176 results about "Decoder architecture" patented technology

In various examples, multimodal image data may be used to generate a set of top-down tile images, which are applied to a deep neural network generator architecture model to produce lane marking-specific heatmap images corresponding to the set of top-down tile images. The multimodal sensor data may include LIDAR-captured intensity channel data, LIDAR-captured feature height channel data, and optical color image channel data. The set of top-down tile images may be processed by the generator model to automatically detect lane boundaries and navigation boundaries to generate pixel-level heatmap images that may classify lane markings by marking characteristics such as line type and / or color. The generator model may comprise an encoder-decoder architecture, with multiscale feature extraction and / or context extraction functional layers intervening between the encoder model and the decoder model.

The invention provides a microcrack precise segmentation method integrating feature fusion and convolution attention, and belongs to the field of image processing. According to the method, a crack segmentation network based on an encoder-decoder architecture is constructed, a convolution block attention module is introduced at an encoder end, background noise is adaptively suppressed and obvious characteristics of cracks are enhanced through a channel and space dual attention mechanism, and the method is suitable for the adaptive segmentation of the cracks on the premise of almost not increasing the calculation overhead. The sensitivity of the model to microcracks is improved; a feature fusion module is introduced at a decoder end, and cooperation of low-layer details and high-layer semantics is realized through cross-layer fusion, so that a semantic gap is effectively bridged, detail loss caused by traditional convolution stacking is avoided, and continuity and a complete topological structure of a long and narrow crack are ensured. According to the method, through collaborative optimization of multi-scale feature extraction and an attention mechanism, accurate capture of the saliency features of the crack and effective suppression of complex background interference are realized, and the detection sensitivity and overall segmentation consistency of the micro-crack are remarkably improved.

Disclosed in the present invention is a non-autoregressive Transformer-based modeling method for a 4-level pulse amplitude modulation high-speed transmitter. The method involves establishing a deep learning model having an encoder-decoder architecture to predict the behavior of a 4-level pulse amplitude modulation transmitter. An encoder processes unordered non-sequential inputs, including an input signal parameter and link parameters, to generate a context vector and then transmit same to a decoder. The decoder uses both the context vector generated by the encoder and a transmitter output signal sequence to generate a categorical probability distribution for each point in the sequence one by one. The model is trained using a random masking strategy, and inference is performed by means of non-autoregressive decoding and filtering, so that the model can perform parallel prediction on an output sequence, and perform a filtering process to predict an output signal. Compared to traditional simulation methods, the present invention achieves a significant acceleration effect, particularly when processing multi-link systems.

The invention relates to the technical field of wireless communication, and discloses an intelligent channel prediction method based on a deep multi-layer perceptron architecture in order to solve the problem of high-dimensional near-field cascade channel prediction in a reconfigurable intelligent surface assisted low-orbit satellite communication system. Designing a frequency domain MLP architecture to learn global time sequence characteristics of a cascade channel; secondly, considering high-dimensional spatial domain characteristics of a near-field cascade channel, decomposing the cascade channel into a plurality of feature dimensions, and proposing a multi-branch MLP architecture to realize global spatial domain correlation modeling; and finally, by taking the frequency domain MLP and the multi-branch MLP as basic network feature extraction modules, designing a channel prediction network model based on an encoder-decoder architecture. According to the method disclosed by the invention, the accurate prediction of the downlink CSI can be realized by utilizing the historical channel state information estimated by the uplink communication, and the proposed channel prediction model has better channel prediction precision and model convergence speed.

The invention belongs to the technical field of artificial intelligence and data science, and discloses a causal discovery method and related device based on a large model and structural constraint, comprising the following steps: according to meta-information of each causal variable, adopting a large language model and combining multiple rounds of dialogues, multi-role play and auxiliary cue word modes, generating an initial potential causal graph of each causal variable; repeating the following steps until a preset condition is met, and outputting a candidate causal graph with the highest value score: generating the candidate causal graph by adopting a policy network based on an encoder-decoder architecture, and carrying out iterative optimization on the candidate causal graph by adopting a reinforcement learning algorithm based on an act-critic framework; and calculating a value score of the candidate causal graph after iterative optimization according to the value network, and updating parameters of the policy network based on the encoder-decoder architecture according to the value score. The efficiency and accuracy of causal relationship recognition are remarkably improved, and a causal relationship graph with high confidence can be generated in a complex data environment.

