313 results about "Feature combination" patented technology

The invention provides a coral reef remote sensing image multi-modal feature generation and restoration method and system, and relates to the technical field of remote sensing image restoration, and the method comprises the steps: obtaining coral reef image data based on a multi-source sensor, and carrying out the preprocessing; performing image morphological feature and spectral feature extraction on the preprocessed coral reef image data by using a convolutional neural network, and establishing a feature database; carrying out weighted fusion on the multi-source features based on a feature combination network, generating comprehensive feature expressions, and storing the comprehensive feature expressions into a feature database; the coral reef image to be restored is matched with the feature database, and guidance parameters are generated; and inputting the guidance parameters into the generative adversarial network, and performing pixel-level reconstruction on the missing region of the coral reef image to be restored. According to the method, a complete closed-loop system is constructed, multi-source information scheduling, pixel-level guide reconstruction and semantic feedback verification are covered, image information can be supplemented, ecological information can be reasonably reconstructed, and a high-quality data basis is provided for subsequent classification, monitoring and protection work.

The invention provides a target tracking method and system based on global and local two-way extraction and asynchronous enhancement, and the method comprises the steps: carrying out the initialization of a template image and a search image, and carrying out the sequence division through block embedding; extracting global features and local features of the template image and the search image at the same time, and performing feature fusion on the global features and the local features of the template image and the search image; the output of the global and local joint module is input into an asynchronous enhancement module for asynchronous interaction and feature enhancement; and inputting the finally enhanced output feature to a prediction head to obtain a tracking result. According to the method, global and local features are extracted at the same time through the feature combination module, and the features after interaction are enhanced by using asynchronous enhancement operation in the feature fusion stage, so that the robustness and the feature expression ability of the model are improved.

The invention relates to the field of nondestructive testing and automatic sorting of agricultural products, and discloses an X-ray-based walnut internal defect feature optimization detection method and device, and the method comprises the steps: collecting an X-ray image of a moving walnut, extracting a region of interest (ROI) through preprocessing, extracting and fusing spatial domain and frequency domain texture features to form a multi-dimensional feature set, and obtaining a feature set; an optimal feature combination is screened out through dimensionality reduction; and finally, the types of the internal defects of the walnuts are identified by a pre-trained classifier, and a result is output. The device comprises a feeding mechanism, a conveying mechanism, an X-ray detection mechanism, a sorting execution mechanism and a processing and control system electrically connected with all the mechanisms. The problems of single identification category and insufficient model stability and precision are solved, and high-throughput and automatic lossless sorting of walnuts is realized.

The invention discloses a terrain classification method based on a random forest algorithm, a server and a storage medium, and belongs to the technical field of terrain classification, and the method comprises the steps: obtaining DEM data and remote sensing image data of a research area, and extracting terrain features, spectral features, index features and texture features; designing a feature combination scheme, and generating an optimal feature combination data set as input data of terrain classification; calculating an optimal parameter combination of a random forest algorithm, and generating an optimal random forest classifier; establishing a terrain classification system suitable for the research area, and constructing a terrain classification training sample set and a verification sample set; performing terrain classification by using a random forest classifier and the optimal feature combination data set to obtain a terrain classification result of the research area; and calculating and evaluating the precision of a terrain classification result by utilizing a verification sample set and a Kappa coefficient evaluation method. By adopting the method, the terrain classification refinement degree and the terrain classification calculation efficiency and classification efficiency can be improved.

The invention relates to the technical field of low-altitude aircraft chip safety monitoring, and discloses a low-altitude aircraft chip safety monitoring system which comprises a state acquisition module, a risk assessment module, an anomaly detection module, a strategy generation module, an instruction execution module, an effect verification module, a threshold setting module, a data output module, a feedback optimization module and a system regulation module. The state acquisition module analyzes the operation parameters and the environment interference to obtain a state value; the risk assessment module extracts and screens abnormal protection feature combinations; the anomaly detection module combines a sample variation matching feature with an evaluation combination; the strategy generation module maps a feature strategy based on real-time data and sets a rule; the instruction execution module captures feature data and analyzes a trend; the effect verification module adjusts rules according to errors; the threshold setting module determines a monitoring threshold interval; the data output module generates standardized data; the system realizes comprehensive monitoring of the chip state through cooperation of multiple modules, and the operation safety and stability are improved.

