172 results about "Frequency analysis" patented technology

The invention belongs to the field of artificial intelligence, particularly relates to a method for predicting the residual life of a key part of a coal mill, and aims to solve the problems of low prediction precision and poor extrapolation caused by working condition disturbance interference and inaccurate degradation characterization in a traditional method. The method comprises the following steps: collecting vibration, temperature, current and acoustic emission multi-source synchronous data; constructing a dynamic working condition decoupling model of fusion of the variational auto-encoder and the attention mechanism, and separating degradation sensitive components; and a high-fidelity degradation index sequence is generated through fusion of time-frequency analysis and a gating circulation unit. According to the method, multi-modal sensing and physical priori knowledge are fused, the monotonicity of a degradation index is remarkably improved to 98% or above, the extrapolation error of an extreme working condition is reduced by 40%, edge-cloud collaborative deployment and digital twinborn visualization are supported, multi-component collaborative early warning is realized, and the early warning amount of non-planned shutdown is improved from 7 days to 21 days or above; and a high-precision, strong-robustness and landing life prediction solution is provided for intelligent operation and maintenance.

The invention belongs to the field of non-intrusive power load monitoring, and discloses a non-intrusive power load identification method, which comprises the following steps of: constructing a load identification framework utilizing frequency domain characteristics, and aims to solve the problem that transient characteristics cannot be effectively applied in non-linear load identification in non-intrusive load monitoring. A current feature extraction technology based on fast Fourier transform and Hilbert-Huang transform and a classification method based on an extreme learning machine as a main body are used for training and testing a public data set to verify that the method is used for extracting transient load features and identifying nonlinear loads.

The invention discloses an ant colony optimization energy focusing cable insulation damage positioning method and device, and relates to the technical field of cable detection engineering, and the method comprises the steps: generating a Gaussian envelope linear frequency modulation chirp detection signal, injecting the detection signal into a to-be-detected cable, and collecting an incident signal and an echo signal containing a reflection component; constructing a parameterized time-frequency analysis kernel function family under a Cohen type time-frequency distribution framework, and searching and optimizing kernel parameters by using an ant colony optimization algorithm; and calculating the time frequency distribution and energy-time curve of the signal based on the energy focusing time frequency analysis kernel function obtained through optimization, extracting the arrival time corresponding to an energy peak value, and realizing cable insulation damage positioning according to the reflection time delay and the propagation wave velocity. According to the method, energy peak focusing and peak position stability are enhanced while cross terms are restrained, defect positioning precision and robustness are remarkably improved, and the method is suitable for engineering online detection and automatic implementation of cable insulation damage.

The invention relates to the field of numerical control machining data processing, in particular to a multi-axis linkage numerical control lathe machining data processing method and system. Comprising the following steps: acquiring vibration information, performing time-frequency analysis, and extracting energy change and instantaneous peak value characteristics in a preset high-frequency band; the features are compared with a preset judgment basis, an early sign of flutter is recognized, and when flutter is recognized, a machining parameter adjusting instruction for the rotating speed of the main shaft, the feeding speed and / or the cutting depth is generated; and the adjusted vibration information is obtained for time-frequency analysis, the flutter suppression effect is evaluated, machining parameters are further adjusted according to the flutter suppression effect, and / or an early warning or shutdown alarm is given out, so that closed-loop active suppression control over microcosmic cutting flutter is formed. The problems that due to the fact that a multi-axis linkage numerical control lathe cannot capture key and instantaneous high-frequency vibration mode information in a data collection and analysis frame, tool abrasion prediction is not accurate, the machining quality is reduced, the workpiece rejection rate is increased, and the production efficiency is reduced are solved.

This invention discloses a multi-scale time-frequency analysis method and system for ocean current meter signal processing, relating to the field of marine information processing. The method includes: S1: acquiring raw data and performing physical benchmark correction to obtain three-dimensional current velocity data and basic mass labels; S2: extracting first and second features from the three-dimensional current velocity data through spatiotemporal decoupling; S3: fusing and separating the three-dimensional current velocity data based on the first and second features to obtain an optimized set of flow field parameters, and generating the uncertainty of each parameter by combining the basic mass labels and error source information; S4: performing comprehensive quality assessment and grading based on the first and second features and the uncertainty, outputting a data product with mass labels and uncertainty information. By extracting spatiotemporal dual-dimensional features and using them to guide signal fusion and quality assessment, highly robust processing of ocean current data and end-to-end quality traceability are achieved, significantly improving the reliability of the data product.

