48 results about "Sequence segmentation" patented technology

The invention relates to the technical field of network space security, and provides a DoH malicious tunnel traffic detection method and system based on feature fusion, and the method comprises the steps: carrying out the preprocessing of obtained to-be-detected DoH traffic data, carrying out the sequence segmentation processing, obtaining a Token sequence, and extracting features based on a byte sequence feature extractor; statistical features are extracted and standardized, features are extracted through a statistical feature extraction sub-network, and statistical feature vectors are obtained; fusing the byte sequence feature vector with the statistical feature vector to obtain a classification result; the byte sequence feature extractor and the statistical feature extraction sub-network extract features, and training is carried out by adopting a semi-supervised learning framework of dynamic pseudo-tag screening for non-tag training samples. According to the method, multi-modal features are fused, a semi-supervised learning mechanism is introduced, malicious DoH traffic generated by multiple DNS tunnel tools is effectively identified under a limited annotation data set, and the detection precision and generalization ability are improved.

The invention provides a bidirectional recurrent neural network acoustic logging curve reconstruction method. The reconstruction method comprises the steps of S1, acquiring data and performing sequence segmentation; s2, carrying out normalization processing on the logging curve; s3, carrying out superposition networking and training on the bidirectional recurrent neural network structure block; s4, model testing and parameter storage; and S5, reading the model parameter file to reload the model, importing the to-be-reconstructed logging data, generating a prediction curve, and storing and exporting a result. A bidirectional recurrent neural network algorithm of artificial intelligence deep learning is adopted, bidirectional depth sequence information contained in a logging curve is effectively captured, the curve reconstruction accuracy is improved, unmeasured, missing and low-quality curves are reconstructed, multi-curve information is fused, and the sensitivity of a reconstructed acoustic curve is improved. The system has the advantages of being low in cost, high in efficiency and high in precision.

The invention discloses a power grid intelligent inspection method and system combined with abnormal positioning, and relates to the technical field of intelligent inspection. According to the method, periodic sequence segmentation analysis is performed on the inspection data, and the rhythm stack difference sequence is constructed, so that burst fluctuation fragments in each period can be captured, and potential abnormal characteristics with repeated offset trends can be mined; through generation of a local offset structure coding sequence, a spatial offset mode between rhythm nodes is accurately described, and quantitative modeling of implicit structure changes in inspection data is realized. In addition, by comparing structural differences in the current period and historical data, abnormal clues with a complementary relationship on the spatial path are explored, on the basis, the judgment capability of an abnormal initial region is enhanced, and powerful support is provided for subsequent path recombination and risk early warning.

The invention provides a self-adaptive control method for improving new energy power generation grid-connected stability of a micro-grid, and the method comprises the steps: carrying out the description according to a corrected phase change trend, fusing the real-time feedback data of voltage deviation, generating a reactive compensation instruction sequence segmentation adjustment scheme, and determining a preliminary compensation capacity dynamic distribution strategy; the microgrid system is regulated and controlled in real time through the optimized control instruction sequence in combination with the power grid stability constraint condition, the actual execution effect of the compensation action is obtained, and whether the system voltage deviation meets the stability requirement or not is judged; according to the stability requirement judgment result of the system voltage deviation, the execution effect is fed back to a micro-grid control layer, a closed-loop mechanism is adopted to carry out iterative updating on the instruction sequence segmentation adjustment scheme, and a final reactive compensation regulation and control scheme is determined.

The invention discloses a large model reasoning method and device based on multi-level cache, a storage medium and equipment, and the method comprises the steps: receiving a reasoning task, and carrying out fixed-length lexical sequence segmentation processing on the reasoning task to determine at least one lexical sequence corresponding to the reasoning task, the lengths of other lexical element sequences except the last lexical element sequence in the lexical element sequences are fixed lengths, and the length of the last lexical element sequence is smaller than or equal to the fixed length; aiming at the lexical element sequence with the fixed length, a target data block corresponding to the lexical element sequence is inquired in a video memory, if the lexical element sequence does not exist in the video memory, the target data block corresponding to the lexical element sequence is inquired in a memory, and when the data blocks in the video memory are evicted out of the video memory, the data blocks are copied into the memory; if the tail lexical element sequence is smaller than the fixed length, the reasoning task is executed according to the target data block and the tail lexical element sequence through the large model; if the tail lexical element sequence is equal to the fixed length, the reasoning task is executed according to the target data block through the large model.

