356 results about "Unsupervised clustering" patented technology

The invention belongs to the technical field of data processing, and particularly relates to a hidden ore body evaluation and positioning method based on multi-source data processing. The method mainly aims at the problems of incompleteness and isomerism of multi-source geological data in acquisition, fusion and modeling. Comprising the following steps: acquiring hyperspectral, geochemical and magnetic anomaly multi-source data of an evaluation area; intelligently complementing missing modal data by using a generative adversarial network based on geological constraints and modal outburst to form a complete multi-source data set; an unsupervised clustering algorithm combining geological correlation and entropy weight analysis is adopted to construct high-confidence-coefficient pseudo-label data, and knowledge mining of unlabeled samples is achieved; feature purification and dimension reduction are carried out through multi-modal feature fusion and hierarchical principal component analysis, and key feature vectors representing the existence of the ore body are extracted; and finally realizing space prediction of the concealed ore body by utilizing the classification model. According to the method, a high-quality data basis and a unified processing framework are provided for intelligent recognition of the hidden ore body, and efficient and accurate positioning of the hidden ore body is achieved.

The invention discloses a data labeling method and system based on user behaviors and attention tracking, and the method comprises the steps: synchronously collecting multi-source behavior signals of a mouse, a keyboard, eye movement and the like of a doctor in real time, combining identity and interface metadata, and carrying out the standardized normalization, abnormality elimination and short time sequence behavior unit division. And extracting individual behavior micro-modes by using unsupervised clustering, and constructing a behavior portrait library. Through multi-modal time sequence modeling and a self-adaptive space-time attention mechanism, behavior characteristics, an interface area and a report text are deeply fused, a multi-level correlation probability is output, and high-precision automatic tagging of content and an image area is realized.

The invention discloses a high and cold arid region slope soil stability safety risk evaluation system and method, and relates to the technical field of slope engineering. The state of each side slope under multi-dimensional indexes such as freeze-thaw cycle frequency, dry-wet alternation index, wind erosion strength, shear strength, water content, porosity and fracture density is quantified, the feature similarity between any two side slopes is calculated, and then a side slope feature similarity matrix is formed. Based on the matrix, an unsupervised clustering algorithm (such as spectral clustering, similarity propagation and the like) can be adopted to divide a plurality of side slopes into similar subsets with structural characteristics similar to environmental response, and category attribution of risks is achieved. On the basis, structural variation analysis, historical instability statistics and central risk difference extraction are performed on similar slope samples, so that the internal instability tendency of the slope can be identified, and a risk prediction model suitable for the type of slope can be constructed through feature training.

The invention discloses a lightweight knowledge graph rapid construction method and system based on an NLP technology. The method comprises the steps of preprocessing an unstructured text and segmenting the unstructured text into semantic segments; unsupervised clustering is combined with the contour coefficient to determine the optimal clustering number, and a semantic association fragment cluster is obtained; performing word segmentation, part-of-speech tagging, NER and entity linking on the fragment cluster, extracting an entity and initial relationship, and fusing semantic similarity, TF-IDF word frequency collaboration degree and co-occurrence frequency to calculate a relationship edge weight; constructing a lightweight knowledge graph; in the question and answer stage, questions are disassembled through a process engine, related sub-graphs are retrieved, and answers with reasoning links are generated. The system correspondingly comprises a text preprocessing module, a clustering module, an entity relation processing module, a graph construction module, a question and answer reasoning module and a storage module. According to the method, the construction cost is reduced, the interpretability and the module coupling degree are improved, multiple scenes such as government and enterprise public opinions and medical assistance are adapted, and the problems of weak generalization, poor real-time performance and'black box 'in the traditional technology are solved.

The invention discloses a federated learning backdoor attack defense method based on layer perception detection, which relates to the technical field of artificial intelligence security, and comprises the steps of client training and uploading, malicious client identification, backdoor key layer detection and deletion, robust aggregation updating and adversarial training for enhancing robustness. According to the method, the model parameters uploaded by the clients are subjected to clustering analysis, and the maximum benign cluster is identified by adopting an unsupervised clustering method, so that instability caused by single threshold judgment is avoided, and the benign client and the malicious client can be distinguished; and meanwhile, layer perception detection and elimination are used in the scheme, so that the backdoor introduced by a malicious client can be effectively identified and eliminated on the premise of ensuring the global model precision, and the influence of backdoor attack on the model is inhibited.

