405 results about "Sequence Feature" patented technology

The invention discloses a text abstract generation method and system based on sparse attention acceleration, and the method comprises the steps: reading long text data, carrying out word segmentation and embedded coding processing, extracting a sequence feature vector, and mapping the sequence feature vector into a query matrix Q, a key matrix K and a value matrix V; constructing an abstract generation network, wherein the abstract generation network comprises a sparse attention calculation module, a feedforward calculation module, a prediction head module and a key value cache module; inputting the query matrix Q, the key matrix K and the value matrix V into an abstract generation network, passing through a backbone network formed by stacking a sparse attention calculation module and a feedforward calculation module for multiple times, processing by a prediction header module, and caching a historical decoding state in real time through a key value caching module to obtain an initial text feature vector; and based on the initial text feature vector, executing an autoregressive decoding process through an abstract generation network, and outputting a final abstract result. According to the method, the decoding process can be accelerated, and the semantic integrity and coherence of the generated abstract are ensured.

The invention provides a multi-modal characterization molecular property prediction method based on hierarchical bidirectional cross attention, and relates to the technical field of machine learning assisted organic chemistry, and the method comprises the following steps: S10, generating same-molecule multiple sequences for data enhancement; s20, coding the sequence features through a pre-trained molecular language model MolBERT; s30, performing multi-modal feature fusion through a layered bidirectional cross attention mechanism; s40, establishing a prediction head; s50, in the reasoning stage, only the feature extraction and fusion steps are executed, and a molecular property prediction result is output through the trained prediction head. According to the method, the molecular sequence, the topological graph structure and the fingerprint features are effectively integrated, so that the prediction precision of the model on a plurality of MoleculeNet (molecular network benchmark) public data sets is superior to that of an existing method.

The invention belongs to the field of bioinformatics, and relates to an improved integrated deep learning cell communication ligand-receptor interaction prediction method. The method comprises the following steps: firstly, carrying out extraction and dimensionality reduction on biological sequence features of a ligand and a receptor, and constructing multi-modal feature input; secondly, constructing an improved deep neural network branch, introducing a batch normalization layer and a Leaky ReLU activation function, solving the problems of gradient disappearance and neuronal necrosis, and improving regularization strength to prevent overfitting; meanwhile, an enhanced heterogeneous graph auto-encoder branch is constructed, the graph embedding dimension is remarkably expanded to improve the feature capacity, and full convergence of the model is ensured by increasing training rounds; thirdly, fusing the improved deep network with the prediction probability of a heterogeneous graph auto-encoder by adopting a weighted integration strategy; and finally, outputting a potential interaction relationship based on the fusion probability. By optimizing the architecture and the strategy, the prediction accuracy and robustness are remarkably improved, and a reliable tool is provided for analyzing a complex cell communication network.

The invention discloses a diopter detection data fusion processing system based on deep learning, and relates to the technical field of medical detection. The system comprises a multi-modal data acquisition module, a data preprocessing and aligning module, a dynamic sequence analysis module, a deep learning regression core module, a personalized calibration module and a result output and visualization module which are connected in sequence. The system obtains objective detection data through a multi-modal data acquisition module, and multi-modal data is formed after preprocessing and space-time alignment; the dynamic sequence analysis module extracts adjustment response time sequence features; the deep learning regression core module calculates an initial diopter parameter; the personalized calibration module performs parameter optimization in combination with the biological characteristics of the user; and finally, a visual quality simulation result is generated through a result output and visualization module. According to the method, the multi-modal data fusion and deep learning technology is utilized, rapid and accurate diopter automatic detection is realized, and the detection efficiency and the individuation level are remarkably improved.