The invention belongs to the technical field of medical image processing, and discloses a medical image sequence semi-supervised segmentation model construction method and application, and the method comprises the steps: constructing a medical image sequence semi-supervised segmentation network which comprises a first segmentation network, a feature projection module, a second segmentation network, a space-time memory module and a channel and space attention module; utilizing a space-time memory module to obtain enhanced features containing semantic information; using double cross attention perception to obtain fusion features, and inputting the enhanced features and the fusion features into a second segmentation network with time perception to generate high-quality sequence segmentation; and training the medical image sequence semi-supervised segmentation network by using the consistency loss of the segmentation labels # imgabs0 # and # imgabs1 # output by the two segmentation networks, the labeled parts in the segmentation probability graphs # imgabs2 # and # imgabs3 # and the supervised loss of the corresponding manual annotation, and obtaining a trained medical image sequence semi-supervised segmentation model. According to the invention, the segmentation performance of the medical image sequence can be improved.

The application discloses a kind of multi-group one-dimensional frequency energy density spectrum ratio measurement precision evaluation method and system, belong to marine observation technical field, comprising: the training data sample of establishing including M group synchronous observation sea wave one-dimensional frequency energy density spectrum sequence;21 similarity evaluation value sequences are constructed;21 similarity evaluation value sequences are sorted according to the order of evaluation value from big to small, and at least two dimensions in sequence segmented average, overall average, fluctuation amplitude, fluctuation rate, distribution concentration and dispersion degree are analyzed, to determine the comprehensive weight coefficient of 21 similarity evaluation values;The one-dimensional frequency energy density spectrum sequence of two synchronous observations to be evaluated is obtained, and the corresponding 21 similarity evaluation values thereof are calculated;The final similarity evaluation value NRD is calculated, to evaluate the ratio measurement precision of two to be evaluated sequences.The application constructs a multidimensional evaluation system, can comprehensively reflect the similarity of two sea wave spectrum sequences.

The application discloses a kind of communication power supply operating state intelligent monitoring methods based on machine learning, comprising the following steps: S1, obtains multiple-source monitoring data and pre-processes;S2, sequence segmentation is carried out using sliding time window and statistical feature vector is extracted, forms state variable set;S3, each state variable is encoded to construct state node and establish weighted directed connection, and construct operating state evolution diagram;S4, state transition matrix is constructed, main transition feature is extracted and low rank is approximately reconstructed, sequence learning is carried out to construct state update function;S5, to state node, multiple-step recursion is carried out, and the Euclidean distance between recursive state and risk boundary is calculated, to determine state monitoring result;S6, error is calculated and state update function parameter is iteratively updated.The application can realize the multiple-step recursion prediction and risk trend identification of communication power supply operating state, improve the accuracy and stability of communication power supply operating state monitoring.

The invention discloses a bit sequence preprocessing method and an OOK symbol generation method and device, and belongs to the technical field of Internet of Things. The method comprises the following steps: acquiring a first bit sequence of which the length is a first number L; under the condition that the first number L is not the integral multiple of the second number M, processing the first bit sequence into a second bit sequence of which the length is a third number L '; wherein the second bit sequence is used for being divided into at least one sequence segment, the length of which is a second number M, for OOK modulation, and M OOK symbols corresponding to each sequence segment are obtained.

This invention discloses a multi-source electromagnetic noise suppression method based on noise classification and deep learning, belonging to the field of geophysical electromagnetic exploration technology. The method includes: segmenting and preprocessing the original electromagnetic observation sequence; using an improved U-Net network with an encoder embedded in a Mamba time-series modeling module for low-frequency noise suppression; identifying strong noise types in the data segments using a ROCKET classifier; based on the classification results, calling a second improved U-Net network trained for the corresponding noise type for class-based targeted denoising; and finally, splicing the data segments to obtain complete, high-quality data. This invention, through a phased processing framework of "low-frequency pre-suppression—noise classification—class-based denoising," combined with the strong time-series modeling capabilities of the Mamba module and the efficient classification performance of ROCKET, significantly improves the suppression accuracy and signal fidelity for complex, multi-source electromagnetic noise, and is particularly suitable for processing ground and airborne electromagnetic exploration data.