The invention relates to a machine tool thermal error modeling method based on a time convolution network and transfer learning, which comprises the following steps: data acquisition: providing a network based on an encoder-decoder architecture based on the time convolution network, inputting the acquired data of source (sufficient) and target (few) machine tools into an encoder, and carrying out data processing on the encoder; the method comprises the following steps of: coding and decoding to output a reconstruction sequence, and adding a reconstruction loss function which is calculated by the reconstruction sequence output by a decoder on a loss function with only one predicted value, so that deep features in thermal error data can be extracted through the architecture, and the prediction precision is improved; in order to achieve enough prediction precision on another target machine tool, transfer learning is used to obtain a loss function adapted to a field, the loss functions are summarized, iterative updating is repeated until deep learning regression algorithm training is completed, and accurate prediction of the thermal error value on the target machine tool is realized.

The invention discloses a multi-scale point cloud pre-training method for sampling from top to bottom, equipment and a medium. According to the method, through a U-Net type asymmetric encoder-decoder architecture and in combination with a multi-scale mask, the extraction capability of local features and global information of point cloud data is enhanced. According to the method, firstly, input point cloud data is subjected to multi-scale division and mask processing, then multi-scale features are extracted through an encoder, a masked point cloud area is reconstructed through a decoder, and a masked local point cloud block is recovered through a linear projection layer. According to the invention, the problem of multi-scale information leakage in the prior art is effectively avoided, and the point cloud data processing precision is improved. After self-supervised pre-training, the model shows excellent performance in a plurality of downstream tasks, especially in applications such as 3D shape classification, object detection and partial segmentation. The method has the advantages of high robustness and wide application prospect.

The invention discloses a large language model pre-training method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining multiple types of text data, and carrying out the preprocessing of the text data, and obtaining an initial pre-training data set; the method comprises the following steps: constructing a generative large language model based on a decoder architecture of a Transform model; according to the initial pre-training data set, performing multi-stage training on the generative large language model; forming a self-distillation mixed data set according to the teacher model and the initial pre-training data set; the trained generative large language model in each stage is used as a teacher model; training a student model by using the self-distillation mixed data set; the student model is a generative large language model at a distance of k stages after the teacher model; and then carrying out cycle training. According to the method, the comprehensiveness of pre-training learning is improved by using the self-distillation mixed data set, and the problem of repeated learning of low-quality pre-training data can be effectively solved.

The invention relates to the technical field of short-term load prediction in a power system, and discloses a short-term load prediction system based on a coder-decoder architecture, and the system is characterized in that a coder is used for extracting local features of a power load mode; and the decoder is used for converting the local features of the power load mode into predicted power load values and outputting the predicted power load values. The encoder is realized by a multi-scale expansion causal convolutional network MSDCC, and the decoder is realized by a bidirectional long short-term memory network BiLSTM. The prediction system construction method comprises the following steps: extracting related data from a historical database, preprocessing and analyzing the data, and constructing a predictor matrix; an MSDCC encoder is constructed; a BiLSTM decoder is constructed; and combining the encoder-decoder architecture to construct a short-term load prediction model. According to the method, the size of the feature map is effectively limited, model parameters are reduced, overfitting is avoided, calculation requirements are controlled, non-linear features are efficiently captured, meanwhile, time keeping complexity is low, and therefore the method has the advantages of being high in prediction efficiency and high in prediction precision.

The embodiment of the invention provides a defect detection method for an ATE wafer graph based on a double-task CNN, and the method comprises the steps: obtaining a wafer graph generated by automatic testing equipment, and carrying out the data processing of the wafer graph; constructing a network model structure which has double task branches and shares one encoder; taking the processed wafer graph data as the input of the network model structure, and carrying out multi-stage strategy training to obtain an optimal network model; and deploying the optimal network model into hardware equipment, taking a wafer graph obtained by the hardware equipment in real time as input of the optimal network model, and performing synchronous reasoning and output to obtain a defect detection result. By using a convolutional neural network model which is based on an encoder-decoder architecture and comprises classification and segmentation dual task branches, wafer defects are automatically, quickly and accurately identified and positioned, the production line process health condition is timely and effectively evaluated, and major production accidents such as large-scale yield loss are avoided.

