2524 results about "Activation function" patented technology

The invention discloses a remote sensing image semantic segmentation method based on Mama and Transform architecture fusion. The method comprises the following steps: S1, data preprocessing; s2, pre-training the model by using the MAE; s3, preliminary feature extraction: performing preliminary feature extraction on the image subjected to MAE processing through linear projection, a convolutional layer and a nonlinear activation function to obtain a low-dimensional feature map; s4, CFF feature fusion: fusing the low-dimensional feature map and the image after MAE processing by using a CFF module; s5, feature enhancement of a Mamb-T module: inputting the feature map of the fused image into the Mamb-T module for operation to obtain further feature enhancement of the image; s6, multi-scale feature extraction; s7, a segmentation module, wherein the feature map is input to a Mask2Former segmentation module for final image segmentation; and S8, outputting and post-processing. According to the method, local and global features in the image can be captured, and the segmentation problem under a small water body and a complex background is effectively solved through combination of a self-adaptive attention mechanism and convolution operation.

The invention provides an intelligent incubation bin anti-interference control method and system based on error self-learning, and relates to the technical field of intelligent control, and the method comprises the steps: respectively calculating a temperature deviation value and a humidity deviation value according to an optimization parameter sequence of a temperature prediction error and an optimization parameter sequence of a humidity prediction error; performing normalized weighted summation on the temperature deviation value and the humidity deviation value, and performing dynamic scaling through a Gaussian kernel function to obtain a correction factor; dynamically adjusting the activation function slope of a neural network hidden layer according to the correction factor and a BP neural network, reconstructing the weight of the fuzzy rule base based on the numerical distribution characteristics of the correction factor, and generating the corrected weight of the fuzzy rule base; and based on the corrected fuzzy rule base weight, constructing a three-dimensional parameter adjustment curved surface, and dynamically adjusting proportion, integral and differential parameters of a PID controller through a curved surface gradient search algorithm to generate a control signal. According to the invention, the control accuracy is improved.

The invention discloses a water quality prediction deep learning model and prediction method based on spatio-temporal feature fusion and LSTM, and the model comprises one or more time convolution networks which comprise a plurality of TemporalBlock modules, and one or more classic convolution networks which comprise one or more convolution layers, a pooling layer, and a full connection layer. A Xavier initialization method is adopted to initialize parameters, a feature fusion mechanism is provided, and each convolution layer comprises a convolution operation, a ReLU activation function, a Dropout layer and a pooling layer; at least one additional temporal convolutional network; and a long-short-term memory network, a feature fusion module and an output module which flexibly adjust the output into single-step prediction and multi-step prediction and are used for generating a final prediction result. According to the method, the time sequence dependence and the nonlinear relation in the water quality monitoring data can be effectively processed, and the deep learning model of the multi-source information can be flexibly integrated, so that the accurate prediction of the water quality change is realized.

The invention discloses a frequency domain enhancement fusion network method for night infrared and visible light images, and relates to the technical field of image processing. According to the method, a coarse enhancement network (CENet) is constructed, the night visible light feature extraction capability is improved through deep convolution and point-by-point convolution, a channel importance score is calculated in combination with SENet, finally, a pixel-level enhancement image is generated by using a Tanh activation function, the enhanced visible light feature image and an infrared light image are respectively sent to FDFNet, and the pixel-level enhancement image is obtained. In FDFNet, a U-shaped feature extractor is constructed based on EFA and D3FN modules of a frequency domain, visible light significant information is injected into an infrared feature map through SENet, so that the feature learning ability of a U-Net network is enhanced, finally, images passing through the FDFNet are spliced, a final fused image is obtained through image reconstruction, and the fusion effect is improved. The method provided by the invention is superior to the most advanced method in the aspect of LLVIP data set, shows the potential of further improving the fusion image precision in the future, and is helpful to promote the progress of advanced visual tasks.