The invention relates to a parameter reduction meta learning load prediction method based on sequence segmentation and a bidirectional GRU network. According to the method, knowledge migration from source load data to a target prediction task is realized based on meta-learning, and the prediction effect under the conditions of small samples and distribution drift is improved. Firstly, a task data division method based on decomposition characteristic time sequence distribution similarity quantization DP is provided, it is ensured that each task has unique time sequence characteristics, and the meta-learning generalization ability is improved. Secondly, feature gradient parameter reduction is introduced into inner layer circulation of meta-learning, and the memory overfitting problem is relieved; besides, a bidirectional GRU network is introduced as a basic model, the long-term and short-term dependency capture capability of the sequence is enhanced, the network structure is kept simple, model parameters can quickly adapt to a target task under limited training data, and good prediction performance is achieved under the conditions of small samples and distribution drift.

This application relates to the field of generator fault identification technology, and discloses a method and apparatus for early fault identification of steam turbine generators based on vibration. The method includes: acquiring training samples; extracting multiple continuous vibration signal segments from each vibration signal sequence sample, constructing grayscale images based on each continuous vibration signal segment, and generating a grayscale image set; training an initial convolutional neural network classifier using an unsupervised sequence segmentation method on each grayscale image set to obtain a target convolutional neural network classifier. The unsupervised sequence segmentation method generates pseudo-labels by automatically determining the optimal early fault occurrence time for each grayscale image set, and uses the pseudo-labels to train the initial convolutional neural network classifier; acquiring the vibration signal to be identified in real time, dynamically constructing the vibration signal segment to be identified based on the acquired vibration signal, constructing the grayscale image to be identified based on it, and inputting the grayscale image to be identified into the target convolutional neural network classifier to determine whether a fault exists.

This invention belongs to the field of acoustic sensing demodulation, specifically a demodulation method for a distributed fiber optic acoustic wave sensing system based on FPGA time-sequence segmentation. This invention uses an FPGA as the processing module for signal demodulation. On the FPGA, three signals with a 120-degree phase difference output from the distributed fiber optic system are acquired. Then, time-sequence segmentation is achieved by FIFO buffering of the spatial point data, ensuring data alignment at each spatial point during signal demodulation. Finally, moving average and other signal demodulation processes are performed on the data at each spatial point in the time domain. This invention requires only a small amount of data buffering to achieve demodulation of the entire system, reaching the theoretical maximum distance for signal demodulation. Ultimately, it achieves signal demodulation of a long-distance distributed fiber optic acoustic wave sensing system with high repetition rate under low resource consumption, providing a new approach and method for hardware-based signal demodulation of distributed fiber optic sensing systems.

The invention relates to an all-sky aurora video unsupervised event segmentation method based on bidirectional perception, which comprises the following steps of: 1, inputting an aurora image sequence, randomly selecting a target frame, forming two subsequences which take the target frame as a center by using the front frame and the rear frame of the target frame as positive samples, and enabling the features of the target frame to have context information; step 2, carrying out comparative learning on the extracted target frame features to obtain a higher-level representation, masking the target frame, then carrying out bidirectional feature reconstruction, reconstructing the target frame, and training a reconstruction model; 3, testing the trained reconstruction model, performing reconstruction and similarity calculation on the aurora image sequence in the test set frame by frame to obtain an error curve, regarding the error curve as a boundary probability curve, performing smoothing processing by using a filter, setting a boundary probability threshold, and regarding a frame which is greater than the boundary probability threshold and is a peak value as a boundary frame; the method has the characteristic of high aurora sequence segmentation accuracy.

The invention relates to the technical field of generator fault identification, and discloses a vibration-based turbonator early fault identification method and device, and the method comprises the steps: obtaining a training sample; intercepting a plurality of continuous vibration signal segments from each vibration signal sequence sample, and constructing a grayscale image according to each continuous vibration signal segment to generate a grayscale image set; and training the initial convolutional neural network classifier by using each grayscale image set and adopting an unsupervised sequence segmentation method to obtain a target convolutional neural network classifier, and generating a pseudo-tag by automatically determining the optimal early fault occurrence time for each grayscale image set according to the unsupervised sequence segmentation method. Training an initial convolutional neural network classifier by using the pseudo labels; the method comprises the steps of obtaining a to-be-recognized vibration signal in real time, dynamically constructing a to-be-recognized vibration signal segment based on the obtained vibration signal, constructing a to-be-recognized grayscale image according to the to-be-recognized vibration signal segment, inputting the to-be-recognized grayscale image into a target convolutional neural network classifier, and determining whether a fault exists or not.