This invention discloses an inversion method for S-wave velocity structure based on dense micromotion observations, belonging to the field of geophysical exploration and engineering geological exploration technology. The method includes: acquiring micromotion data; performing segmented preprocessing and cross-correlation calculations; constructing a phase consistency metric function based on the instantaneous phase information of the cross-correlation function for each time period; generating adaptive weighting coefficients; and then constructing an empirical Green's function through weighted superposition. Subsequently, time-frequency analysis is performed to obtain a time-frequency energy map; preliminary extraction of candidate dispersion velocities for each frequency is performed, and their comprehensive confidence level is calculated; after verification, a reliable dispersion curve is obtained. Based on the dispersion curve, a model parameter vector is established, and a joint objective function is constructed. A differential evolution algorithm is used for parallel global optimization inversion to obtain a one-dimensional S-wave velocity structure model below each station pair. The one-dimensional models are interpolated and fused to generate a two-dimensional S-wave velocity structure profile. This invention has the advantages of high inversion stability, high computational efficiency, and a high degree of automation.

The invention relates to the technical field of roadbed engineering monitoring, in particular to a roadbed engineering settlement deformation monitoring device and method, and the device comprises an optical fiber sensor network, a data collection module, a signal processing module, a model matching module, a frequency analysis module, a data fusion module, an early warning trigger module and a prediction report module. The optical fiber sensor network is deployed in a roadbed to collect an optical signal delay sequence, fluctuation characteristics are extracted through data collection and signal processing, settlement acceleration risks are judged through model matching, the speed trend is determined through frequency analysis, historical and real-time data are integrated through data fusion, an alarm is triggered when an early warning module exceeds a threshold value, and a risk assessment report is generated through a prediction module. According to the device, roadbed three-dimensional global monitoring and full-process automation are achieved, monitoring blind areas are eliminated, early warning lag is shortened, and accurate support is provided for roadbed stability operation and maintenance.

The invention discloses a sparse inversion time-frequency analysis method based on a semi-norm, and the method comprises the steps: designing a window function for original seismic data, so as to reduce the influence of a window truncation effect on time-frequency analysis; constructing a Fourier forward matrix and a windowed Fourier matrix according to the size of the window function and the size of the frequency spectrum; in-window seismic data extraction is carried out on sample points of single-channel seismic data, and a semi-norm constrained time-frequency spectrum sparse inversion objective function is constructed; performing time-frequency spectrum coefficient inversion on the time-frequency spectrum sparse inversion target function to obtain a sparse inversion time-frequency spectrum result of the seismic data in the window; and sliding the sample points of the seismic data to repeat the process, and performing time-frequency spectrum sparse inversion on all the seismic data to obtain a final time-frequency analysis result. According to the method, the time-frequency spectrum precision and the time-frequency focusing property of the seismic signal can be effectively improved, and a time-frequency analysis result with higher precision is provided for subsequent reservoir identification and oil and gas prediction work.

The invention discloses an SAR intermittent sampling forwarding interference suppression method and system based on pulse compression and time-frequency analysis, and a storage medium, and the method comprises the following steps: carrying out the range pulse compression of SAR echoes containing interference pulse by pulse, so as to accurately obtain an interference-containing pulse sequence; and then, carrying out short-time Fourier transform on the pulse with interference after pulse compression and the nearest pulse without interference, and carrying out superposition and differential operation of two different dimensions to realize time-frequency domain positioning of the 2D-ISRJ. And designing a time-frequency domain filter, and carrying out time-frequency domain wave trapping on the positioned 2D-ISRJ. And finally, recovering the missing part of the signal after wave trapping by using the time-frequency data containing the interference pulse and the nearest non-interference pulse after pulse compression. According to the invention, the interference detection precision of typical two-dimensional intermittent sampling forwarding interference forwarding types can be relatively high under the condition of relatively weak interference, and the suppression method not only can achieve a relatively good suppression effect, but also can better protect useful signals.