The invention relates to the technical field of infrared light machine system thermal deformation prediction, in particular to a multi-material structure thermally induced stress deformation prediction method based on a graph neural network. The method comprises the steps of data set establishment, graph structure establishment, graph neural network model establishment and training and model and parameter optimization. Finite element nodes correspond to graph nodes, finite element edges correspond to graph edges, an encoder-message passing-decoder architecture model is established, and node states are updated through a three-layer physical symmetry message passing mechanism. Physical constraint loss including minimum displacement smoothness constraint and stress continuity constraint is innovatively added into a loss function. Compared with traditional finite element calculation, the method has the advantages that the speed is increased by more than 100 times, high hardware adaptability is achieved, the black box limitation of a data-driven neural network model is broken through, thermally induced stress deformation analysis caused by different material coefficients can be processed, the adaptability to geometric changes is high, and good engineering application value is achieved.

The invention provides an industrial defect detection method based on a multi-scale Transform, and the method comprises the steps: employing a pre-training network to extract multi-scale deep features, introducing a multi-scale Transform encoder-decoder architecture, achieving the long-distance dependence modeling through a multi-head self-attention mechanism, and obtaining the global representation capability exceeding a convolution network. Compared with CNN-U-Net, lossless jump connection is designed, 512-dimensional features generated by an encoder can be completely transmitted to a decoder, and tiny defect information loss caused by a traditional dimension reduction mode is avoided; the calculation complexity of global attention is remarkably reduced through a three-level local attention decomposition mechanism and is only 0.47% of that of a standard Transform, and the industrial real-time detection requirement is met; and self-adaptive error normalization and multi-scale smooth fusion are combined, so that false alarms at complex textures are effectively inhibited, and patch boundary artifacts are eliminated. Finally, according to the method, a pixel-level abnormal score graph is generated based on reconstruction deviation, and high-precision positioning and detection of multi-scale defects are achieved.

The invention discloses a voice model compression method, electronic equipment and a storage medium. The method is specially designed for a large-scale sequence-to-sequence speech recognition model with an encoder-decoder architecture. According to the method, the decoder and the encoder are pruned in sequence, so that heavy back propagation calculation is avoided. According to the method, under the condition of no back propagation or retraining, the parameters of the Whisper-large model can be reduced by about 60%, and the performance of the model on various data sets is hardly influenced. Meanwhile, the method is suitable for a multi-language data set, and the pruned model keeps good robustness and generalization in the aspect of multi-language ability. According to the innovation, the threshold of large-scale model deployment is greatly reduced, and the method is easier to apply in a resource-limited environment.

The invention discloses a CBCT tooth segmentation method and system based on an anatomical perception cascade network, and the method comprises the steps: a first stage, carrying out the simplified dichotomy segmentation based on an original CBCT image and a coarse segmentation network taking 3D U-Net as a trunk, outputting maxillary and mandibular tooth probability graphs, and taking the maxillary and mandibular tooth probability graphs as prior information to guide the generation of an SDM; in the second stage, the original CBCT image and the calibrated maxillary tooth probability graph and the calibrated mandibular tooth probability graph are spliced together, a formed multi-channel input tensor is input into a fine segmentation network, the fine segmentation network takes Residual U-Net as a trunk, and an improved AGBR module and an improved SDMAA module are integrated in an encoder-decoder architecture of the fine segmentation network; and the decoder fuses all refined and re-calibrated feature maps, and upsamples and reconstructs 42 types of instance segmentation results with correct topology and clear boundaries. According to the method, the problem that in the prior art, when the inherent and local boundary fuzzy defect in CBCT is overcome, an effective pertinence mechanism is lacked is solved, and precise and robust 42-class instance segmentation can be achieved.

A method for learning robotic grasping of objects includes inputting images of scenes with objects in random configurations from a training dataset to a grasping neural network having an encoder-decoder architecture. Color frames of the images are input to a first encoder and corresponding depth frames are input to a second encoder. Features extracted from the first and second encoders are fused via a third encoder to produce a fusion encoding. The fusion encoding is inputted to first and second decoders to respectively produce a grasp map output and a segmentation map output. The first and second decoders are jointly trained in a supervised process using a common loss function defined by a grasp loss of the first decoder and a segmentation loss of the second decoder. During the training, the grasp map output of the first decoder is guided by the segmentation map output of the second decoder.