The invention discloses a large language model battery evaluation method and device based on multi-modal fusion, and relates to the technical field of battery intelligent management. The method comprises the steps that according to time sequence operation data, an improved multi-scale gating time sequence model is used for time sequence feature extraction, and operation time sequence features are obtained; according to the log text data and the monitoring image data, performing semantic feature extraction by using a multi-modal large language model to obtain fused semantic features; performing medium-term fusion reasoning by using an improved large language model according to the operation time sequence features and the fusion semantic features to obtain a battery evaluation result; generating a battery risk detection result and a corresponding explanatory text according to a preset alarm threshold value and a battery evaluation result based on the post-masking time sequence attention flow and the post-masking semantic attention flow; and carrying out risk visualization according to the battery risk detection result and the corresponding explanatory text. The invention relates to a high-reliability battery evaluation method based on a multi-modal large language model.

The invention discloses a vibration event early warning diffusion method based on edge calculation, and relates to the technical field of computer application, and the method comprises the following steps: constructing a topological phase coordinate under a unified time baseline, playing back a historical track of early warning diffusion based on the topological phase coordinate, recognizing a continuous path segment with a closed trend in the track, and carrying out the early warning diffusion. Outputting a loop suspicious region; a fingerprint code of early warning information is constructed on the basis of a loop suspicious area, a three-dimensional feature set containing time sequence features, path sequence features and phase sequence features is extracted, and a corresponding loop suspicion degree spectrum is generated. According to the method, through space-time modeling, path fingerprint construction, causal discrimination and duplicate removal mechanisms, accurate identification and repeated suppression of a vibration early warning information diffusion path are realized, dynamic threshold regulation and reverse intervention are combined, loop closed-loop treatment and topology self-healing are realized, and early warning accuracy, timeliness and system stability are remarkably improved.

The invention relates to the technical field of optical communication, and discloses a multi-modal optical device production line risk identification and prediction method and system, and the method comprises the steps: collecting multi-modal data in real time from each process of an optical device production line, carrying out the preprocessing of the multi-modal data, carrying out the feature extraction, constructing a unified feature matrix, respectively extracting depth time sequence features and depth space features, and carrying out the recognition and prediction of the risk of the multi-modal optical device production line. The depth time sequence features and the spatial features are fused, a risk identification prediction decision model carries out risk prediction through the fused features, process nodes are obtained, target features of the process nodes are screened through an SHAP method, abnormal areas in the target features are positioned through gradient mapping, and an early warning threshold value is dynamically set. According to the method and the device, the technical problems that risk early warning and identification cannot be carried out in time and decisions cannot be taken in time during production of the optical device production line are solved.

The invention discloses an oil chromatography data trend classification method and system based on depth feature enhancement and an attention mechanism, and the method comprises the steps: carrying out numeralization conversion, deletion detection and grouping trend calculation on oil chromatography original gas component data, and generating a basic feature vector; executing multi-scale sliding statistics, change rate and subsequence feature enhancement, and calculating comprehensive similarity and attention weight based on a template library to generate a weighted similarity vector; splicing the enhanced feature and the weighted similarity vector into a time sequence input sequence, and outputting an oil chromatogram trend classification result after attention expansion and long and short term memory network processing. According to the method, structured processing and basic trend extraction of data are realized, adaptive matching and weighted aggregation of historical operation modes are realized, and a multi-dimensional dependency relationship and time sequence dynamic change are captured, so that accurate classification of oil chromatogram trends is realized.

The invention provides an intention recognition response method and system based on forest farmer question and answer data, and the method comprises the steps: firstly obtaining an initial question and answer data set of forest farmer users, the initial question and answer data set comprises a plurality of question and answer interaction records, and then carrying out the semantic standardization processing of the initial question and answer data to obtain a structured question and answer data set; then semantic association features and interaction time sequence features are extracted from the structured question and answer data, the semantic association features reflect the semantic matching degree of user questions and system responses, the interaction time sequence features reflect the context dependency relationship of adjacent question and answer records, and dynamic intention mapping processing is conducted based on the semantic association features and the interaction time sequence features; and finally, according to the target intention recognition result, matching a preset response knowledge base, generating optimized response data, and sending the optimized response data to the forest farmer user terminal to update the response, so that the accuracy of forest farmer question and answer intention recognition and the accuracy of the response are improved.