Embodiments of the present disclosure provide a centrifugal compressor fault diagnosis method, device and equipment. The method comprises: acquiring a current vibration signal of a centrifugal compressor to be diagnosed; extracting a current fault characteristic parameter of the current vibration signal based on a linear transformation time-frequency analysis method; acquiring a pre-stored fault characteristic sample library; and performing analogy analysis on the current fault characteristic parameter and the fault characteristic sample library to determine a fault type of the centrifugal compressor. In this way, the current fault characteristic parameter and the fault characteristic sample library can be automatically subjected to analogy analysis, so as to accurately and quickly locate the fault type of the centrifugal compressor. Since this fault analysis method does not require model training, the fault analysis result is not affected by the model precision, and the algorithm is simple to use, so the identification accuracy of the fault type of the centrifugal compressor can be ensured.

An electric power industrial control system equipment anomaly detection method based on protocol use frequency analysis belongs to the technical field of electric power industrial control system security, and comprises the following steps: firstly, extracting the use frequency of various communication protocols in a system, and carrying out anomaly judgment on the use frequency of the protocols at each moment; then, calculating a protocol use history characteristic value of the corresponding time period of the previous day, taking the characteristic value as a reference, and judging the abnormal state of the protocol in the time period by comparing the protocol use characteristics of the day and the history; and finally, starting from the overall perspective of the equipment, constructing an equipment anomaly feature vector according to various protocol state distributions contained in the equipment, calculating a comprehensive anomaly score of the equipment according to the equipment anomaly feature vector, and dividing equipment operation state grades, thereby effectively improving the accuracy and real-time performance of equipment anomaly detection of the electric power industrial control system.

The invention discloses a notebook computer fault trend prediction system based on time sequence modeling, and the system comprises a period marking module which is used for constructing a life cycle feature data set, carrying out the clustering operation through an improved HDBSCAN algorithm, and constructing a life cycle state tag; the trend prediction module is used for outputting a standard trend prediction result through the improved DRN model; the state control module is used for executing state change frequency analysis and generating a state smoothing factor; the burst identification module is used for constructing a sudden change signal channel and outputting a burst trend prediction result; and the trend correction module is used for constructing an abnormal memory cache region, generating an abnormal early warning bias vector and outputting a corrected trend prediction result. According to the method, the fine-grained identification capability of trend prediction and the response sensitivity of abnormal change are effectively improved, and continuous, stable and accurate modeling and prediction of the potential fault trend in the full life cycle of the notebook computer are realized.

The application provides a method for trajectory midcourse target recognition based on a low-data-rate HRRP sequence, and relates to the technical field of radar signal processing.The method obtains target signals through steps one and two, and inputs the HRRP sequence obtained by observing the target through a radar sensor as sample data; the method extracts features of the target through steps three and four: the method establishes a precession model by taking a cone-cylinder target as an example, uses the periodicity of scattering center points in the HRRP sequence to perform time-frequency analysis, and realizes one-dimensional feature distribution of energy strength by using a projection method, so as to extract a precession frequency; and the method extracts target structure features, entropy value features and period features; the method performs signal feature fusion processing through step five, classifies the features by performing feature importance analysis and using a Random Forest algorithm, obtains the category of a target to be recognized, and outputs a final fusion result.

A bearing damage detection system includes a rotation sensor that detects rotation of a rolling bearing and outputs a rotation signal, a rotation fluctuation extraction section that extracts rotation fluctuation of the rolling bearing from the rotation signal and generates a rotation fluctuation signal, a frequency analysis section that performs frequency analysis on a waveform of the rotation fluctuation signal and derives a frequency characteristic, a peak intensity calculation section that derives a peak intensity corresponding to bearing damage from the frequency characteristic, and a damage detection section that detects bearing damage of the rolling bearing based on the peak intensity. The rotation fluctuation extraction section derives a waveform of the rotation fluctuation signal from which various noises are removed.

The acoustic signal detection method comprises the following steps: generating a transmitting signal of a constant envelope system according to a signal modulation and time-frequency analysis principle; processing the received underwater acoustic signal data on the basis of beam forming and matched filtering theories to generate a matched related azimuth history graph; executing open peak retrieval in the matched related azimuth history graph, and extracting a peak point set corresponding to a suspected target; according to the azimuth and the time position of the peak point, extracting a time-frequency characteristic pattern of each peak point from the original data; and processing the time-frequency feature map by using a pre-trained dual-channel neural network model to determine a final detection result. The method can improve the accuracy of echo signal detection.