The invention relates to the technical field of intelligent driving, and particularly provides a personalized driving behavior identification method based on deep embedded clustering. Extracting a low-dimensional depth feature vector of the driving time sequence segment through a deep neural network auto-encoder, performing end-to-end joint optimization on the auto-encoder and unsupervised clustering by using a deep embedded clustering method, generating a driving style clustering result, and labeling a clustering result label; freezing the trained auto-encoder parameters, constructing a driving style classification model, and establishing a mapping relation between driving time sequence fragments and style labels; and distributing semantic tags for clustering results by extracting interpretable features. According to the end-to-end joint optimization technology based on deep embedded clustering, the convergence speed and generalization ability of the model are remarkably improved, and the driving style identification precision is far higher than that of a traditional machine learning method.

The invention discloses an LCD display screen production quality collaborative management and control method based on a detection data chain, particularly relates to the technical field of industrial production quality management, and is used for solving the problem of quality root cause misjudgment caused by complex association of multi-source data in the existing liquid crystal display screen manufacturing process. The method comprises the following steps of: constructing a detection data chain covering an array, a box forming process and a module process, firstly performing multi-source data correlation analysis to obtain a variable pair set, then analyzing an intermediary influence path between variable pairs to eliminate false correlation, and then determining a real causal direction by applying a causal discovery algorithm based on conditional independence. Carrying out unsupervised clustering analysis on the basis of defect characteristic indexes, verifying the consistency of causal directions in different data subsets, and finally outputting a stable quality root cause analysis result to realize accurate positioning and effective management and control of key process parameters.

The invention relates to the technical field of federated learning security, and discloses a federated learning poisoning defense method based on time-frequency spectrogram and comparative learning, which comprises the following steps: receiving model update uploaded by each client, grouping and vectorizing parameters according to model layers, and generating a time-frequency spectrogram by applying short-time Fourier transform to parameter vectors of each layer; based on the time-frequency spectrogram, constructing a positive sample pair through data enhancement, carrying out difficult negative sample mining, and training an encoder by using a contrast loss function to extract an embedded vector with high discriminant power; and performing unsupervised clustering on the embedded vector by using a DBSCAN clustering algorithm, judging the maximum cluster as a benign client, performing final judgment in combination with historical malicious records, and only aggregating model parameters of the benign client to update a global model. According to the invention, high-precision detection of attack features can be realized, and a more universal, more efficient and more practical federal learning poisoning attack defense method is realized.

The invention discloses an automatic cable surface defect detection method and system based on a convolutional neural network, and belongs to the technical field of computer vision and deep learning, and the method comprises the following steps: collecting an initial qualified segment image during the production of a new batch of cables, and extracting a library texture feature vector through a pre-trained texture feature coding network, unsupervised clustering is carried out, and cluster centers left after the abnormal clusters are removed and statistics of the cluster centers are stored in a normal texture self-adaptive reference library of the batch; collecting a cable surface expansion image on line, dividing the image into image blocks, inputting the image blocks into the texture feature coding network to obtain a detection texture feature vector, obtaining a texture deviation degree according to the detection texture feature vector and each cluster center in the normal texture adaptive reference library, and judging whether the texture deviation degree is a texture abnormal point or not; and carrying out connected domain analysis on texture abnormal points of the same-frame expanded image to obtain candidate abnormal regions. The problem that real defects and instantaneous noise are difficult to distinguish in the prior art is solved.

The invention belongs to the technical field of remote sensing, and relates to a multispectral remote sensing atmospheric correction method based on fusion of a physical model and a neural network, which comprises the following steps: S1, classifying aerosol parameter data obtained based on foundation observation, remote sensing inversion or meteorological model and satellite data fusion inversion by adopting an unsupervised clustering algorithm, extracting a typical aerosol mode with physical representativeness; s2, in combination with the typical aerosol mode, simulating the apparent reflectivity of an observation channel under a plurality of atmospheric states and observation geometric conditions by using a radiation transfer model, generating a lookup table covering a wide parameter space, and constructing a training data set; and S3, constructing a nonlinear regression model, taking the training data set as a training sample, learning a mapping relation among an observation angle, an aerosol condition and surface reflectance, performing atmospheric interference correction on an actual multispectral remote sensing image under a pollution condition, and outputting a surface reflectance result.