The invention relates to the field of traditional Chinese medicine pulse condition collection, and provides a traditional Chinese medicine six-channel identification cognition method and system based on heart rate variability, and the method comprises the steps: obtaining heart rate variability data of a detected object, and extracting parameters such as time frequency from the heart rate variability data; constructing a hierarchical feature extraction network, inputting parameters such as time frequency into the hierarchical feature extraction network, and distributing weights for the parameters such as time frequency based on a qi-blood-body fluid theory to obtain a fusion feature vector; inputting the fused feature vector into a particle swarm optimization algorithm, and performing feature selection by adopting a six-channel transmission constraint function and a syndrome affinity particle update strategy to obtain an optimized feature combination; the optimized feature combination is converted through the semantic mapping relation between the heart rate variability parameters and the pulse condition descriptors, and pulse condition feature parameters are obtained; and performing syndrome classification calculation based on the pulse condition characteristic parameters, and outputting six-channel syndrome types. According to the invention, automatic identification conversion from physiological signals to traditional Chinese medicine syndromes is realized, and the precision and reliability of six-channel syndrome identification are improved.

The invention discloses a large language model robustness visual diagnosis method based on multi-dimensional features and adversarial attacks. The method aims to break through a mode that traditional evaluation only depends on a single aggregation index, and a multi-dimensional text feature exploration system covering vocabularies, syntax, semantics and a structural layer is constructed, and a large-scale antagonism disturbance mechanism and a task self-adaptive quantification strategy are combined. And generating structured feature-adversarial instruction-robustness diagnosis data comprising the cue word to be evaluated and the corpus. On the basis, an interactive visual analysis system is constructed, and through bidirectional linkage of a feature statistical view and a semantic projection view, a user is supported to realize progressive exploration from macroscopic feature screening to microscopic semantic attribution under the double view angles of cue words and corpora, so that a root cause causing the fragility of the model is deeply diagnosed. According to the method, the key feature combination influencing the stability of the model can be identified, so that a basis is provided for directional optimization of the model, and the diagnosis depth of robustness evaluation is improved.

The embodiment of the invention provides an intelligent monitoring method and system based on multi-source heterogeneous data fusion. The intelligent monitoring method based on multi-source heterogeneous data fusion comprises the steps that production data sets in different production scenes are collected and stored in a classified mode; mining association features of equipment operation behavior records and material circulation track information through semantic association analysis, and extracting influence features of production environment perception data in combination with environmental factor mapping to obtain multi-source qualitative features; generating a node relation chain according to the business process association relation, and configuring a matching algorithm of qualitative scene features; screening feature subsets conforming to association rules, and combining and associating the feature subsets to obtain a scene feature combination; creating a scene exception reasoning model by using the scene feature combination and performing exception feature evolution analysis to obtain an exception reasoning result; and the abnormal propagation path is traced, the root cause node is positioned, and the abnormal monitoring label is generated, so that the accuracy of abnormal root cause positioning and the scene adaptability are improved.

The invention discloses a hydrological forecasting method based on multi-feature combination and a Transform model. The method comprises the following steps: acquiring measured data of a target watershed hydrological station, building a physical hydrological model, acquiring output data and derivative feature data of the physical hydrological model, and integrating to obtain basic hydrological data; creating enhanced hydrological physical features, and forming a multi-dimensional original feature pool; constructing a plurality of combination strategies based on the basic hydrological data and the multi-dimensional original feature pool; capturing a long-term dependency relationship of the hydrological time sequence by using an improved Transform model; and the model performance is improved through automatic hyper-parameter optimization. According to the method, the influence of different input feature combinations on the flood forecasting precision is highlighted, and effective technical support is provided for water resource management and flood control and disaster reduction.

The invention provides a converter valve key component burning defect identification method based on a neural network, and belongs to the technical field of power electronic equipment fault diagnosis, and the method comprises the steps: collecting multi-modal data through an infrared thermal imager and a plurality of sensors, and inputting the pre-processed multi-modal data into a specially designed neural network architecture; the framework comprises an image feature extraction module, a time sequence feature extraction module, a feature fusion module and a classification positioning module. According to image processing, improved ResNet50 is combined with an attention mechanism, time sequence features are extracted through a bidirectional long-short-term memory network and a time convolution network, and effective feature combination is achieved through a dynamic weight fusion mechanism. Meanwhile, a double-phase heat conduction model is introduced to analyze temperature distribution, accurate identification of burning defects of key components of the converter valve in a complex environment is realized through large-scale data set training and a two-stage optimization strategy, and key technical support is provided for safe operation of a power system.