The invention discloses similar object group expansion model training and similar object group expansion methods and devices, and belongs to the technical field of similar object group expansion, and the training method comprises the steps: obtaining a training set and a target time point corresponding to each training sample in the training set; a positive sample in the training set is a seed user, a target time point corresponding to the positive sample represents a time point of the seed user meeting a preset behavior condition, and a target time point corresponding to the negative sample is consistent with the target time point corresponding to the positive sample in distribution; dividing the user behavior sequence of the training sample before the corresponding target time point into a plurality of subsequences by using different preset time intervals, respectively extracting sequence features from the plurality of subsequences, and generating user features of the training sample based on a splicing result of the sequence features; and training a preset model based on the user features of the training samples to obtain a trained similar object group expansion model. Therefore, the extraction efficiency of the long-term and short-term behavior characteristics can be optimized.

The invention relates to the technical field of biological information, and discloses a messenger RNA stability prediction method, model, device, equipment and medium, the method comprises the following steps: obtaining target sequence information of a target messenger RNA; acquiring at least two of the following target feature information based on the target sequence information by using a feature extraction module in the messenger RNA stability prediction model: first sequence feature information, Kozak sequence feature information, Motif attention feature information and manual feature information; and predicting the stability of the target messenger RNA based on the target feature information by using a prediction head in the messenger RNA stability prediction model. According to the method, the mRNA stability is predicted by fusing the universal sequence feature of the mRNA, the Kozak sequence feature of the learnable position weight, the Motif attention feature based on the hash k-mer and the manual feature, and the accuracy of mRNA stability prediction is improved.

The invention discloses a lightweight multi-mode lower limb motion intention recognition method and system, and relates to the technical field of biomedicine. The method comprises the steps that multi-channel surface electromyogram signals sEMG of a target lower limb and joint angle signals and joint torque signals of lower limbs on the same side are synchronously collected; performing feature extraction by using a double-branch structure to obtain muscle-related deep time sequence features and joint-related high-level semantic features; performing feature fusion through a bidirectional cross attention mechanism to obtain cross attention fusion features; and performing flattening, nonlinear mapping, regularization and Softmax classification on the cross attention fusion features, and outputting the motion intention of the target lower limb. Through double-branch input, depth feature extraction and a bidirectional cross attention mechanism, on the premise of ensuring recognition precision, a lightweight attention module, a residual structure and a cross-modal interaction mechanism are introduced, and the parameter quantity and calculation overhead are reduced.

The invention relates to a tumble risk assessment method and system based on a multi-modal deep learning network, and relates to the technical field of human body posture recognition, and the method comprises the steps: collecting a scene motion image sequence, a plantar pressure feature sequence and personnel basic information; analyzing the scene motion image sequence by using a human body posture estimation network model to determine a human body key feature point sequence; performing feature extraction on the human body key feature point sequence and the scene motion image sequence to generate human body posture sequence features; performing feature extraction fusion on the human body key feature point sequence and the plantar pressure feature sequence to generate human body motion fusion time sequence features; performing feature extraction on the personnel basic information to generate personnel basic information features; and inputting the human body posture sequence features, the human body motion fusion time sequence features and the personnel basic information features into a multi-modal deep learning network for analysis so as to determine a fall risk score. The method and the device have the effect of improving the precision of fall risk assessment.