The invention discloses a hydropower station operation state intelligent prediction method and system based on multi-source data fusion. The method comprises the steps that heterogeneous data including but not limited to water level, flow, generating capacity, equipment operation parameters, environmental meteorological data, hydrological data and the like are collected in real time from multiple data sources of a sensor network, an SCADA system, a meteorological station and a hydrological station of a hydropower station; the method comprises the following steps of: preprocessing collected original data, extracting time sequence features, fusing multi-source data, taking a fused multi-dimensional feature vector as input, and learning and training through an encoder-decoder architecture of a model so as to capture a complex dynamic mode of an operation state of a hydropower station; and inputting the data acquired and preprocessed in real time into the trained AI prediction model, and generating prediction results of key operation parameters, equipment states, residual life and potential risks of the hydropower station. The method can more accurately capture the complex nonlinear relation in the operation of the hydropower station, and significantly improves the prediction precision of key parameters and risks.

The invention discloses a code abstract generation method oriented to semantic structure feature joint enhancement. According to the method, an encoder-decoder architecture is followed, and a multi-form combined semantic feature enhancement method is adopted to automatically generate a code abstract. Wherein the encoder learns code semantics at token and statement levels, and the decoder receives code embedding and generates a code abstract in combination with an attention mechanism. Unlike a single tree or graph coding process, the method uses graph representation to enhance the interval divided tree structure. The joint decoder aggregates the text sequence information and the structure sequence information containing the location node features, and predicts and generates a model understood code summary text. According to the method, the problems of repetition and redundancy of code texts with different lengths caused by disordered code text features are relieved. In a relatively complex code text data set, a word ambiguity phenomenon caused by insufficient knowledge information understanding is reduced, and a code abstract with relatively high quality is generated. Meanwhile, the method has good convergence performance.

The embodiment of the invention provides an encrypted traffic classification model training method, an electronic device, a storage medium and a program product, through a self-supervised learning framework, an unmarked patch is utilized to embed and express a training encoder-decoder structure, so that the model can autonomously learn universal spatio-temporal features from an encrypted traffic grey-scale map. The mask reconstruction task promotes the model to understand the internal structure rule of the traffic data, the understanding depth of the model for the input data is enhanced, the adaptability of the model to the distributed data is improved, and the generalization ability, adaptability and robustness of the model are also improved.

The invention discloses a Conformer-based mixed self-attention and convolution improved speech recognition method and system. The recognition method comprises the steps of obtaining speech feature representation; the speech feature representation is input to a Conformer encoder for encoding processing, high-dimensional feature representation is obtained, and the Conformer encoder comprises a convolution module, an attention linear enhancement module and a feedforward network residual scaling module; according to the high-dimensional feature representation, bidirectional decoding is carried out through a bidirectional Transform decoder, and bidirectional decoding output is obtained; and carrying out merging processing on the bidirectional decoding output to obtain a speech recognition result. According to the method, the modeling capacity for long-time dependence and global context is enhanced through a bidirectional decoder architecture, the processing capacity of the model in a long sequence is improved through ALiBi relative position coding, and the model can be flexibly adjusted in a noise environment through a right decoder weighting adjustment mechanism.

The invention provides an underwater video enhancement method based on semantic guidance, and the method comprises the following steps: S1, carrying out the global feature enhancement of a multi-frame video, and obtaining a feature map after global enhancement; s2, performing region enhancement on the context features of the multi-frame video to obtain a feature map after region enhancement; s3, performing semantic guidance on the feature map after region enhancement to obtain a region enhancement feature map after semantic guidance; and S4, fusing the semantically guided region enhanced feature map and the globally enhanced feature map to obtain an output effect of the intermediate frame. According to the global feature enhancement method disclosed by the invention, an encoder-decoder architecture is mainly used, multi-frame time sequence information is fully utilized to perform spatial alignment of adjacent multi-frame features by using grouping spatial shift, context information is fully learned to suppress irrelevant noise, the robustness of feature representation is enhanced, and the effect of global feature enhancement is achieved.

The invention discloses an overhead power line icing thickness detection method and system based on a structure perception framework, and the method comprises the steps: firstly obtaining point cloud data of a power transmission line, preprocessing the point cloud data, and inputting the preprocessed point cloud data into a structure perception semantic segmentation network; the network adopts a four-level encoder-decoder architecture, an encoder extracts multi-scale features through local space encoding, cross-scale space attention fusion and local aggregation operation, and a decoder realizes feature fusion in combination with a jumper connection mechanism and outputs a lead semantic segmentation result; extracting a conductor point set based on a segmentation result and carrying out Euclidean clustering strand splitting; performing biplane projection and polynomial fitting on each strand of wire to construct a three-dimensional center line; and finally calculating the distance from the point to the center line, selecting a peripheral point set to calculate the wire envelope radius, and combining the radius of the bare wire to obtain the icing thickness. Measurement deviation caused by point cloud shielding, wire bending and asymmetric icing is effectively solved, millimeter-level precision non-contact ice thickness detection in a complex environment is achieved, and reliable technical support is provided for power grid disaster prevention.