The invention provides an automatic driving takeover quality evaluation method based on fusion of driving behaviors and physiological and psychological characteristics, and the method comprises the steps: carrying out the multi-source data collection and time synchronization of the behavior response and physiological change of a driver in the automatic driving takeover process, and obtaining multi-modal data; performing signal cleaning and adaptive segmented filtering on the multi-modal data to obtain filtering data; differential dynamic feature extraction and time window construction are performed on the filtering data, and multi-modal feature fusion is performed after key features are obtained, so that comprehensive features can be obtained; performing unsupervised clustering on the comprehensive features to obtain different takeover modes; and constructing an intermediary effect model based on the structural equation model and a Bootstrap test method, and performing quality evaluation on the takeover mode to obtain an evaluation result. According to the method, an intelligent evaluation framework combining Gaussian mixture model clustering and intermediary effect analysis is introduced, so that multi-dimensional and interpretable modeling and quality judgment of driver takeover behaviors are realized.

The invention relates to an operation and maintenance alarm root cause positioning method and system fusing a knowledge graph and a large model, and belongs to the technical field of data processing and intelligent operation and maintaining.The method comprises the steps that original alarm data are obtained from a plurality of heterogeneous monitoring sources based on alarm data query service of time period parameters and subjected to cleaning and structural processing, and the original alarm data are obtained; obtaining alarm data to be analyzed; performing unsupervised clustering analysis to realize alarm automatic classification and noise filtering to obtain classified alarms and category labels; extracting corresponding object topological relation information from a pre-constructed knowledge graph; and inputting the classified alarms, category labels and object topological relation information into a large language model, guiding the large language model to perform semantic reasoning through a preset cue word project, and generating a structured diagnosis report containing root cause diagnosis, an influence range and repair suggestions. According to the method, the alarm storm is inhibited, the fault root cause is quickly, accurately and interpretably positioned, and the operation and maintenance efficiency and the system stability are improved.

The invention relates to a manufacturing system risk control knowledge matching method based on semantic embedding and clustering analysis, and the method comprises the following steps: collecting and preprocessing risk control text data: collecting unstructured text data of a manufacturing system history record, and obtaining preprocessed risk control text data, constructing a professional corpus for a discrete manufacturing scene; text semantic embedding generation; semantic clustering modeling: performing unsupervised clustering modeling on all semantic vectors, mining semantic association and potential structures between texts, obtaining semantic representations of risk control knowledge through a clustering algorithm, and assisting in generating clustering tags; and a risk knowledge matching mechanism.

The invention relates to the technical field of intelligent switching, in particular to a multi-mode dynamic coupling intelligent switching method and system based on road condition recognition and a vehicle. The method comprises the steps that gradient, vehicle speed, accelerator and brake signals are collected in real time, and a multi-dimensional feature vector fusing driving habits is constructed after preprocessing; a real-time road condition label is automatically generated through unsupervised clustering, and the clustering effectiveness is verified online by using a contour coefficient. Enabling the road condition label and the current power mode to form a system state, and inputting the system state into a self-learning decision network based on reinforcement learning; and the network updates the state-action value matrix through a reward function in a time sequence difference mode, and outputs a target power mode with the highest accumulated reward. And calculating the motion confidence coefficient of the target mode based on the value matrix, if the motion confidence coefficient exceeds a threshold value, generating a switching instruction, and smoothly adjusting the torque by an execution mechanism through a torque transition algorithm to realize impact-free switching. Automatic road condition recognition, self-adaptive decision making and non-inductive execution are achieved, and smoothness and adaptability are improved.

The invention provides a fault positioning method and device, electronic equipment and a medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: carrying out the sampling of optical power data based on a preset sampling frequency, and determining optical power sequence data, the optical power sequence data comprises the identification of a PON port, the equipment identification of a user, and an optical power parameter, each PON port corresponds to a plurality of optical power sequence data; based on the optical power sequence data, determining a similarity value between any two users through unsupervised clustering and time sequence alignment in sequence; and correlation analysis is carried out on the similarity value between any two users, fault information is determined, and the fault information comprises a fault position, a fault equipment identifier and an identifier of a fault PON port. According to the method, through clustering and time sequence alignment, similar value calculation of optical power sequences and accurate delimitation of faults are realized, and the accuracy of fault positioning is improved; and the fault analysis and positioning efficiency is improved. Furthermore, the network stability and the user experience are guaranteed.