The invention provides a knowledge graph and text fused railway equipment fault diagnosis method and system, and the method comprises the steps: inputting a natural language fault description, analyzing the natural language fault description into a text, preprocessing the text to obtain a text sequence, inputting the text sequence into a pre-training BERT, extracting a context semantic vector of each word, inputting the context semantic vector into a BiLSTM to capture a context dependency relationship and time sequence features, and carrying out the recognition of the text sequence. Generating a sequence feature vector of fused semantics; the method comprises the steps of constructing a knowledge graph, carrying out token-level embedding fusion on text semantic features and knowledge graph entities and relationships, carrying out sequence relationship modeling by utilizing BiLSTM, calculating weights of fused semantic features by utilizing an attention mechanism, obtaining weighted semantic vectors, executing Softmax classification, and outputting fault category labels and confidence coefficients thereof; and outputting a diagnosis result, reasons and recommended measures, and providing an explanation path of related knowledge nodes. According to the method, the interpretability of the fault diagnosis process and the reusability of knowledge are realized, and the diagnosis accuracy is improved.

The invention discloses a deep learning method for uniform background light and shadow based on a Transform model. The method comprises the steps that an input image and a corresponding portrait main body mask are acquired, and the input image comprises an area with uneven background light and shadow; preprocessing the input image and the portrait main body mask, including image standardization and feature fusion, to obtain a fused feature map; based on a patch segmentation method, converting the fusion feature map into sequence features; encoding the sequence features to output enhanced sequence features, modeling a light and shadow distribution dependency relationship of full image pixels through a global attention mechanism, and distinguishing portrait main body features and background features based on the portrait main body mask; and based on the inverse logic of patch segmentation, recovering the enhanced sequence features into a spatial feature graph and the like. According to the method, semantic mask prior, global context modeling and adaptive residual correction are organically integrated, and a solution is provided for solving the core problem in background light and shadow homogenization.

The invention relates to an enzyme turnover number prediction method and system fused with multi-modal characteristics of an enzymatic reaction, and belongs to the technical field of bioinformatics. Comprising the following steps: respectively constructing an optimized protein pre-training model and an optimized chemical reaction pre-training model; extracting sequence features of a to-be-detected amino acid sequence by utilizing the optimized protein pre-training model to obtain an enzyme sequence feature matrix; extracting sequence features of an enzymatic reaction SMILES sequence to be detected by utilizing the optimized chemical reaction pre-training model to obtain an enzymatic reaction SMILES feature matrix; using molecular fingerprints to represent substrate and product molecular sets, and extracting a set feature matrix of the two molecular sets through a graph attention network; the method comprises the following steps: representing substrate and product molecule sets by using a molecular graph, and extracting internal feature matrixes of molecules in the two molecule sets through a graph isomorphic network; and carrying out feature enhancement on the obtained feature matrix so as to obtain an enzyme turnover number prediction value. The method improves the stability and precision of the prediction result, and has the advantages of low cost and short period.

The invention discloses an Internet of Things anomaly detection method and system based on quaternion state space diffusion enhancement, and belongs to the technical field of network security and artificial intelligence. The method comprises the following steps: mapping a flow time sequence feature into a quaternion tensor to maintain an internal coupling relationship of a multi-dimensional feature; a double-flow encoder is designed, a quaternion selective state space model is adopted to extract continuous fluid features, and a dynamic hypergraph neural network is adopted to model discrete protocol features; carrying out self-supervised pre-training on a resistance pseudo-anomaly sample by utilizing potential diffusion model generation, and optimizing characterization by combining quaternion cepstrum distance loss; the injected learnable prompt vector is optimized in the small sample fine tuning stage, and a category prototype is corrected by using a semi-supervised expectation maximization algorithm; and calculating a sample anomaly score based on an energy model to realize known attack classification and unknown anomaly judgment. According to the method, the generalization ability of the model under the small sample condition and the unknown threat detection ability are remarkably improved.