The purpose of the present disclosure is to provide a frequency analysis device capable of determining a frequency with a higher accuracy than a frequency resolution. A frequency analysis device according to the present disclosure divides a signal wave in the time domain into a plurality of sections at predetermined time intervals, and performs discrete Fourier transform on each of the plurality of sections into which the signal wave has been divided. Furthermore, the frequency analysis device calculates a frequency spectrum by combining the signal waves subjected to the discrete Fourier transform in the plurality of sections, and determines a provisionally determined frequency from the frequency spectrum. Furthermore, the frequency analysis device calculates a section phase difference, which is a phase difference between the signal waves in adjacent sections, for at least one set of the plurality of sections. The frequency analysis device corrects the provisionally determined frequency using the section phase difference.

A method for controlling an OIS system of an image capturing device, the method comprising: obtaining a first and a second time-series of samples derived from a motion signal of a motion sensor of the image capturing device, wherein the motion signal indicates a motion of the image capturing device, and wherein the first time-series spans a longer time period of the motion signal than the second time-series; determining a current dominant frequency for the motion signal based on a frequency analysis of the first time-series; and determining a setpoint for the OIS system by: fitting to the second time-series a model function based on a sine and / or cosine function and the current dominant frequency, extrapolating the fitted model function past a last sample of the second time-series to predict a successive sample, and determining the setpoint based on the predicted successive sample.

The present disclosure relates to the technical field of geological exploration, and particularly relates to a multi-scale stratigraphic division method, device, equipment, storage medium and computer program. The method comprises the following steps: acquiring well logging data, correcting the well logging data to obtain corrected data; performing empirical wavelet decomposition on the corrected data to obtain an empirical wavelet decomposition result, and performing time-frequency analysis on the empirical wavelet decomposition result to obtain a corresponding time-frequency graph; performing feature extraction on the time-frequency graph to obtain a time-frequency feature, and constructing a frequency spectrum combination feature according to the time-frequency feature; and dividing a scale stratum according to the frequency spectrum combination feature to obtain a division result. The present application can adapt to the division of multi-scale strata.

The invention belongs to the technical field of electric energy quality monitoring, and particularly discloses an electric energy quality disturbance identification method and device based on multi-modal feature fusion. According to the invention, the method comprises the steps: obtaining a time-frequency feature outputted by a target time-frequency analysis model; carrying out feature fusion on the initial multi-modal features and the time-frequency features based on a multi-branch multi-modal feature fusion network; and determining a power quality disturbance identification result of the current power signal according to the target fusion feature. Through the above mode, the target time-frequency analysis model composed of the improved adaptive transformation function and the target window function is utilized to analyze and output the time-frequency characteristics, so that the time-frequency focusing capability of the disturbance signal can be improved, and accurate positioning of the disturbance characteristics is ensured. And then feature fusion is carried out through a multi-modal feature fusion network of a multi-branch parallel structure, differential information among all modals is effectively mined, and then an electric energy quality disturbance identification result of the current electric energy signal is identified, so that the accuracy of electric energy quality disturbance identification can be effectively improved.

The application relates to the field of financial text verification, in particular to a financial knowledge base construction method and system based on a large model, which comprises text conversion, obtaining a first semantic unit with a corrected term and a determined second semantic unit, establishing a scene-concealed meaning double-layer correction model with the first semantic unit as input, outputting the first semantic unit with determined semantics and a scene label, and determining unique definition; the term in the first semantic unit with a to-be-corrected term does not have a definition in a scene definition mapping library, a target word mapping is established, and explicit semantics are mapped based on semantic similarity; and each semantic unit is combined according to the original text position. The application binds unique definition for polysemous terms, realizes explicitness of the concealed meaning terms without definition in the scene definition mapping library through semantic similarity and co-occurrence frequency analysis, effectively eliminates the ambiguity of financial semantics and the difficult problem of concealed meaning interpretation, and avoids knowledge base data deviation or even errors caused by semantic confusion.

The application provides a kind of dry reactor sound signal visual detection method, medium and system, belong to dry reactor sound signal detection technical field, include: first, the sound signal of dry reactor is collected, and pretreatment is carried out to eliminate environmental noise.Then the time-frequency analysis is carried out to the sound signal after pretreatment, and the time-frequency spectrum is obtained.Next, adopt the way of bayesian probability inference and adaptive threshold increase, and highlight the small change in time-frequency spectrum, and obtain the increased time-frequency spectrum.Subsequently, energy distribution, peak frequency and harmonic structure are extracted from the increased time-frequency spectrum, and combined into a multi-dimensional feature vector.Apply dimension reduction algorithm, map high-dimensional feature vector to two-dimensional or three-dimensional space, and obtain the second feature vector.Finally, use unsupervised learning algorithm to carry out cluster analysis on the second feature vector, and assign color or label according to the clustering result for different categories, generate visual classification image output.