The invention discloses a rotary machine fault diagnosis method based on residual contraction convolution and an attention mechanism, and belongs to the field of rotary machine fault diagnosis. According to the method, firstly, a soft threshold filtering activation function is designed based on Meta-ACON, a soft threshold filtering module is developed around the activation function, and the module can adaptively learn a filtering threshold and a forward dynamic weighting coefficient, so that irrelevant features are effectively filtered, and the expression ability of effective features is enhanced. And then, in combination with multi-channel and multi-scale residual convolution, a multi-channel residual shrinkage convolution unit is constructed to realize multi-scale and multi-channel effective feature extraction. Besides, according to the method, a high-order statistical modeling idea is utilized, a GSoP attention mechanism is introduced behind a multi-channel residual shrinkage convolution unit, and the ability of extracting discriminative features is enhanced through second-order statistical information of a high-level channel feature map. Test results show that the method has good fault identification capability and generalization capability.

The invention discloses a water quality prediction method based on a Transform-LSTM fusion model, and belongs to the technical field of water quality time series data prediction and artificial intelligence. Comprising the following steps: (1) acquiring water quality data from a water quality monitoring station; (2) carrying out pretreatment; (3) screening out water quality characteristic data; (4) dividing into a training set, a verification set and a test set; (5) inputting the data into a Transform-LSTM (Long Short Term Memory) fusion model; (6) embedding water quality data time sequence information by a Transformer encoder through position coding, extracting a global dependency relationship among features by utilizing a multi-head attention mechanism, and optimizing gradient propagation by combining residual connection and layer normalization; (7) the LSTM layer receives the high-order features after Transform coding, and captures a local time sequence dynamic mode; and (8) mapping the extracted water quality time sequence characteristics to a specific prediction result by a regression output layer by adopting a linear activation function, and calculating an evaluation index. According to the method, the water quality change trend of the surface water body can be effectively predicted, and powerful support is provided for water ecological protection and sustainable development.

An electronic circuit system implementing and executing machine learning inference engines. While ML inference engines are based on (architectures and parameters defined by) configured, trained and tuned machine learning models, our design has the novel ability to support data driven, on-the-fly-reconfigured model runs. Reconfiguration and tuning operations include dynamic computational graph modifications, define-by-run alterations, changes to network depth (number of layers) and width (neurons per layer), and adjustments to weights, biases, plus activation function parameters. Neural networks supported include Feed-Forward, RNN, CNN, and Hopfield architectures, plus Ensemble, Federated, Cooperating, Adversarial, and Swarm collections. Decision Trees and Forests are also supported, as are more esoteric approaches such as ART and KAN. Our invention is capable of running both standalone and cooperatively, the cooperative processing being local and / or remote / cloud based, interfacing with telemetry applications to feed data, and machine learning software to feed new or updated models.

The invention discloses a multi-modal sentiment analysis model construction method based on text-driven cross-modal fusion and mutual information estimation, and the method comprises the steps: extracting text features, audio features and video features of a to-be-analyzed video, and constructing a feature extraction module with context time sequence information; on the basis of a cross-modal attention mechanism, other modal features can perceive text emotion information, and a stacked text driving cross-modal fusion module is constructed; constructing a mutual information estimation module for performing mutual information upper and lower bound estimation on the fusion feature vector in order to reserve task related information as much as possible and filter noise information irrelevant to the task; constructing a loss function module in combination with the mutual information estimation module and a main task loss function; and constructing a prediction module for inputting the fusion feature vector into an activation function for emotion polarity judgment. According to the multi-modal sentiment analysis model constructed by the method, comprehensive experiments are carried out on two public and available data sets, and experiment results show that the sentiment analysis performance of the model is remarkably improved.