The invention discloses a conveying system and method based on AI commodity recognition, and relates to the technical field of conveying recognition, and the system comprises a collection module which is used for collecting image data and sensor data of a to-be-conveyed commodity in conveying equipment in real time; the AI feature combination module is used for carrying out feature extraction on the image data of the to-be-transmitted commodity and the sensor data based on an AI model, and combining the extracted features to obtain a first feature; the commodity recognition module is used for inputting the first feature into a pre-trained commodity recognition model for recognition and determining a recognition result; the adjusting module is used for adjusting the transmission information of the to-be-transmitted commodity based on the identification result; the cooperation mechanism of the conveying mechanism equipment and the recognition system is optimized, and the conveying efficiency of the conveying mechanism equipment is improved.

The invention discloses a real-time network intrusion detection method based on a genetic algorithm and a bidirectional long-short term memory network, and belongs to the technical field of network security, and the method comprises the following steps: 1, extracting continuous and category network traffic features from a constructed data set, and carrying out the feature preprocessing and preliminary screening; step 2, performing feature selection optimization based on a genetic algorithm so as to screen out an optimal feature combination with high accuracy and low dimension; 3, adopting a bidirectional LSTM algorithm to construct an abnormal traffic detection model based on the feature subset selected by the genetic algorithm, wherein the abnormal traffic detection model is used for identifying normal and abnormal samples in the network traffic; 4, evaluating the performance of the abnormal traffic detection model by adopting the confusion matrix; and step 5, deployment and real-time detection of an abnormal flow detection model. According to the invention, a lightweight and traceable intrusion detection framework is constructed. The generalization ability and precision of the detection model are improved; the method gives consideration to accuracy, interpretability and system response capability.

The invention discloses a new energy vehicle modeling design method and system, electronic equipment and a storage medium, and belongs to the technical field of vehicle industry design, and the method comprises the steps: carrying out the weight calculation and sorting of user perceptual vocabularies obtained through an evaluation construction map based on an interval type 2 trapezoidal fuzzy Kano model, so as to recognize key perceptual demands; establishing a nonlinear mapping relation between the key perceptual vocabularies and the form design characteristics of the new energy automobile based on a limit gradient lifting model optimized by a He-Marx optimization algorithm; predicting and generating an optimal morphological feature combination with the highest perceptual evaluation value by using the mapping relation; and performing subjective and objective comprehensive evaluation on the generated design scheme in combination with an eye movement tracking technology and a discrete information data fluctuation weighting method, and screening out a final design scheme. According to the method, subjective and fuzzy perceptual requirements of the user can be accurately quantified and sorted, a nonlinear mapping relation between the perceptual requirements of the user and specific product design characteristics is established, and the modeling design efficiency of the new energy automobile is improved.

The invention relates to the technical field of health risk prediction, in particular to a metabonomics-radiomics prediction method for chronic subdural hematoma recurrence risk, which comprises the following steps: acquiring a CT image and extracting edge gray fluctuation, constructing fluctuation parameters in combination with metabolome data, screening coordination characteristics to generate a risk combination, and predicting a chronic subdural hematoma recurrence risk. Feature pairs consistent in trend are extracted to form a collaborative channel, and a feature matrix is constructed to generate an input vector set; according to the method, disturbance features are extracted through a CT image edge gray level path, a cross-modal fluctuation trend comparison mechanism is established in combination with patient brain metabolism indexes, biological consistency between the features is enhanced, feature combinations with uncoordinated changes are eliminated, feature pairs with collaborative structure and function trends are screened, and a linkage path is constructed. The evolution relation from structural disturbance to metabolic response is reflected, channel data sorting and recombination improve the difference of input characteristics, the stability and accuracy of recurrence discrimination are enhanced, and the systematicness and interpretability of risk assessment are improved.

The invention relates to the technical field of risk control scene feature development, in particular to a multi-source data intelligent feature development system based on large language model driving, which comprises a data preprocessing module, a feature generation module and a feature screening optimization module, the data preprocessing module realizes automatic processing and multi-language support of unstructured data; the feature generation module generates complex features and advanced semantic features according to business scene requirements by using a large language model, so that the feature dimension and depth are greatly expanded; and the feature screening optimization module performs multi-stage filtering on the high-dimensional candidate features, and determines an optimal feature subset through feature combination optimization. The system can intelligently and efficiently extract high-quality features from multi-source data, significantly improve the data processing efficiency and the quality of model features, enhance the adaptability to a new business scene, reduce the labor cost of feature engineering, and have good expansibility and application value.