The invention discloses a printed circuit board element placement method and device based on a traveling salesman problem and a medium, belongs to the technical field of traveling salesman problems and printed circuit boards, and aims to solve the technical problem that diversified constraint conditions and efficiency are difficult to consider when the placement sequence of traditional printed circuit board elements is optimized. According to the technical scheme, the method comprises the following steps: constructing a training data set; performing feature coding on the placement requirement of the printed circuit board element by using an encoder to obtain a coding result; decoding a coding result through a decoder based on a double-pointer network, and generating an element placement sequence meeting the priority constraint; and training the encoder-decoder architecture by adopting a near-end strategy optimization algorithm based on reinforcement learning, and finally obtaining a deep reinforcement learning model for solving the priority constraint traveling salesman problem for placing the printed circuit board element.

The invention discloses an improved UNet type RGB-D (Red Green Blue-Decision) based saliency target detection system. According to the method, a feature relevance perception fusion module (CPFM) is introduced to fully mine prior relevance of significant features between RGB-D modes, and it is ensured that fused feature representation is more complete; by adopting a double-flow encoder-decoder architecture, an RGB image and a depth map are respectively processed, and the detection performance of a saliency target is collaboratively optimized through modularization. Compared with an existing single-flow or simple fusion architecture, the PAliUNet can more fully mine complementary information between the RGB and the depth map through independent processing and fusion of the two branches, and the detection effect is improved. The PAliUNet combines a CPFM module, an EnASP module and an SADM module, cooperative work is carried out in the three stages of coding, fusion and decoding, the robustness and detection precision of the model are greatly improved, and especially under the conditions of low-quality depth maps and complex backgrounds, remarkable performance advantages are shown.

The invention discloses a robot control system and method based on AI voice, and belongs to the technical field of robot intelligent control. The system comprises a local voice processing module, a dialogue management and context understanding module, an instruction mapping and generating module and a robot communication and control module. The local voice processing module is integrated with a 16 kHz sampling ASR engine, voice-to-text conversion is realized through a 6-layer LSTM network and beam search decoding, and the TTS engine adopts a WaveNet architecture to synthesize voice. A 7 billion parameter large language model is arranged in the dialogue management module, a 32-layer Transform decoder architecture is adopted, 10 rounds of dialogue history is maintained, and intention recognition and entity extraction are achieved through a multi-head self-attention mechanism. And the instruction mapping module performs trajectory planning by adopting quintic polynomial interpolation and quaternion spherical linear interpolation. According to the invention, full-link localization processing is realized, the speech recognition delay is 180 milliseconds, the position error is less than 0.5 mm, and the problems of high delay, high privacy risk and weak complex instruction understanding ability in the prior art are effectively solved.

The invention discloses a deep learning-based bridge scour damage identification method and system, and solves the problems of poor universality and limited scour damage identification efficiency and precision of the existing bridge scour damage identification method. The method comprises the following steps: constructing an original data set; preprocessing the data; the method comprises the following steps: constructing a sequence-to-sequence multi-parameter model of an encoder-decoder architecture by adopting a double-layer long short-term memory network, and adding a Dropout layer and a LayerNorm layer in the model for standardization; a loss function of model training and optimization is defined, an Adam optimization algorithm is adopted, L2 regularization is added, an early stop method is utilized to train and optimize the Seq2Seq multi-parameter model, and a model with the minimum verification set loss is selected as an optimal prediction model to be used for recognizing the bridge scouring damage. According to the method, the complex scouring damage condition of the multi-span bridge can be rapidly identified, the scouring damage identification efficiency and precision are improved, the universality is higher, and the identification efficiency is higher.