The invention discloses a wind power blade state monitoring and fault early warning system based on a wavelet clustering algorithm, and relates to the field of wind power blade fault diagnosis. According to the system, blade operation data are synchronously acquired through a vibration sensor and an acoustic emission sensor, time domain, frequency domain and time-frequency domain features are fused by adopting a multi-domain joint feature extraction method, unsupervised clustering analysis is performed on feature vectors by utilizing an improved wavelet clustering algorithm for automatically acquiring local clustering information, and a damage mode is automatically identified. Fault classification is realized by establishing a support vector machine diagnosis model, an acoustic emission event rate fluctuation verification mechanism based on blade rotation periodicity is innovatively introduced, and the authenticity of damage is verified through mutual verification of vibration and acoustic emission signals. According to the method, the problems of insufficient feature extraction, poor clustering effect and high false alarm rate of a traditional method are solved, accurate identification and early warning of multiple damage modes are realized, and the accuracy of fault identification is effectively improved.

The invention relates to the technical field of safety monitoring artificial intelligence, and discloses a safety management method and system based on artificial intelligence and big data. The method comprises the steps that multi-modal real-time monitoring data is collected through a distributed sensor array, and a standardized sequence is generated after preprocessing; the system accesses a historical security event library, adaptively divides risk category clusters by using an unsupervised clustering algorithm, and extracts key risk features and influence domains for each cluster. On the basis, a plurality of special self-adaptive security analysis engines are configured to perform parallel analysis on real-time data to generate a plurality of security analysis feature maps. And carrying out weighted fusion and enhancement on the feature maps through a feature fusion network to obtain a unified security situation feature vector, and finally outputting a security level and a management and control instruction through a decision model. According to the method, autonomous identification, dynamic evaluation and accurate response of complex risks are realized, and the intelligent level of safety management and the decision accuracy are improved.

The invention provides an operation and maintenance performance assessment data management system based on dynamic weight optimization, and relates to the technical field of operation and maintenance management, and the system comprises a quantification module which is used for obtaining an original operation and maintenance data sequence through a data collection interface, and mapping the operation and maintenance data sequence to a multi-dimensional feature space; in the multi-dimensional feature space, quantitative analysis is carried out on spatial distribution of data points, and data density distribution characteristics of all areas are determined; the calculation module is used for executing unsupervised clustering operation based on the density distribution characteristics and the proximity relation of the data points to obtain a group of data sets with internal consistency; core statistical features are calculated for each generated data set, and a specific calibration coefficient is derived for the data set based on the core statistical features. According to the invention, precision, scenario and high efficiency of operation and maintenance performance assessment are realized.

The invention discloses a dynamic visual guidance method and system for cooperative assembly of multiple mechanical arms, and belongs to the technical field of industrial robot visual guidance, and the method comprises the steps: synchronously collecting an image sequence through a binocular visual system, synchronously recording the pose sequence of each mechanical arm through a robot controller, and obtaining an observation data set; obtaining a two-dimensional trajectory flow of vehicle body assembly area feature points and instrument body feature points through an optical flow tracking algorithm, and obtaining a hand-eye matrix correction through a nonlinear optimization algorithm; obtaining an updated hand-eye matrix and a local pixel compensation field through a spatial interpolation algorithm based on the re-projection error of each feature point; according to the method, self-adaptive compensation of flexible deformation of a vehicle body and dynamic reflection of an instrument is realized by adopting unsupervised clustering and spatial interpolation according to a re-projection error, and the precision of cooperative assembly of multiple mechanical arms and the self-adaptive capability of a system in a complex industrial environment are improved.

The invention discloses a digital project management method and system based on artificial intelligence, and relates to the technical field of project management, and the method comprises the steps: constructing a project digital portrait and multi-dimensional feature analysis framework, determining an analysis domain boundary of intelligent auditing, and automatically generating an intelligent analysis grid adaptive to multi-source heterogeneous data; and performing grid division on project progress data according to an analysis domain boundary, performing adaptive partition on a project grid through an unsupervised clustering algorithm, dynamically determining a critical threshold value of risk early warning based on a project digital portrait and intelligent analysis grid and comprehensive risk stress analysis, and outputting a risk early warning decision management conclusion containing the critical threshold value. According to the management system, an intelligent analysis grid adaptive to multi-source heterogeneous data is generated by constructing a project digital portrait and a multi-dimensional feature analysis framework, and a risk early warning critical threshold value is dynamically determined, so that real-time perception, accurate analysis and intelligent decision support of the full life cycle state of a project are realized.