The invention relates to the technical field of ground surface settlement monitoring, and discloses a ground surface real-time monitoring system based on a shield construction tunnel. A data acquisition unit of the system obtains a displacement measurement value from an earth surface sensor and generates time sequence data; the sequence segmentation unit analyzes mutation points in the sequence and segments the sequence into a plurality of data segments according to the mutation points; an exception marking unit performs statistical test on each data segment to identify an exception data segment set; the trend evaluation unit performs trend line fitting on the abnormal set and evaluates the fluctuation degree of fitting deviation; the matrix conversion unit organizes the time sequence data into a specific matrix structure and performs matrix decomposition; a feature derivation unit derives feature parameters from the decomposition result; and the monitoring control unit adaptively adjusts a monitoring strategy according to the characteristic parameters and generates a settlement report. According to the system, accurate and self-adaptive monitoring and early warning of ground surface settlement are realized through multi-stage data processing and intelligent analysis.

The invention provides a neural network action recognition data processing method, and solves the problems that feature extraction in a traditional action recognition system depends on manual design, the model training efficiency is low, cross-platform deployment is difficult and the like. According to the method, by constructing an end-to-end deep learning model, direct mapping from original sensor data to gesture categories or sentences is achieved, meanwhile, multi-device deployment is supported by using cross-platform characteristics of TensorFlow, and an efficient and stable processing scheme is provided for wearable action recognition devices. The system specifically comprises a data preprocessing module, a 1D CNN model construction module, a model training module and a model deployment module. The data preprocessing module supports loading of an open source triboelectric sensor action data set from multiple formats, and signal standardization, data enhancement and sequence segmentation are achieved. The 1D CNN model construction module adopts a multi-layer 1D convolution structure, and introduces a time sequence attention mechanism and multi-scale feature fusion, thereby improving the recognition precision. The model training module uses an Adam optimizer and a learning rate attenuation strategy to realize early stop and check point storage. And the model deployment module supports model export, model compression and cross-platform deployment. Experimental results show that the recognition accuracy of the 1D CNN model on a test set reaches 91.3%, which is superior to that of a traditional method, and the 1D CNN model shows good performance in a continuous action sentence recognition task.

This invention relates to the field of fault location technology, and discloses a method and system for rapid fault location in operation and maintenance. The method includes: acquiring operational data and denoising it to obtain a denoised data sequence; segmenting the sequence and identifying anomalies, determining the anomaly range through trend offset, and obtaining anomaly markers; if the waveform distortion metric of the anomaly marker exceeds a threshold, obtaining the anomaly location through time series analysis; extracting periodic sequences, grouping similar behaviors among devices to obtain associated device clusters; constructing a causal graph among devices, obtaining a list of potential anomalous devices through propagation path backtracking and node dependency sorting; assessing the impact degree of each device and identifying the faulty device; and obtaining the final fault location based on the topology information of the faulty device through cross-cluster association decomposition, causal edge weight adjustment, and graph structure simplification. This method can accurately and quickly locate the source of a fault in complex network environments with severe noise interference and tight device coupling.

The application discloses a kind of multi-source data's reliable alignment method, device, equipment, storage medium and product.Identify the discontinuous fragment of the measurement data sequence of each sensing device and associate sensing device identification, obtain measurement data variation section set;Determine the data channel of each associated sensing device based on sensing device identification and determine overlapping path section through network diagnosis, generate path network atlas;The target measurement data sequence that flows through each overlapping path section is segmented and processed to obtain target data segment, and different target data segments that flow through each overlapping path section at the same time are subjected to data flow consistency analysis to determine an interference channel list;Each target data segment in the target measurement data sequence via the interference channel is subjected to validity determination, and the effective data segment and the measurement data sequence of the non-interference channel are aligned and combined according to time to obtain multi-source reliable alignment data.It ensures data quality, avoids distortion information mixed into fusion results, and improves the accuracy and reliability of downstream decision-making.

This application relates to the field of image processing technology, specifically to a high-voltage power supply DNA sequencing visual inspection method and system. The method includes: acquiring a four-channel DNA electrophoresis image using Sanger sequencing technology; acquiring contours in the DNA electrophoresis grayscale image; constructing a DNA local density confidence coefficient and local DNA density parameters for each pixel based on the grayscale values ​​of each pixel within a channel and its neighboring pixels; acquiring the density run matrix for each pixel; constructing a DNA density distribution heterogeneity coefficient and channel tailing effect evaluation parameters for each pixel, and based on these, constructing a sequencing segmentation probability coefficient for each pixel; acquiring a decision threshold for each pixel; using the center of each contour as a seed point, and acquiring each DNA fragment based on an adaptive threshold using a region growing method, thereby obtaining the DNA sequence. This application can improve the accuracy of detecting the base sequence of a DNA sequence.