The invention discloses a GRU-based vehicle-mounted CAN intrusion detection method and system, a storage medium and a computer system. According to the method, an automatic encoder (AE) is introduced to deepen the understanding of a model on input sequence characteristics, a sliding window is used for selecting batch CAN data to be preprocessed to obtain 13-dimensional time sequence data, and a scalar value within the range of [0, 1] is obtained through processing of the encoder, a GRU, a decoder, a full connection layer and a sigmoid activation function and used for classification of abnormal data. The Conv1D is used as a hidden layer, and compared with two-dimensional convolution, the one-dimensional convolution parameter quantity is smaller, and the calculation is simpler and more convenient. An attack message and a normal message can be completely distinguished, the precision and the accuracy rate reach 100%, and the precision in Fuzz detection is 0.9983; compared with the prior art, the method has high accuracy and reliability in the aspect of intrusion behavior detection, can effectively identify most intrusion events, and can keep a relatively low overall error rate, so that good balance between safety and availability is realized.

The invention relates to the technical field of artificial intelligence, and provides a protein function prediction method and device based on multi-modal protein data, and the method comprises the steps: obtaining protein multi-source data, carrying out the feature extraction of a protein sequence in the protein multi-source data, and obtaining a protein sequence feature; constructing a heterogeneous graph based on the protein multi-source data; performing feature coding on the heterogeneous graph by adopting a graph attention mechanism to obtain protein graph features; performing multi-modal fusion on the protein sequence features and the protein map features by adopting a gating fusion mechanism to obtain fusion features; and performing multi-label classification prediction based on the fusion features to obtain a protein function annotation result. The accuracy and robustness of protein function prediction can be improved, and the problems that in the prior art, multi-source protein data cannot be effectively integrated, and the method is sensitive to data noise are solved.

The invention discloses a memory diagnosis method and device, equipment and a medium, which are applied to the technical field of servers, and the method comprises the following steps: acquiring a memory data sequence collected according to a preset time interval within a recent preset duration; memory data features are extracted based on the memory data sequence, and the memory data features comprise change trend features; inputting the memory data features into a target memory state prediction model to obtain a memory state prediction result of a future preset duration corresponding to the memory data sequence; wherein the target memory state prediction model is obtained by training a memory data sequence feature sample and a memory state label, and the memory state label represents a memory state of a future preset duration corresponding to the memory data sequence feature sample. Therefore, the memory fault can be predicted in real time, so that problems can be positioned and solved, and the server performance is improved.

The invention provides a motor intelligent diagnosis method and system and a storage medium, and the method comprises the steps: obtaining the multi-source heterogeneous information and real-time working condition parameters of a motor, converting the multi-source heterogeneous information to a frequency domain to obtain a frequency spectrum tensor, mapping the real-time working condition parameters to a learnable spectrum attention mask, and obtaining a learnable spectrum attention mask; weighting the spectrum tensor by using the spectrum attention mask to obtain a mechanism weighted spectrum; processing the multi-source heterogeneous information and the real-time working condition parameters through physical simulation and a generative model to obtain targeted enhanced data; constructing a hybrid diagnosis network, taking a mechanism weighted frequency spectrum and targeted enhancement data as input, and taking a spectrum attention mask as a bias of a self-attention mechanism to extract global time sequence features; inputting the global time sequence features into a multi-task decoder based on the global time sequence features, outputting fault type probability, fault quantification parameters and a fault feature heat map, and obtaining a diagnosis result based on the fault type probability, the fault quantification parameters and the fault feature heat map. According to the invention, motor diagnosis can be realized more accurately.

The invention discloses a drug-target interaction prediction method and system, and belongs to the technical field of biological information. The method comprises the following steps: firstly, acquiring molecular structure data of a drug and sequence and structure data of a target spot; then, respectively extracting molecular map structure characteristics and SMILES sequence characteristics of the medicine, and amino acid sequence characteristics and three-dimensional space structure characteristics of a target spot; further, taking drugs and targets as nodes, taking the fused multi-modal features as node features, and combining known interaction and similarity information to construct an initial heterogeneous graph; inputting the heterogeneous graph into a dynamic graph neural network, dynamically learning an inter-node connection weight by using a graph attention mechanism, and iteratively updating node representation through multi-layer message transmission to obtain depth feature representation of drugs and targets; and finally, splicing the depth features, inputting the depth features into a multi-layer perceptron classifier, and predicting the drug-target interaction probability. According to the method, through multi-modal feature fusion and dynamic graph structure learning, the prediction accuracy and robustness are remarkably improved.