The invention discloses an unbalanced node classification method based on graph contrast learning, and belongs to the technical field of artificial intelligence. According to the method, a graph comparison learning framework of adaptive balance data is provided, minority classes can be automatically identified, the minority class performance is improved, and then the overall performance of the model is improved. Firstly, an Encoder-Decoder architecture is used for pre-training, and compared with a traditional pseudo tag generation method, an unbalance rate self-adaptive sampling strategy is designed, the unbalance rate of data is calculated according to pseudo tags, and the sampling strategy is selected in a self-adaptive mode. For a data set with a low unbalance rate, a simple downsampling method is adopted, and the proportion of minority class information is increased; for a data set with a relatively high unbalance rate, a mixed sampling strategy is adopted, and over-sampling and down-sampling are combined, so that the information loss of majority of nodes is reduced while the information proportion of minority of nodes is increased. In addition, the pre-training model used in the invention can provide more accurate label information, thereby improving the distinguishing ability of the model in subsequent GCL training. Then, a new data augmentation technology is designed, in the node masking process, pseudo label information is utilized, information of minority class nodes is reserved preferentially, and meanwhile majority class nodes are masked; the method is helpful for the model to better capture minority class features in an unbalanced data set. And finally, a linear classifier is used for classification. According to the method, the unbalanced node classification performance under the self-supervision condition can be effectively improved.

The embodiment of the invention relates to a processing method and device for a dense encoder realized based on a large language model, and the method comprises the steps: selecting a large language model which has completed pre-training and NLP task fine tuning and is realized based on a pure decoder architecture as a target model, the bidirectional encoders are obtained through a transformation mode of solidifying a causal mask matrix used by a target model decoder in a reasoning process into an all-one matrix, and the embedded encoding module of the target model and the plurality of bidirectional encoders are connected in sequence to form a dense encoder; performing first-stage fine adjustment on the dense encoder through a shielding word prediction task, and performing second-stage fine adjustment on the dense encoder through an unsupervised contrast learning mechanism; and after the fine tuning is finished, constructing a document vector library for a target document library appointed by the user by utilizing the dense encoder, and providing retrieval service for the target document library based on the document vector library and the dense encoder. The dense encoder is used for processing a text retrieval task, so that the retrieval accuracy can be improved.

The invention discloses a multi-person redirection walking and position prediction method based on an artificial potential field, and belongs to the technical field of virtual reality and robot navigation. Aiming at the problems of fixed repulsive force weight, high collision risk and frequent resetting in a multi-user dynamic scene in a traditional redirection walking algorithm and collision caused by insufficient short-term position prediction precision of predictive redirection, the weight distribution of repulsive force between an environment and users is adjusted in real time according to user density through a dynamic situation field weight optimization mechanism; the repulsive force intensity is enhanced when a user approaches an obstacle or a dense area, meanwhile, a safety distance and direction sensing mechanism is introduced, the user movement direction and obstacle included angle analysis is combined, deviation from a high-risk area is guided preferentially, and the invalid reset frequency is reduced. The prediction model adopts an encoder-decoder architecture, an attention mechanism is embedded in a multi-layer long-short-term memory network, key time sequence characteristics in a space-time trajectory are dynamically captured, and the short-term position prediction precision is improved.

A method for testing and debugging interactions of collaborative mixed reality objects is disclosed. In one embodiment, the method includes receiving an input including a first mixed reality object expressed by a first set of attributes, a second mixed reality object expressed by a second set of attributes, a first individual test case associated with the first mixed reality object, and a second individual test case associated with the second mixed reality object. The method automatically generates a collaborative mixed reality test case from the input to evaluate an interaction of the first mixed reality object with the second mixed reality object within the collaborative mixed reality environment. In some embodiments, a generated AI-based encoder-decoder architecture is used to generate collaborative mixed reality test cases from inputs. A corresponding system and computer program product are also disclosed.

The invention discloses an epileptic seizure video detection method based on a memory enhanced double-branch auto-encoder. The method comprises the following steps: firstly, preparing and preprocessing data; secondly, constructing an ME-DBAE model which comprises a query generation module, a multi-head memory module and a double-branch decoder module; training the model by using training set data, finally inputting a to-be-detected video into the trained model to obtain an epileptic seizure score, and judging whether the video belongs to an epileptic seizure sample or not through a threshold value. By designing the multi-head memory module, diversified normal motion prototypes are stored in parallel, and the limitation that a single memory structure is difficult to adapt to complex normal actions is broken through; and by constructing a double-branch decoder architecture, double tasks of input reconstruction and future frame prediction are cooperatively realized, and meanwhile, space structure and time trend characteristics of normal motion are captured, so that epileptic seizure actions and normal actions are effectively distinguished.