The invention discloses a remote sensing target detection method and system based on selective feature space fusion, and belongs to the field of remote sensing target detection of computer vision, and the method is characterized in that firstly, multi-scale features, extracted by a backbone network, of an input remote sensing target image are subjected to mapping processing through a convolution layer and a group normalization layer; the method comprises the following steps: obtaining feature maps of different stages, then carrying out spatial dimension alignment and channel splicing operation on the feature maps of different stages, refining spliced features by using a channel attention mechanism, and carrying out pooling, large kernel convolution and activation function processing on the refined features to obtain a selective feature space attention weight; according to the method, the target detection capability in remote sensing target detection is improved, so that the problems of diversity of remote sensing images, large target size and complexity are solved.

The invention relates to the technical field of mine safety analysis, and discloses a mine personnel trajectory prediction method, system and device based on artificial intelligence. The method comprises the following steps: collecting mine personnel positioning data, and performing standardization processing; calculating trajectory quality through a neural network; converting the spatio-temporal feature vector to perform image recognition; performing multi-time scale prediction by applying an activation function; matching with a mine space structure, and clustering to identify a group behavior mode; and calculating a safety index, and generating a safety scheduling strategy. According to the method, the future trajectory of the mine personnel can be accurately predicted through the special environment constraint of the mine and the working behavior mode of the miner, and the prediction result is converted into a safe scheduling strategy.

Provided herein are systems, methods, computer-readable media, and techniques for providing trusted artificial intelligence (AI) using Fully Homomorphic Encryption (FHE), comprising: (A) providing a deep neural network (DNN)-based model with modified architecture, wherein the modified architecture at least (i) uses a Gaussian function as an activation function and (ii) removes one or more pooling layers; (B) obtaining encrypted data, wherein the encrypted data are generated by applying the FHE to plaintext data; and (C) generating an inference with the DNN-based model based on the encrypted data. Further provided herein are systems, methods, computer-readable media, and techniques for providing trusted AI using stochastic computing, using noise based computing, using an artificial immune system, and with secure multi-party computation (SMPC) based at least in part on watermarking.

The invention discloses a visual language multi-modal fusion method based on parameter-free cross attention, and belongs to the field of computer vision. The implementation method comprises the following steps: using a fixed pre-training language model as a trunk, using a visual encoder to extract image features, and calculating a cross attention weight between language query and visual features through a parameter-free activation function, replacing a plurality of groups of learnable projection matrixes introduced by a traditional cross attention module, and significantly reducing the model parameter scale. A multi-scale visual feature generation mechanism based on pooling operation is introduced, and rich visual semantic prompt information is provided for a language model. A dynamic feature selection module is designed in combination with cross attention, visual areas corresponding to all text tokens are screened, low-correlation areas are discarded, only visual content more contributing to the current language context is reserved, accurate information matching and efficient fusion between modals are achieved, and the accuracy and efficiency of information fusion are improved. And the performance of the visual language model in tasks such as image-text question answering, image generation and multi-modal instruction understanding is improved.

The invention discloses an industrial time sequence prediction method based on adaptive continuous learning. The method comprises the following steps: firstly, dividing a non-stationary industrial time series data set to obtain a plurality of domains with the maximum distribution difference; different time domains are then modeled in sequence, and an improved empirical playback (DER + +) method is used to avoid catastrophic forgetting of previously accumulated knowledge. Meanwhile, a soft sample buffer area is introduced to promote memory and learning of key modes in the current field. And finally, the time-sensitive activation function TimeRelu enables the time convolutional network (TCN) to have a time evolution property, and the generalization ability of the prediction model is enhanced. According to the method, the continuous learning normal form is introduced into the time sequence prediction task, the limitations of huge resource overhead of traditional cumulative training, disastrous forgetting of an incremental learning mode and the like are overcome, and the method has theoretical and practical significance on industrial time sequence prediction.