The invention discloses a double-branch neural network dynamic RCS sequence classification method and system based on time-frequency feature combination. The method comprises the following steps: S1, carrying out target modeling and simulating static RCS data; s2, motion parameters are initialized, orbit data are generated, and dynamic RCS data are generated in combination with the attitude angle sequence; s3, preprocessing the dynamic RCS data; and S4, constructing a double-branch neural network model based on time-frequency feature combination, inputting the data obtained in the step S3 into the double-branch neural network model for training and storing optimal parameters, and performing dynamic RCS target classification by using the optimal parameters. Through dynamic RCS data generation and double-branch deep learning model construction, high-precision classification of complex moving targets is realized, and the method has the characteristics of light weight, high efficiency and high real-time performance.

The invention provides a multi-scene-oriented commodity style feature extraction and same style retrieval method and system, a storage medium and a program product. The method comprises the following steps: acquiring a plurality of images of a to-be-identified commodity from at least two different observation perspectives; respectively extracting a corresponding feature combination from each image; on the basis of the commodity type and the feature combination under each view angle, the feature matching number of the to-be-identified commodity and candidate commodities in a commodity library is calculated, and an initial matching sequence under multiple view angles is obtained; fusing the plurality of initial matching sequences to generate a comprehensive matching sequence; and outputting the comprehensive matching sequence as a final recognition result. The method is simpler and lower in cost under the conditions of fixed business scenes, infringement risk control and selection based on clear attributes, and the recognition effect and practicability of the method are far better than those of a large model scheme needing regular training in the aspects of avoiding the long tail effect and realizing real-time updating.

The invention discloses a rehabilitation evaluation system and method based on data feedback, and the method comprises the steps: obtaining a continuous online rehabilitation video as rehabilitation training video frame data, and constructing a rehabilitation training motion data set; extracting joint angle features and skeleton point distance features from the action data set by adopting a label classification technology, and obtaining distance measurement of each human body joint skeleton point coordinate; matching the time sequence with the action features by adopting a dynamic adjustment algorithm for the action feature combination, and constructing a classification evaluation model of rehabilitation training actions; a data feedback mechanism is introduced into the classification evaluation model of the rehabilitation training actions, the relation between rehabilitation requirements and rehabilitation evaluation is mapped, the whole system achieves detailed analysis and dynamic adaptation of the rehabilitation training actions, accurate rehabilitation suggestions can be provided according to individual differences, the effectiveness of a treatment strategy is ensured through real-time monitoring and adjustment, and the rehabilitation training effect is improved. And the rehabilitation effect and efficiency are greatly improved.

The invention discloses a cross-field feed conversion ratio analysis method based on deep learning. The method comprises the following steps: accessing historical data, and calculating individual and group FCR data; the method comprises the following steps: performing feature extraction on data elements influencing FCR in multi-field domain data, constructing feature vectors, quantifying single features and contribution degree of interaction between the features to the FCR based on a Sobol index, and then identifying key features and combining the features according to gradient weights of the key features by taking the features as input and the FCR as output through a deep learning model; using a clustering algorithm to use a data element set corresponding to the plurality of feature combinations for representing the overall influence importance on the FCR; by taking the element set as a unit, generating a strategy set and a corresponding strategy set weight after assignment, and solving an optimal production strategy by adopting a non-dominated sorting genetic algorithm; according to the method, data of different pig farms are analyzed, and a universal rule and personalized features for a specific field area can be identified.

An apparatus for processing an information signal has: a feature extractor for extracting a set of features having a first dimension; a feature segmenter for segmenting into a first subset having a second dimension and a second subset having a third dimension, which overlap, both being lower than the first dimension; a neural network processor for processing the first and second subsets using a first and a second neural network to obtain a first and a second result, respectively; a feature combiner for combining the first and second results using a third neural network, having a third complexity lower than a first or a second complexity of the first and second neural network to obtain a result set of features having a result dimension; and an output post-processor for post-processing the result set of features to obtain a processed information signal.