The invention discloses an anti-noise robust modulation signal identification method and system. The method specifically comprises the following steps: preprocessing an IQ signal to generate an AP feature; inputting the normalized IQ signals and AP features into a GRU guide feature fusion module to generate fusion features; based on the fusion features, channel calibration is carried out through a first-level SEBlock channel attention module, feature extraction and down-collection are carried out through a convolutional network, and coding features are obtained in combination with position coding parameters; the coding features are input to a multi-attention collaborative enhancement module for feature enhancement, and deep time sequence features are output; and inputting the deep time sequence characteristics into an LSTM network for sequence modeling, and outputting a prediction result of the modulation mode of the IQ signal through a classification head. According to the method, high-precision identification of modulation signals in a complex noise environment is realized, and the anti-noise robustness and generalization capability of modulation signal identification are effectively improved.

The invention discloses a ring stationary signal extraction method based on a multi-scale gating attention mechanism and BiLSTM, and belongs to the technical field of mechanical fault diagnosis and signal processing. The objective of the invention is to solve the problem that weak fault features are difficult to accurately extract in a strong background noise and non-Gaussian abnormal interference environment in the prior art. Comprising the steps that a deep neural network model is constructed, local waveforms of different receptive fields are captured through parallel convolution branches, channel masks are generated through an attention mechanism, and hierarchical adaptive suppression of background noise is achieved; biLSTM is embedded in a bottleneck layer, a bidirectional memory unit of the BiLSTM is used for carrying out complete sequence time sequence modeling on compression features, and random abnormal pulses which do not conform to periodic logic are accurately eliminated through a long-term evolution rule of the sequence features; and finally, a decoder is adopted to carry out nonlinear shaping and detail repairing on the up-sampling features after transposed convolution. Gaussian white noise and non-Gaussian outliers can be efficiently filtered out, and pure ring stationary fault impact signals are reconstructed in a high-fidelity mode.

The invention belongs to the technical field of computer biology, and relates to a lightweight drug-target interaction prediction method based on double pre-training language models. Downloading a protein information data set of required drug compounds and drug action targets from a public database; sending the protein sequence into an encoder to obtain a generated vector; the method comprises the following steps: labeling a compound sequence of a drug by using a tokenizer, pre-defining a corresponding relation between a compound sequence label and an integer, then converting into an integer value, and inputting the integer value into an encoder to generate a vector; splicing the class tokens to obtain a drug target pair, and feeding the drug target pair into an interaction head; and calculating an interaction probability between the two. According to the method, the drug pre-training language model and the protein pre-training language model are introduced at the same time, deep semantic coding is performed on the SMILES sequence and the amino acid sequence, molecular structure characteristics and protein sequence characteristics can be comprehensively captured, and the expression ability is remarkably enhanced.

The invention relates to the technical field of cloud computing resource scheduling, in particular to a resource allocation method and device and electronic equipment, and the method comprises the steps: inputting current time sequence feature data and current static feature data into a pre-trained resource demand prediction model, and outputting a predicted overflow probability and a predicted overflow probability confidence interval of the first resource pool within a preset duration, and if it is judged that the first to third resource pools meet a preset resource dynamic allocation condition based on the predicted overflow probability and / or the predicted overflow probability confidence interval, determining a dynamic allocation strategy according to the priority of the current to-be-processed task, and distributing the current to-be-processed task to at least one resource pool for task processing. Therefore, the problems of resource response lag, task processing delay, difficulty in optimizing the overall resource efficiency while guaranteeing the key task and the like due to the fact that a static priority mechanism cannot dynamically adapt to the vehicle type strategic weight and the data timeliness and resource allocation lacks an elastic mechanism in the prior art are solved.