The invention discloses a pipeline anomaly detection method based on VQSAE and DCGCN, and the method comprises the steps: 1, collecting the time series data of a sensor on a pipeline, constructing a data set, and carrying out the preprocessing of the data set; 2, constructing an improved VQSAE model, and introducing a new activation function SELU to carry out local time feature extraction on the data; step 3, constructing a DCGCN global space-time relation model according to features extracted by the VQSAE model, and introducing a graph attention mechanism GAM to perform space-time graph convolution operation; and 4, detecting the pipeline anomaly through the DCGCN model, and outputting the pipeline anomaly detection accuracy. According to the method, data are fused through the VQSAE and the DCGCN, complex pipeline topology and multi-modal data can be processed, the false alarm rate and the missing report rate are reduced, more accurate, efficient, high-robustness and extensible pipeline anomaly detection is achieved, managers are helped to recognize potential pipeline risks earlier, maintenance decisions are optimized, and the pipeline anomaly detection efficiency is improved. And economic loss and potential safety hazards caused by faults are reduced.

The invention discloses an intelligent question and answer method and system based on semantic vector optimization and dynamic prompt, and the method comprises the steps: inputting a target question into a trained question and answer model, and obtaining an answer corresponding to the target question; wherein the trained question and answer model comprises an input layer, an embedded layer, a semantic vector generation module, a feature compression module, a dynamic prompt generation module, an answer generation module, a linear layer and an activation function layer which are connected in sequence; wherein the output end of the feature compression module is further connected with the input end of the RAG retrieval module, and the output end of the RAG retrieval module is connected with the input end of the dynamic prompt generation module; according to the trained question and answer model, the training process comprises the steps that a semantic vector generation module is constructed, and the semantic vector generation module is pre-trained; and setting the pre-trained semantic vector generation module in the question and answer model, training the question and answer model, and when the total loss function value does not decrease any more, stopping training to obtain a trained question and answer model.

The invention relates to an AI computing subsystem, end-side equipment and a control method of the end-side equipment, and the AI computing subsystem comprises at least one FLASH storage chip which is used for pre-storing weight parameters of at least one AI network model; the at least one DRAM storage chip is used for storing weight parameters and activation functions of the AI model; and the logic chip is connected with the FLASH storage chip and the DRAM storage chip, responds to an external instruction, reads data of the FLASH storage chip so as to write the data into the DRAM storage chip, and / or reads weight parameters and activation functions stored in the DRAM storage chip and executes a calculation task. The SoC system only needs to send parameters or start commands of the AI network, and the whole network runs in the AI computing subsystem; and the most complex part of the AI network part can be deployed in the AI computing subsystem, and the SoC system only needs to compute the simple operator of the part, so that the SoC system has larger space to complete other responsible tasks.

A device implementing a nonlinear activation function including a material having nonlinear susceptibility phase-matching a coherent nonlinear interaction involving a signal comprising a signal wavelength and a bias comprising a bias wavelength, so that (1) a first phase difference between the signal and the bias induces the interaction comprising second harmonic generation (generating a second harmonic of the bias wavelength) or sum frequency generation (generating a sum frequency of the bias and the signal, and (2) a second phase difference between the signal and the bias induces the interaction comprising parametric amplification amplifying the bias and attenuating the signal. A positive input to the nonlinear activation function is represented by the signal having an input energy and the first phase difference. A negative input is represented by second phase difference. The output is an output energy of the signal as function of the input energy.

The invention discloses an infrared and visible light fusion method based on multi-feature extraction. A multi-feature extraction network is composed of a fusion network, a semantic segmentation network, a content loss function and a semantic loss function. The fusion network comprises a feature extraction part and a feature reconstruction part, and the feature extraction part comprises a convolutional layer with an activation function, an adaptive gradient feature extraction module and a space channel reconstruction convolution and residual gradient extraction module; the feature reconstruction part is composed of an attention mechanism and four convolution layers connected in series; the method comprises the following steps: S1, inputting an infrared image and a visible light image into a fusion network, and generating a fusion image through feature extraction and feature reconstruction; s2, transmitting the fused image to a segmentation network for segmentation; and S3, comparing the segmentation result with a semantic tag, and evaluating richness of semantic information contained in the fused image. The contour of the fused image is clearer, the target is clearer, and the index is remarkably improved.