The application discloses a cross-sequence multi-behavior sequence recommendation method based on a Transformer. The method is as follows: firstly, a user historical interaction sequence is segmented, and then cross-sequence heterogeneous relationship propagation is performed; then, short-term interaction mode learning is performed; next, global representation aggregation is performed; finally, a recommendation content list is generated by combining user recent preferences and item embedding, and meanwhile, content priority is considered; and model parameters are adjusted by using user historical interaction for label data enhancement. The application proposes a user sequence segmentation method, guarantees that a subsequence maintains relatively concentrated and stable preferences of a user, and enables the model to more accurately capture dynamic changing preferences of the user, proposes a similar subsequence determination mechanism and a method for performing cross-sequence information propagation between similar subsequences, and preserves time coding information in the information propagation process, thereby improving the diversity and accuracy of recommended contents.

The application provides an ultra-long sequence processing method based on an improved Transform model, and the method is implemented based on a WSformer model which is constructed by improving the Transform model, and specifically comprises the following steps: S1, pre-training of the WSformer model; S1.1, an ultra-long sequence is divided into small sequences by using a sequence segmentation method; S1.2, feature extraction based on a double-layer attention mechanism; word-level feature extraction and sentence-level feature extraction are performed by using the double-layer attention mechanism, and feature encoding of the entire sequence is obtained; S2, improved position vector coding; in the coding process of step S1.2, the position coding of the word is adjusted by using a trigonometric function, and the coding of the position vector is realized. The method can effectively reduce the time cost of long text sequence coding, more accurately depict the position relationship between words, realize effective modeling of long sequence texts, and improve the prediction accuracy and the calculation efficiency.

This invention discloses a path tracking method for tracked vehicles based on a liquid time constant neural network, belonging to the field of intelligent control technology for tracked vehicles. The method includes: establishing an expert demonstration database based on MPC (Multi-Process Control) and collecting path tracking error and drive motor torque data; performing robust scaling normalization and long sequence segmentation preprocessing on the data; building an LTCNN model with an NCP (Non-Conceptual Processing) architecture in the PyTorch framework and training it using the AdamW optimizer through MPC behavior cloning; inputting the real-time path tracking error into the trained model and outputting the drive motor torque to control the vehicle to track the desired path. This invention ensures path tracking accuracy through MPC behavior cloning, achieving a lateral error RMSE of 0.089m under small curvature conditions and 0.13m under large curvature conditions; the forward inference of LTCNN improves real-time performance, saving 34.8% of runtime compared to MPC; data preprocessing and the NCP architecture enhance the model's robustness and interpretability, resulting in smoother output torque, making it suitable for path tracking of tracked vehicles in complex off-road environments.

The invention provides a rapid gas detection method, system and equipment for an electronic nose and a storage medium, and the method comprises the following steps: carrying out sequence segmentation on an input gas sensor signal, and extracting a time sequence feature; the time sequence features are embedded into the potential space, embedded information is obtained, and the embedded information carries the local time sequence features of the gas sensor signals; carrying out global mode feature extraction and pooling fusion on the embedded information, and integrating global features of the gas sensor signals to obtain gas signal characterization; the gas signal characterization is mapped to a semantic space, a gas classification is identified, and a corresponding gas concentration is determined. Embedded information rich in gas semantic information is obtained through sequence segmentation, signal redundancy and noise can be reduced, and overfitting is prevented. Therefore, the method can realize high-precision rapid gas classification and concentration estimation in the gas detection process.

The invention discloses an archive data structured extraction method based on machine learning, and the method comprises the following steps: carrying out the normalization processing, noise suppression and sequence segmentation of an archive text, and obtaining a processable text sequence; the method comprises the following steps: generating a plurality of candidate field fragments in a text sequence, extracting multi-level features for each fragment, carrying out joint evaluation on boundary combination and feature representation of the field fragments by introducing a minimum description length criterion, determining an optimal fragment division result, generating compressed feature representation, and constructing sequence input by taking the field fragments as an integral unit. And utilizing a semi-Markov conditional random field model to carry out overall modeling and joint decoding on the variable-length field to obtain a corresponding relationship between the field tag and the field fragment, and finally generating a structured archive data result. According to the method, the dependence on manual rules can be reduced, and the accuracy and adaptability of structured extraction of the archive data are improved.