The invention provides a multi-modal false news detection method based on unsupervised clustering and frequency domain information, and relates to the technical field of image text data. The method comprises the following steps: based on multi-modal sample data, respectively extracting text features, visual features and image text features, obtaining text feature clustering labels according to text feature global semantic correlation, inputting the text features with the labels into an unsupervised clustering learning network, and obtaining an unsupervised clustering learning result; semantic consistency is enhanced through a bidirectional gating loop unit and a multi-head attention mechanism, and enhanced text features are obtained; extracting frequency domain features based on the visual features, and fusing the frequency domain features with the spatial domain features to obtain visual joint features; and after image text features are also enhanced by the unsupervised clustering learning network, the image text features are fused with text and visual features through an attention fusion module by taking the image text features as a bridge to obtain multi-modal fusion features, and the multi-modal fusion features are input into a full connection layer to complete false news classification. According to the method, the accuracy of false news detection is effectively improved.

The invention discloses a segmented clustering-feature mapping combined voice abstract method. The method comprises the following steps of: segmenting long voice according to a timestamp, extracting a fragment-level semantic vector by using an audio encoder, clustering the fragment-level semantic vector into N classes through an unsupervised clustering algorithm (K-means), selecting representative fragments closest to a cluster center from each class, and splicing the representative fragments into compressed voice segments according to a time sequence; then, semantic features of the spliced voice segments are extracted through an audio encoder, and the semantic features are mapped to a feature space capable of being recognized by a large language model through a Project module; and finally, generating an abstract text by the combined fine-tuning large language model. Core semantics are reserved through clustering screening, redundancy is reduced, error propagation of a traditional cascading mode is avoided, side language information is reserved, various audio encoders and large language models are compatible, and the method can be expanded to a multi-language voice abstract task.

A method to identify which safety leading indicators are not capturing hazards during safety inspections that later contribute to safety incidents and to identify blind spots in safety inspections. The method includes inputting a plurality of incident reports from the worksite and a plurality of inspection reports from the worksite into a system that uses machine learning (ML) and artificial intelligence (AI); identifying socio-technical factors influencing safety risks at the worksite using unsupervised clustering for topic modeling; evaluating semantic similarity between the plurality of incident report and the plurality of inspection reports conducted in a preceding time period to detect missing socio-technical factors not captured during safety inspections; and providing the identified socio-technical factors previously not captured during safety inspections to a stakeholder for development of modified inspection checklists tailored to capture the identified socio-technical factors previously not captured during safety inspection.

The embodiment of the invention provides a fault positioning method and device based on association rules, equipment and a program product. Relates to the field of energy storage system fault diagnosis. The method comprises the following steps: acquiring a state code, a fault code and an operation parameter of the energy storage thermal management equipment; based on the state code, the fault code and the operation parameter, generating a fault rule base by adopting an association rule mining algorithm; unsupervised clustering analysis is carried out on the operation parameters, and abnormal data deviating from the standard are screened out; fusing the abnormal data with the fault rule base to generate a fault map; and based on the fault map, a fault positioning report of the energy storage thermal management equipment is output, the fault positioning report at least comprises a fault propagation path and a maintenance scheme, and the fault propagation path is a path of a causal relationship between the fault code and the abnormal mode. The method is used for achieving the effect of improving the fault positioning efficiency and accuracy.