The invention relates to an adaptive feature weighted fusion fault diagnosis method for multi-source heterogeneous data of a motor, and the method comprises the steps: constructing a multi-head deep convolutional neural network, designing a plurality of parallel convolutional channels in the neural network, deeply mining the abnormal features of original signals of current and vibration, and focusing on the signal features in different dimensions, thereby achieving the fault diagnosis of the multi-source heterogeneous data of the motor. And comprehensive capture of key feature information is ensured. And adaptive weighted fusion is performed on the extracted multi-dimensional features by adopting a correlation analysis method to highlight abnormal key features, so that the feature fusion effect of the multi-source heterogeneous data and the accuracy of anomaly detection are improved. An SELU activation function is adopted and a hyper-parameter optimization strategy of a dynamic attenuation learning rate is combined so that an overfitting problem which is easy to occur in a model training process can be effectively solved and convergence stability of the model and high precision of fault identification can be enhanced. Through adaptive fusion of multi-source heterogeneous data features, key fault features are effectively extracted, and the diagnosis accuracy in an electromechanical coupling fault scene is improved.

The invention relates to a wellbore multiphase flow model numerical solution and gas-liquid distribution state inversion method and system, and belongs to the technical field of petroleum engineering, and the method comprises the steps: 1, constructing and training a physical information neural network for drilling wellbore multiphase flow dynamic simulation and overflow gas distribution state inversion; determining input and output of the physical information neural network; determining a loss function of the physical information neural network; training a physical information neural network; 2, designing an adaptive optimization algorithm, optimizing the final solution precision and convergence speed of the physical information neural network, and obtaining an adaptive physical information neural network; designing an adaptive activation function; designing a self-adaptive sampling mechanism based on residual errors; 3, based on the self-adaptive physical information neural network, numerical solution and gas-liquid distribution state inversion of the shaft multiphase flow model are achieved. According to the method, the problem that a traditional numerical method usually needs high-precision grid division and a large number of computing resources is effectively solved.

The present application relates to the technical field of damaged codes, and specifically relates to a method for recognizing a damaged quick-response code. The method comprises: first acquiring a sample data set, inputting the sample data set into a CenterNet target detection model to be trained, and training the CenterNet target detection model; and then acquiring a target damaged quick-response code to be recognized, and recognizing the target damaged quick-response code by means of the trained CenterNet target detection model, so as to acquire a damage rate of the target damaged quick-response code. In the solution, a target damaged quick-response code is recognized by using a trained CenterNet target detection model, thereby reducing the amount of and complexity of calculation; and the CenterNet target detection model is optimized by means of an activation function, and the performance of the trained CenterNet target detection model is then evaluated, thereby improving the accuracy and efficiency of damage rate recognition results.

The invention provides an energy storage system power grid allocation method based on a dynamic space-time diagram convolutional neural network. The energy storage management method comprises the following steps: S1, acquiring time sequence data of a power grid node; s2, sequentially executing a graph convolution layer operation and a time convolution operation by using a dynamic space-time graph convolution neural network, and extracting space-time features of the time sequence data; s3, calculating attention weighted features of the spatio-temporal features; s4, processing the new energy output uncertainty based on a fuzzy rule by using the spatial-temporal characteristics and the attention weighted characteristics; s5, realizing distributed model training based on a federated learning framework, and dynamically updating model parameters through an attention mechanism; s6, integrating the geographic positions and the connection relations of the nodes, and outputting enhanced features; and S7, the enhanced features pass through a full connection layer and an activation function, a power grid load prediction result is output, and a charging and discharging strategy of the energy storage system is dynamically regulated and controlled. Power supply and demand dynamic balance in a new energy high-permeability scene is realized, and the operation cost and the environmental influence are reduced.