The present application relates to the technical field of database concurrency control, and particularly relates to a multi-instance concurrent sequence segment competition method, system and computer readable storage medium based on a TDSQL database. The method first acquires sequence segment information corresponding to a current instance from a sequence list; then calculates a target sequence segment range based on the sequence segment information, and occupies the target sequence segment in the TDSQL database through an update operation with an original value condition, the update operation including a conditional judgment of whether the starting value of the current sequence segment is consistent with a pre-stored value; finally, it is judged whether the update operation successfully occupies the target sequence segment, the successfully occupied sequence segment is cached to the local instance, and an incremental sequence number is generated one by one through a thread safety mechanism. The present application realizes the generation of a globally unique and strictly incremental sequence number in a multi-instance concurrent environment under the constraint of only using a TDSQL database.

A radar pulse stream sorting method based on wide-range embedding and mask interleaving includes the following steps: Step 1: preprocessing the target template time series signal; segmenting the time series data through a sliding window; Step 2: processing the data to generate simulated data, upgrading the data from the original multivariate time series to a wide-range embedding space as the input of the network; Step 3: masking the wide-range embedding space generated after conversion using the wide-range embedding method with a probability; Step 4: determining the loss function, optimizer, and learning strategy to obtain a network model; Step 5: inferring the model and adding a correction algorithm to the output result, performing sliding window sequence segmentation on the multivariate time series signal data composed of interleaved radar pulse descriptors in the inference stage, and serving as the input of the model inference stage. The present invention realizes high-granularity, sample-level pulse labeling and sorting for high-density interleaved radar pulse sequences.

The present application relates to a wavelet transform-based multifunctional radar state switching point detection method and system, which comprises the following steps: collecting a multifunctional radar pulse signal sequence with multiple working modes, obtaining a single radar emitter pulse signal sequence after feature parameter extraction and pulse sorting; detecting and processing the single radar emitter pulse signal sequence for missing pulses and false pulses and normalizing to obtain a preprocessed single radar pulse signal sequence; detecting the state switching point based on the multi-dimensional parameters of the preprocessed single radar pulse signal sequence to obtain a multi-dimensional parameter mutation point matrix; determining the final mutation point of the single radar emitter pulse signal sequence based on the multi-dimensional parameter mutation point matrix to realize the state switching point detection of the single radar emitter pulse signal sequence and sequence segmentation based thereon. The single working state radar pulse signal sequence obtained by the present application provides data support for working state recognition.

The application discloses a kind of based on noise classification and deep learning's multi-source electromagnetic noise suppression method, belong to geophysical electromagnetic detection technical field.The method includes: original electromagnetic observation sequence is segmented and preprocessed;Low-frequency noise is suppressed using the improved U-Net network of encoder embedding Mamba time series modeling module;The strong noise type in data segment is identified by ROCKET classifier;According to the classification result, the second improved U-Net network trained for the corresponding noise type is called to carry out class-oriented denoising;Finally, complete high-quality data is obtained by splicing data segments.The present application uses the processing framework of "low-frequency pre-suppression-noise classification-class denoising", combines the strong time series modeling capability of Mamba module and the high-efficiency classification performance of ROCKET, significantly improves the suppression accuracy and signal fidelity of complex, multi-source electromagnetic noise, especially suitable for ground and air electromagnetic exploration data processing.

The invention provides a natural language processing method based on a self-adaptive double-memory hybrid architecture, which comprises the following steps of: embedding conversion: acquiring an input text lexical element sequence, converting the input text lexical element sequence into embedding representation through an embedding layer, and providing input for a self-adaptive double-memory hybrid architecture model; complexity evaluation: based on the embedded representation, a sequence complexity evaluator of a dynamic memory layering mechanism is adopted; memory allocation; sequence segmentation; performing short-term processing; middle-stage treatment; long-term treatment; bidirectional fusion; performing forward optimization; outputting a task; precise modeling is kept in a key short-term window, an efficient mechanism is adopted in a medium-and-long-term window, optimal balance of efficiency and performance is achieved, a three-layer memory system is adopted, active information is captured in a short term, semantic association is maintained in a medium term, historical knowledge is compressed for a long term, multi-scale memory organic unification, two-way feedback and self-adaptive fusion are achieved, and the method is suitable for large-scale application. Different mechanisms are complementary to each other and are dynamically coordinated, and optimal configuration is automatically achieved in different task scenes.