The present application relates to the technical field of jitter analysis, in particular to a high-speed oscilloscope eye diagram real-time jitter detection method and system, comprising the following steps: high-speed acquisition waveform zero-crossing detection obtains timestamp calculation offset, grouping obtains cross-period edge time sequence, according to sequence hierarchical clustering merges jump cluster, calculates deviation, obtains drift trend vector, based on trend vector, assigns compensation weight, weightedly corrects time coordinate, calls corrected coordinate to detect period boundary, calculates difference, triggers correction when difference exceeds threshold, obtains period-corrected edge coordinate, and according to correction coordinate, Fourier analysis is carried out to extract jitter frequency and amplitude to generate eye diagram jitter detection result, in the present application, phase correlation is established based on cross-period sequence, and a resolvable offset mode is formed by identifying the distribution structure through clustering, the cumulative deviation is suppressed through drift trend and weight correction, the abnormal period is calibrated through period dynamic discrimination, and the jitter structure and amplitude relationship are separated through frequency analysis, so that the overall stability and integrity are strengthened.

Method and systems are described that obtain or collect time-series vibration data related to operation of a downhole drilling system and convert the time-series vibration data to frequency domain vibration data. The frequency domain vibration data can be evaluated to determine one or more vibrational resonance frequencies as well as corresponding non-aliased harmonic frequencies (if and when present) and aliased harmonic frequencies (if and when present) in the frequency domain vibration data. Change in the one or more vibrational resonance frequencies in the frequency domain vibration data over time can be monitored during the drilling operations to detect or predict the occurrence of a defect or failure or other unwanted conditions (e.g., HFTO) in the drilling system. A health alert can be generated in response to the detection or prediction of the occurrence of a defect or failure or other unwanted condition in the drilling system.

The invention relates to the technical field of artificial intelligence and distributed machine learning, in particular to a model decoupling federated learning method and system based on frequency analysis. When the model is trained, firstly, an image classification model deployed by a client is divided into a public feature layer and a personalized classification layer; carrying out aggregation training on the public feature layer by adopting a federal learning method; during local training, a client carries out segmented training on a public feature layer and a personalized classification layer of an image classification model on a local data set {an image sample and an identification result}. According to the method, the adaptation capability of the model to the local heterogeneous data is improved through a personalized mechanism of model decoupling. Meanwhile, due to the adoption of parameter decoupling, the client only uploads the parameters of the public feature layer, and the personalized classification layer is reserved locally. An attacker can only obtain incomplete model information, the difficulty of reversely deducing user privacy data from model parameters is increased, and privacy protection is enhanced.

The invention relates to the technical field of noise suppression, and discloses a self-adaptive multi-agent noise suppression system and method based on event triggering. The system comprises an event detection module, a self-adaptive noise suppression module, a multi-agent collaboration module and a dynamic feedback module. The event detection module monitors a noise signal through a time-frequency analysis algorithm and generates an event sequence according to a trigger condition; the adaptive noise suppression module updates parameters through an adaptive filter model and an online learning algorithm according to the sequence adjustment strategy; the multi-agent collaboration module deploys nodes to execute local tasks, and collaboratively optimizes the local tasks through a distributed communication protocol; the dynamic feedback module collects suppression effect data, adjusts trigger conditions and filter parameters, and forms closed-loop control. In addition, a signal preprocessing module is arranged to reduce the calculation complexity, and an exception handling mechanism is provided to guarantee the reliability. The method can accurately suppress noise, improves the system efficiency and stability, and is suitable for various noise suppression scenes.

The invention discloses an event camera calibration method based on event stream time-frequency analysis. The method comprises the following steps: S1, establishing an event camera calibration system; comprising an event camera and an LCD display screen. S2, acquiring an event stream; reconstructing an event stream output by the event camera into a 2D stripe intensity image by using a time-frequency analysis algorithm; s3, extracting feature points; a multi-step phase shift method is used to calculate wrapped phases, wrapped phase diagrams in the horizontal direction and the vertical direction are obtained respectively, a horizontal straight line group and a vertical straight line group are fitted according to the wrapped phase diagrams, and intersection points are feature points; s4, moving the LCD display screen back and forth or at different angles to obtain data of at least three poses, repeating the steps S1 to S3, and respectively capturing horizontal and vertical stripe images of different poses; and S5, according to the image information obtained through processing, working out world coordinates corresponding to the feature points, then obtaining internal and external parameters of the event camera based on the feature points and the world coordinates thereof, and finally optimizing the obtained parameter values.