The invention discloses a social robot identification method and system based on deep learning, and relates to the technical field of robots, and the method comprises the steps: carrying out the smooth processing of a tweet through a large language model, and generating a tweet fused with expression semantics in combination with a natural language processing model; capturing an emotion expression difference between a robot account and a real user; generating a comprehensive feature vector of global context sensing; outputting fusion features; the output fusion features are mapped to a high-dimensional space through a linear layer, and the detection probability of a robot account is output through an activation function; and based on adaptive moment estimation, gradient back propagation is carried out by using a cross entropy loss function, and parameters of the graph convolutional network model are optimized to obtain a final classification result. According to the method, the problems of sparse text expression and emotion information loss are effectively relieved, and the separability of the social robot and the real user in emotion behavior modes is enhanced.

A method for recognizing a motor imagery electroencephalography (MI-EEG) signal based on a capsule network (CapsNet) is provided, and relates to the technical field of deep learning and brain-computer interfaces (BCIs). An electroencephalography (EEG) time series is mapped into a three-dimensional (3D) array form based on a spatial electrode distribution. By combining 3D convolution, a CapsNet constructs a three-dimensional capsule network (3D-CapsNet) for recognizing an MI-EEG signal. A 3D convolution module performs feature extraction from both a temporal dimension and an inter-channel spatial dimension through a plurality of layers of 3D convolution to obtain a low-level feature. The low-level feature output by the 3D convolution module is integrated through the CapsNet to obtain a high-level spatial vector containing an inter-feature relationship. A primary capsule and a motor capsule are connected through dynamic routing, and finally a CapsNet module outputs a classification result through a nonlinear activation function squash.

The invention relates to the technical field of energy storage frequency modulation, in particular to a prediction method and system for an energy storage frequency modulation instruction, and the method comprises the steps: decomposing a frequency modulation instruction sequence Pt through VMD, and obtaining a plurality of IMF subsequences; according to the total area Sii of the current IMF sub-sequence and the test group error Wi, judging whether the current IMF sub-sequence is a false sequence or not by using an activation function; and improving the IMF sub-sequence of which the judgment result is false, and performing prediction output of a frequency modulation instruction by utilizing the improved IMF sub-sequence. The method has the advantages that false sequences possibly generated in the variational mode decomposition process can be effectively recognized and improved. According to the method, the false sequence can be accurately found out and processed by calculating the area defined by the adjacent data points in the subsequences and the X axis, the prediction error is reduced, and the prediction precision is improved. In addition, a false sequence is optimized by means of generating a random sequence, performing mirror image transformation and the like, and the response speed and the frequency modulation efficiency of the system are further improved.

The invention discloses a road defect detection method based on an improved RT-DETR-R18 model. A backbone network adopts a CSPNet architecture. According to the invention, innovative improvement is carried out on an original C2f module, and Bottleneck in the original C2f module is replaced by DynamicIncMixerBlock to form a C2fDCMB module. The DynamicIncMixerBlock is characterized in that a DynamicIncMixerBlock is fused with a DynamicInceptionMixer component and a ConvolutionalGLU component, and the DynamicIncMixerBlock and the ConvolutionalGLU component are fused with each other. A dynamic Inception deep convolution structure is adopted by the DynamicInception Mixer, and features of different scales and directions are adaptively captured through dynamic kernel weight distribution; according to the method, AIFI (intra-scale feature interaction) in a high-efficiency hybrid encoder is improved, a module is combined with an EfficentAdditiveAttach and a feedforward network structure to form TransformerEncoder LayerEfficentAdditiveAttach, a RepC3 module is replaced by a RetBlockC3 module in a cross-scale feature fusion module (CCFM) through a multi-head attention mechanism and a nonlinear activation function, the RetBlockC3 is improved based on the RepC3, RetBlock and RelPos2d are introduced, and the RetBlock and RelPos2d are introduced into the RetBlockC3 module to form a multi-scale feature fusion module. According to the method, the accuracy, recall rate and detection speed of road defect detection are obviously superior to those of a traditional detection model, and a more accurate and efficient technical solution is provided for maintenance of road infrastructures and traffic safety guarantee.