The application discloses a telephone diversion method and system based on machine learning, and a storage medium, comprising: constructing a sample feature library of harmful and normal numbers; preprocessing feature data in the sample feature library, and dividing the feature data into a feature training set and a feature test set; constructing a decision tree based on the feature training set; screening decision rules in the decision tree based on the feature test set, obtaining a feature combination with the best performance on the test set and a segmentation threshold corresponding to the feature combination; collecting new batches of call numbers, obtaining feature data of the call numbers, identifying the feature data of the call numbers based on the optimal feature combination and the segmentation threshold corresponding to the feature combination, and predicting whether the call number is a harmful number or a normal number; and updating the optimal feature combination and the corresponding segmentation threshold regularly, so that when the features of harmful numbers change, a good diversion effect can still be achieved. The application classifies sample feature data by using a machine learning algorithm, screens a feature combination and a segmentation threshold with the best diversion effect, and can more accurately identify harmful telephone numbers, thereby reducing misjudgment and missed judgment.

The invention discloses a customer purchase willingness evaluation method and device, equipment and a storage medium, belongs to the technical field of artificial intelligence, and is applied to a purchase willingness prediction scene in the financial field. According to the method, the most representative key features are extracted by integrating historical purchase records, customer behavior logs and market trend data and adopting dimension reduction technologies such as principal component analysis, and noise and redundancy influences caused by high-dimensional data are avoided. Meanwhile, by identifying and quantifying the interaction effect between the features and constructing a weighted interaction relation matrix by using a random forest algorithm, the comprehensive influence of different feature combinations on the purchase intention is deeply revealed, so that the perception ability of the model to the real intention of the customer is improved. In addition, through clustering analysis and dynamic behavior capture, a customer group behavior mode is constructed, and demand characteristics and transformation potentials of different customer groups are accurately described. According to the method, the explanation depth of the customer purchase behavior is improved, and the accuracy and the stability of the prediction model are remarkably enhanced.

The invention relates to the technical field of user behavior analysis, and provides a user behavior anomaly detection method and system based on a multi-dimensional baseline, and the method comprises the steps: calculating the similarity of any two historical behavior feature combinations, dividing the historical behavior feature combinations into multiple classes according to the similarity, and carrying out the statistics of the frequency of each class; for each category, summarizing a plurality of feature modes, and counting the frequency of each feature mode; based on the frequency of the corresponding category and the frequency of the feature mode, selecting one feature mode as a primary normal behavior baseline; calculating the weight of each feature by using a TF-IDF algorithm, and selecting the feature with the highest weight as a key feature; combining the initial normal behavior baseline with the key features to serve as a normal behavior baseline; obtaining a user behavior log, and extracting a plurality of behavior features through the large model to obtain a behavior feature combination; and comparing the behavior feature combination with the behavior baseline, and judging whether the user behavior is abnormal or not. And false alarms of abnormal behaviors are reduced.

The present application relates to a kind of based on multimodal feature combination image uniqueness anti-counterfeiting identification method, belong to image processing and anti-counterfeiting identification technical field.It includes: constructing anti-counterfeiting carrier, include by random microstructure Random texture area And the readable identification area of associated digital identity information;Registration stage, based on texture feature extraction model and morphological feature extraction model respectively, the deep texture feature of random texture area and morphological structure feature are extracted, the unique feature fingerprint is generated by the multi-modal fusion of the two, and is bound and stored with digital identity information;Verification stage, through terminal equipment acquisition image to be measured, locate random texture area and extract measured texture feature, based on metric learning algorithm, measured texture feature and the unique feature fingerprint in database are mapped to the same feature space for comparison, and output true and false determination.The present application realizes the anti-counterfeiting identification of high security and high identification accuracy by multimodal feature combination and asymmetric registration verification architecture.

The application discloses a kind of underwater vehicle degradation characterization method, belong to underwater vehicle technical field, steps are as follows: the full life cycle data collected by sensor carried on underwater vehicle is collected, data is preprocessed;Feature extraction is carried out to the data after pre-processing;Design a kind of layered cooperation particle swarm optimization algorithm based on fuzzy analytic hierarchy process, the selected feature is obtained by screening the extracted feature, and the final optimal feature combination is obtained;Dynamic correlation mahalanobis distance is used to weight and fuse the selected feature, and the fusion feature is obtained;Using the UMAP dimension reduction method with degradation perception, the degradation process is divided into three stages of early, middle and late according to the full life cycle data, an independent UMAP submodel is trained in each stage, and then the degradation index is fused by transfer learning.This application breaks through the limitation of single signal, realizes multi-component coupling degradation recognition, strengthens early degradation capture, reduces the rate of missed judgment and fault risk, has environmental dynamic adaptability, and adapts to complex deep-sea working conditions.