The invention relates to the technical field of image processing and deep learning, in particular to an automatic surface wave inversion method and system, and the method comprises the steps: receiving original seismic information, sequentially carrying out the mean value removal, trend removal and spectrum whitening preprocessing of surface wave information, calculating a single-component noise cross-correlation function and a nine-component noise cross-correlation function based on the preprocessed data, and obtaining a single-component noise cross-correlation function and a nine-component noise cross-correlation function; and based on a time-frequency analysis method, calculating a surface wave group velocity and a velocity frequency dispersion energy diagram through the cross-correlation function, automatically extracting a frequency dispersion curve from the velocity frequency dispersion energy diagram by using an artificial intelligence frequency dispersion extraction method, screening frequency dispersion files of which effective frequency dispersion points and signal-to-noise ratios meet a preset standard, and calculating the surface wave group velocity and the velocity frequency dispersion energy diagram through the cross-correlation function. And inverting the parameters of the three-dimensional shear wave velocity model based on the underground interface constraint three-dimensional shear wave velocity inversion method to obtain a three-dimensional shear wave velocity model. The problems that in the prior art, efficiency is low, the operation process is complex, and the professional threshold is high are solved.

This detection system comprises: a detection device having a pulse wave sensor and an acceleration sensor; a data processing device capable of communicating with the detection device; and a management device that performs notification according to information received from the data processing device. The pulse wave sensor and the acceleration sensor alternately repeat a measurement state and a standby state. The data processing device has a pulse wave data processing unit that calculates an acceleration pulse wave from a pulse wave, a body movement data processing unit that performs frequency analysis of acceleration, and a determination unit that determines whether a person wearing the detection device is suspected of having developed sepsis. When at least one of a first condition and a second condition is satisfied, the determination unit transmits, to the management device, information indicating that there is a suspicion of sepsis. The first condition is that the spectral intensity in a specific frequency band among the frequency components of the acceleration obtained during one measurement state is equal to or greater than a predetermined threshold. The second condition is that the average value of acceleration pulse wave amplitudes of the acceleration pulse wave calculated from the pulse wave obtained during one measurement state is equal to or greater than a predetermined prescribed value.

The invention provides an aero-engine strong non-stationary cross frequency signal time-frequency analysis method and system, and relates to the technical field of aero-engine signal processing and fault diagnosis, and the method comprises the steps: obtaining a vibration signal, and obtaining initial time-frequency distribution; performing adaptive ridge line extraction on the initial time-frequency distribution, and determining a time-frequency block range of an intersection region; carrying out local sub-blocking on the time frequency distribution in the intersection area to obtain optimized time frequency distribution; performing component reconstruction on the vibration signal by adopting Vold-Kalman order filtering to obtain final time frequency distribution; according to the method, the ridge line fracture time-frequency block is determined through bidirectional retrieval repair, the intersection area and the hierarchical cascade reconstruction component are optimized through comprehensive evaluation indexes, the ridge line continuity and instantaneous frequency extraction precision of the intersection area are remarkably improved, and high-resolution representation and accurate separation of strong non-stationary intersection frequency signals are achieved.

The invention discloses a stamping die wear prediction method based on multi-sensor fusion, and relates to the technical field of die state monitoring. Comprising the steps that S1, multi-source sensing data in the operation process of a stamping die are collected, and the multi-source sensing data comprise a vibration signal, a temperature signal, a pressure signal and an acoustic emission signal; s2, performing data preprocessing on the acquired multi-source sensing data, wherein the data preprocessing comprises denoising, dimension unification and data standardization on each path of sensing signals; through adaptive time sequence alignment, the problems of feature extraction deviation and context information loss caused by difficulty in accurate alignment of a multi-source signal time sequence in an existing method are solved; time domain statistics, frequency domain energy, time frequency analysis and deep time sequence characteristics are extracted at the same time, and the deep coupling relationship of the heterogenous data is mined through a cross-modal attention and gating mechanism of the multi-modal fusion network, so that the limitation of simple weighted fusion is avoided.