The invention discloses a liquid nitrogen preservation effect evaluation method and system based on big data. The method comprises the steps of data acquisition, data preprocessing, establishment of a liquid nitrogen preservation effect evaluation model and liquid nitrogen preservation effect evaluation. The invention belongs to the field of effect evaluation, and particularly relates to a liquid nitrogen preservation effect evaluation method and system based on big data, according to the scheme, a dynamic activation function is designed by introducing learnable parameters, and the prediction stability is improved; a regularization item is constructed based on the module weight so as to design a loss function, and the prediction accuracy of liquid nitrogen preservation effect evaluation is improved; through gradual updating module by module, the capability of capturing subtle changes of liquid nitrogen preservation effect evaluation data is enhanced; and the liquid nitrogen preservation effect evaluation efficiency is improved based on secondary dynamic optimization correction parameters.

The invention discloses a lightweight remote sensing target detection method and system, and relates to the technical field of target detection, and the method comprises the steps: obtaining a to-be-detected target remote sensing image; and inputting the target remote sensing image into the trained improved YOLOv8n network model, and performing target detection on the target remote sensing image through the improved YOLOv8n network model to obtain a target detection result. According to the method, the lightweight MobileNetV3 network is adopted as a backbone network, and the reasoning speed is increased by using a deep separable convolution and efficient residual connection structure; through a nonlinear activation function and linear projection in residual connection, the feature expression capability is improved; a multi-scale attention expansion mechanism is introduced in front of a detection head, the receptive field of the model is expanded while the calculation cost is reduced and the processing efficiency is improved, and finally, a DWR module is introduced, feature fusion is optimized, and the detection capability of small targets and shielded targets in remote sensing images is improved.

The invention belongs to the technical field of artificial intelligence information security, particularly discloses a neural network privacy reasoning method and device based on winograd sparse convolution and differentiable activation function approximation, and aims at solving the problems that existing neural network privacy reasoning is high in calculation time delay and large in communication cost. And in combination with a multi-granularity weight pruning algorithm, the sparsity of a linear convolutional layer is maximized, and the communication cost required by the convolutional layer with dense multiplication is reduced. Meanwhile, collaborative optimization is carried out on a nonlinear activation function, a new activation function sensitivity and importance quantitative evaluation algorithm is designed, and an adaptive learnable polynomial function is introduced, so that the fitting precision of the activation function is improved, and the polynomial replacement loss is reduced; and the powerful prediction performance of the neural network model is reserved by using a multi-stage knowledge distillation method, and the calculation and communication efficiency of privacy protection neural network reasoning is effectively improved on the premise of not influencing the prediction precision.

The invention discloses an aero-engine missing value filling method based on multi-scale time sequence and attribute feature extraction. Firstly, for data collected by each sensor of an aero-engine, a temporary replacement value of a missing value is generated by adopting linear interpolation; extracting a plurality of significant cycles of the aero-engine sequence data as a time scale, decomposing the sequence into trend terms and season terms according to the significant scale, and splicing all the trend terms and season terms in the time dimension; then, in-scale and inter-scale time sequence dependence is captured in a time dimension by using cross-scale graph convolution so as to carry out missing value filling in time sequence. In the attribute dimension, local neighborhood information is accumulated by adopting a cross-attribute graph volume, global attribute features are calculated by utilizing a relational network, node features are updated, and attribute association is supplemented on the basis of a time sequence filling result. And finally, aggregating the node features of a specific scale through a maximum pooling layer to represent the features, and calculating by adopting two linear layers and an activation function to obtain a final filling result.