processing is carried out to original log file, and saved as aggregation log file.The application can solve the technical problems that application log format is not standard, content is difficult to analyze, and mass log is difficult to compress.

The invention discloses a high-order grid parallel generation method based on GPU acceleration and fine projection, and the method comprises the steps: firstly inputting an original volume grid and a surface grid, and carrying out the order lifting processing, and obtaining a high-order volume grid and a high-order surface grid; secondly, calculating physical offsets of all nodes on the high-order surface grid relative to the high-order surface grid through projection by utilizing an AABB tree structure, and storing the offsets as displacement potential function values; then, on the basis of a boundary integral equation, physical offset vectors of all internal nodes in the volume mesh are calculated in parallel through a GPU; and finally, traversing all grid nodes, and updating all point coordinates according to the physical offset vectors of the internal nodes to obtain a generated high-order grid. According to the method, the calculation time of the boundary offset is greatly shortened, and the method has the advantages of high efficiency, high expandability and high precision and is suitable for high-precision CFD simulation pretreatment of industrial-grade complex shapes.

Techniques for parallel generation of blocks using a compiler are disclosed. A processing unit comprising compute slices, barrier register sets, a control unit, and a memory system is accessed. Each compute slice includes an execution unit and is coupled to other compute slices by a barrier register set. A compiler evaluates a compiled program that includes basic blocks, based on a control flow graph. A first hyperblock is created from at least two basic blocks. One or more branch instructions are replaced with skip instructions that direct instruction execution between basic blocks in the hyperblock. A first slice task is allocated to a first compute slice. A second slice task is allotted based on branch prediction. Pointers to the first compute slice and the second compute slice are initialized. The compiled program is executed, beginning with the first compute slice.

The invention relates to a large model autoregressive Transform parallelization generation method, which comprises the following steps of: dividing an input long text sequence into a plurality of local context windows with fixed lengths, and ensuring that an overlapping part exists between adjacent windows by adopting a sliding mechanism so as to maintain semantic coherence; parallel encoder construction and decoding: after a plurality of sliding windows are generated by adopting a sliding mechanism, performing parallel encoding processing on each window by utilizing a multi-GPU parallel computing mechanism; part of generated results and attention key value pairs are stored in an autoregressive decoding maintenance historical state cache region, and repeated calculation is avoided; a cross-window attention mechanism is introduced in the decoding stage, so that a current decoded word can pay attention to context information in a plurality of sliding windows; task scheduling and load balancing control: distributing window processing tasks to different computing units through a dynamic task scheduler, and optimizing the resource utilization rate in combination with a load balancing algorithm; and parallel generation and result splicing output: splicing and integrating the results generated by the windows in sequence.

The invention discloses a long-time-history large model processing method based on asynchronous reinforcement learning, and relates to the technical field of large language models.The method comprises the steps that the problem complexity is improved through dual-operation iteration, high-quality training data is obtained through triple verification, a training end and a track generation end are decoupled through a two-stage track parallel generation mechanism, and the problem is solved; after a training end collects a preset batch track, long-time-history large model updating is started, then historical data is managed by depending on a basic tool set sub-model, end-to-end asynchronous reinforcement learning optimization is executed on the whole process of an intelligent agent, finally, the problem of long-time-history reward sparseness is solved based on a GRPO algorithm, and a model strategy is optimized. According to the scheme, the core bottlenecks of weak long-range exploration, low training efficiency and insufficient data quality of an open-source search intelligent body are solved, the search intelligence of fuzzy query analysis, key information extraction by noise, cross-document reasoning verification and super-long-range tool calling is unlocked, the limitation of a traditional search round is broken through, and a long-time-range task is supported.

The present invention is applicable to the field of data processing, and discloses a method for parallel generation of response texts, a terminal device, and a storage medium. The method for parallel generation of response texts includes: when input information is detected, generating a global vector corresponding to the dialogue context and generating an initial response sequence corresponding to the global vector; inputting the global vector and the initial response sequence into the A-G module for parallel processing to obtain multiple candidate sequences; and generating a response content according to the candidate sequences. Through the parallel processing mechanism of the A-G module, multiple candidate sequences are generated in a single iteration, shortening the time delay required by the traditional word-by-word generation method. The problem of delay caused by multiple iterations in the traditional sequence generation method is solved. In the stage of generating the global vector, through the cooperation of positional embedding and multi-head self-attention, the temporal relationship and deep semantics of the dialogue context can be captured. Combining with the overall evaluation mechanism of the candidate sequences of the A-G module, the phenomenon of semantic discontinuity caused by local optimal solutions in the traditional method is effectively avoided.

The invention relates to a seismic wave simulation method based on terrain point cloud data hexahedral mesh parallel generation. The method comprises the following steps: acquiring terrain point cloud data, and setting parallel hexahedral mesh generation parameters; in each process, a terrain quadrilateral surface grid is generated based on the point cloud, a non-terrain surface grid is obtained through 2D over-limit interpolation, then the longitude upper and lower scale range of 3D over-limit interpolation is calculated, and a geologic body hexahedral grid is generated through 3D over-limit interpolation. And determining the number of grids in the longitude direction of each process according to the superscript and superscript ranges, calculating the number of grids in the latitude direction and the number of grids in the elevation direction, further obtaining each subdivision grid and the number of subdivision grids generated by the process, and finally generating a geologic body hexahedron subdivision grid. And finally, importing into seismic wave numerical solution software, setting parameters, carrying out parallel solution, and analyzing seismic wave propagation characteristics according to a result. By adopting the method, the seismic wave simulation precision can be improved.

The invention provides an RS and 64B / 66B coding combined circuit and a parallelism degree configuration method, and relates to the technical field of parallelism degree configuration of coding combined circuits, and the circuit comprises a 64B / 66B coding module which is used for coding input 64-bit data to generate 66-bit data, and outputting the 66-bit data to a bit width conversion module; the bit width conversion module is used for performing bit width adjustment on 66-bit data to obtain w-bit data and outputting the w-bit data to the configurable RS parallel coding module, and w is the bit width of an RS code symbol; the RS coding parallel generator polynomial coefficient calculation module is used for outputting a parallel coefficient to the configurable RS parallel coding module according to configuration information, and the configurable RS parallel coding module is used for completing configuration and coding based on w-bit data and the parallel coefficient and outputting a coding result to the output cache module; and the output cache module realizes the output of the data and the verification data. The joint coding circuit has the advantages of being high in resource efficiency and excellent in performance.

The application discloses an anti-interference millimeter wave radar dynamic target detection system based on deep learning, relates to the technical field of radar signal processing and deep learning fusion, and comprises a preprocessing module, a feature extraction module and a long-time sequence global dependence modeling module.The preprocessing module is used for high-fidelity preprocessing of original millimeter wave radar signals, removing static clutter, completing phase correction, and outputting high signal-to-noise ratio dynamic signals after double-sided filtering denoising.The feature extraction module is used for front-end feature extraction of the high signal-to-noise ratio dynamic signals, parallel generation of a modulated phase stream and a feature stream jointly represented by covariance and trajectory through separation of amplitude and phase information, and output of two heterogeneous feature streams.The long-time sequence global dependence modeling module is used for feature fusion of the two heterogeneous feature streams, and the two heterogeneous feature streams are fused into a unified fusion feature map through an element-by-element addition operation.The fusion feature map is input into a Transform network, and a dynamic target detection result is output.The dynamic target detection precision and robustness in a complex environment are effectively improved.