The present application relates to the technical field of gene mutation prediction, and discloses a gene mutation prediction method generated by multi-mode annotation, and the specific process comprises the following steps: carrying out mutation type annotation on input single-base mutation position information to obtain mutation basic information containing mutation types, then using an ANNOVAR annotation tool, SpliceAI splicing effect prediction software and reference mutation information of a function effect database to carry out multi-dimensional feature annotation, using Bayesian PCA based on the obtained multi-dimensional feature mutation data set to fill in the annotation data, then using an automatic engineering feature list and a separated feature selection list to carry out feature combination and screening, and obtaining a gene mutation prediction score after gradient generation tree algorithm. The present application can be used for predicting all non-synonymous exon mutations, has good performance in classifying rare benign mutations, and can identify a small amount of mutations with high pathogenic probability from a large amount of candidate mutations.

The present disclosure relates to systems, methods, and non-transitory computer-readable media for performing interactive digital image editing operations with machine learning models and feature backpropagation refinement layers. For example, the disclosed systems perform interactive digital image editing operations by incorporating a feature backpropagation refinement layer into a non-interactive machine learning model that utilizes a consistency loss to adjust the feature backpropagation refinement layer according to one or more user interactions. In some embodiments, the disclosed systems utilize a feature backpropagation refinement layer that includes a bias sublayer for localizing changes to a digital image and a convolution sublayer for per-channel scaling and feature combination across channels. In some cases, the disclosed systems utilize a consistency loss that facilitates localized modifications to a digital image based on a distance of various pixels or features from a user interaction.

The invention discloses a training method and device of a business handling evaluation model, equipment and a storage medium, and can be applied to the field of financial science and technology. The method comprises the steps of obtaining historical business handling data of a target bank business, and constructing business sample data based on the historical business handling data; based on the business sample data, performing first model training on a pre-constructed machine learning model to obtain a candidate business handling evaluation model; adopting a candidate business handling evaluation model to perform business handling prediction on different business dimension feature combinations in the historical business handling data to obtain candidate predicted business handling tags, and determining feature importance degrees of different business dimension features in the historical business handling data based on the predicted business handling tags; and performing second model training on the candidate business handling evaluation model based on the feature importance and historical business handling data to obtain a target business handling evaluation model. According to the technical scheme, the model performance and the prediction precision are improved.

The application discloses a test method, device, storage medium and computer equipment, the method comprises the following steps: obtaining target order information and to-be-tested state characteristics; performing feature combination processing according to the to-be-tested state characteristics and the target order information, and determining state characteristic samples; generating test cases according to the state characteristic samples; performing simulation processing according to the test cases, and obtaining simulation results; and determining test log information of the state characteristic samples according to the hit times in the simulation results. Thus, all possible state characteristics generated in the order life cycle are used as the basis of sample permutation and combination, the scene measurement of non-terminal states such as intermediate states and process states is realized, the perception of the test personnel to scene omissions is improved, and online problems caused by test omissions are effectively reduced. Meanwhile, test cases are generated based on real order data, the test case generation efficiency is improved, the test cases can cover various real business scenes, and the test cost is reduced.

The invention relates to the technical field of intelligent sensors, and provides an operation fault diagnosis method and system applied to a power transformer, and the method comprises the steps: obtaining a plurality of audio signals under various power transformer fault working conditions; obtaining a plurality of feature points and feature combinations thereof of each audio signal; importance parameters of each feature combination in each working condition are obtained; distinctive evaluation parameters of different working conditions are obtained, and then working condition evaluation indexes of all frequencies are obtained; obtaining a plurality of sample combinations of the audio signal, calculating the similarity between different sample combinations, and constructing a neighbor matrix; performing dimension reduction on each sample combination of the audio signals through an LPP algorithm to obtain a final dimension reduction result of each audio signal; and an operation state diagnosis model is constructed, and online fault diagnosis is carried out through the operation state diagnosis model after dimension reduction is carried out on real-time audio signals of the power transformer. The invention aims to solve the problem of diagnosis result deviation caused by different fault detail features.