The invention provides a drug target prediction method based on cross-modal attention and uncertainty evaluation, and belongs to the technical field of drug target prediction. In order to solve the technical problems that the existing drug target prediction lacks quantitative evaluation on the reliability of a prediction result and a nonlinear interaction relationship between a drug and a target is difficult to establish, the method comprises the following steps: collecting data, and fusing graph structure features and sequence features of extracted drug molecules to obtain a final code of the drug molecules; extracting amino acid sequence characteristics of the target protein, and constructing protein sequence characteristic expression; inputting the drug molecular features and the protein sequence features into a cross-modal attention module, and aligning and fusing the drug features and the protein features by using a bidirectional cross attention mechanism to obtain drug-target combined feature representation; introducing an uncertainty quantification mechanism, and outputting uncertainty estimation of a drug-target interaction prediction label and a prediction result; the method is used for predicting drug target interaction.

The invention relates to an antifungal peptide recognition method and system based on a multi-graph neural network and a cross attention mechanism. The method comprises the following steps: extracting sequence features from a processed data set by adopting a protein language model, and extracting structural features; taking each amino acid residue as a graph node, establishing an undirected edge, constructing a peptide feature graph, inputting the peptide feature graph into a multi-graph neural network, extracting features from different dimensions through each graph neural network, and generating a graph-level feature vector through a global pooling layer; and calculating a cross-branch weight through a hierarchical cross attention mechanism to realize feature fusion, and inputting the fused features into a binary classifier to obtain an antifungal peptide recognition result. Feature extraction can be carried out from different dimensions by using the multi-graph neural network, and a hierarchical cross attention mechanism is introduced to carry out dynamic fusion on output features of the multi-graph neural network, so that key feature information can be highlighted, and accurate recognition and classification prediction of antifungal peptides can be realized.

The invention requests to protect a rotating machine fault diagnosis method based on multistage attention fusion 2D CNN. According to the method, a multi-level attention fusion mechanism is innovatively introduced for the problems of poor noise robustness and excessive parameters of a traditional convolutional neural network in rotating machine fault diagnosis. The method comprises the following steps: firstly, converting a signal from a time domain to a frequency domain by adopting fast Fourier transform, and analyzing the frequency domain characteristics of the signal; according to the method, vibration signals are decomposed into intrinsic mode functions through variational mode decomposition, the characteristics of the signals in different frequency ranges are better reflected, the transformation results of the intrinsic mode functions and the intrinsic mode functions are spliced to serve as multi-channel input of a model, the noise influence is reduced, and the noise robustness of initial characteristic extraction is enhanced. Secondly, feature extraction is carried out through channel attention and a global attention mechanism, the channel attention mechanism identifies key features, and the global attention mechanism captures a long-range dependency relationship; then, heterogeneous sequence features are projected to a unified space through an improved cross attention mechanism for efficient fusion, and the calculation overhead is reduced. And finally, inputting the fusion features into a full-connection layer for classification. According to the method, high diagnosis accuracy can be kept under different noise intensities.

The invention provides a construction method and application of a well site real-time data pre-training model, and belongs to the crossing field of petroleum engineering and artificial intelligence. Carrying out discrete marking; the method comprises the following steps: constructing a self-supervised time sequence model based on a Transform architecture, and fusing position coding and an embedding mechanism; designing a self-supervised task, and setting multiple task targets such as mask prediction, comparative learning, sequence judgment and future state prediction; performing end-to-end training on the model through a joint loss function, and extracting deep time sequence features; deploying the pre-training model in an actual well site environment, and executing tasks such as anomaly detection, working condition recognition and trend prediction; and carrying out model optimization and iteration. The invention further provides application of the method in the field of intelligent monitoring of oil and gas drilling engineering. According to the method, the problems of high dependence on manual labeling, weak model generalization ability, insufficient real-time adaptability and the like of a traditional method are solved, and the method has good engineering deployment and cross-well-site popularization value.