The lithium battery internal short circuit diagnosis method and system based on the double-state cycle equivalent circuit of the application relate to the lithium battery internal short circuit detection technical field. Firstly, a hybrid model of a battery second-order RC equivalent circuit embedded in an RNN is constructed, the state update is constrained to conform to a physical equation through customization of a cycle unit, and an MLP subnet is used to fit an OCV-SOC relationship; a voltage interval weighted loss function is designed to strengthen the accuracy of a sensitive area, and learning rate decay and gradient clipping are combined to optimize physical parameters and MLP parameters; a trained model is used to predict a healthy voltage and generate a residual error; a sliding window is used to extract a residual error subsequence, the DTW distance of the residual error subsequence and a zero vector is calculated, and a mean value and a maximum value are used to form a two-dimensional feature; and automatic fault identification is realized based on Gaussian mixture model unsupervised clustering. The application combines the advantages of physical mechanisms and data driving, can detect early internal short circuits with high precision without fault samples, and solves the problems of high false alarm rate of the threshold method, poor adaptability of the model method, and dependence of the data driving on labeled samples.

The invention relates to the technical field of intelligent power grid monitoring, in particular to a measuring and testing method for an intelligent low-voltage distribution box line, which comprises the following steps of: acquiring and decomposing an original signal of a reconstructed line to form a characteristic image; steady-state and outlier feature clusters are autonomously emerged through unsupervised clustering, and anomaly recognition without a preset threshold value is achieved. And parameterizing the steady-state features into a dynamic template for real-time comparison and marking of feature deviation events. For a deviation event, an original signal is deeply deconstructed to extract energy distribution and phase synchronization distortion characteristics, and the characteristics, a switching event and a load record are subjected to time sequence alignment and coupling analysis, so that the logic relevance of an abnormal source is locked. Based on the relevance, the sensing device is driven to adaptively adjust a sampling strategy, a new round of verification measurement is started, and a closed loop from sensing, diagnosis to optimization tracking is formed. According to the method, unknown abnormity can be adaptively identified, and automatic analysis of abnormity reasons and closed-loop optimization of a monitoring strategy are realized.

The invention provides a double-current comparison and DHI combined equipment fault detection method and system, and belongs to the technical field of power equipment state monitoring and fault diagnosis. The method comprises a time sequence perception adversarial enhancement module, a generative adversarial network (GAN) data expansion module, a double-flow contrast attention network (D-CAN) feature extraction module and a dynamic health index (DHI) calculation module. The core lies in that a fault sample with time sequence correlation is supplemented through a physically constrained GAN, robust features are extracted by using a ResNet1D and Transform fused double-flow network, and a health state is quantified in combination with unsupervised clustering and mahalanobis distance. Through simulation and experimental verification, zero-delay detection of the early fault of the transformer can be realized, the output health index and the 3D visualization result can provide an accurate basis for operation and maintenance of the transformer, and the diagnosis accuracy, the data utilization rate and the dynamic adaptability of state evaluation are remarkably improved.

The invention relates to the technical field of power system automation, and discloses a distributed photovoltaic spatio-temporal data recovery and unsupervised clustering method and system.The distributed photovoltaic spatio-temporal data recovery and unsupervised clustering method comprises the steps that observation data of a distributed photovoltaic system are collected, and the observation data comprise historical distributed photovoltaic output time sequences; the collected historical distributed photovoltaic output time sequence is used as an input signal to be transmitted to each module of the GRAIL framework so as to drive a network to jointly learn time sequence dynamic characteristics and a clustering structure; wherein the GRAIL framework is a closed-loop iterative structure, and comprises a clustering sensing data generation module, a latent feature representation module and a self-adaptive clustering and refining module. The method can be widely applied to distributed photovoltaic station operation monitoring and abnormity identification scenes, and intelligent regional state recovery and mode grouping are realized.

The invention discloses a special equipment safety monitoring method based on the Internet of Things, and provides a multi-mode sensing data acquisition, noise removal, data standardization and label completion method. Rare abnormal samples are screened through unsupervised clustering and anomaly detection, and diverse pseudo-abnormal data are expanded through a conditional generative adversarial network. And deep representation learning, multi-modal feature mapping and migration fusion are further applied to construct a multi-modal collaborative abnormal feature set, and the multi-modal collaborative abnormal feature set is incorporated into an actively optimized security event discrimination model to realize dynamic rule adaptive evolution.

The invention discloses an unsupervised silicon wafer surface defect detection method and device based on hierarchical clustering, relates to the field of image processing and computer vision, and utilizes an unsupervised clustering technology to detect silicon wafer surface defects. The method comprises the following steps: extracting multi-scale features from a plurality of normal silicon wafer images through a pre-trained convolutional neural network; integrating the multi-scale features of the plurality of pictures by adopting a hierarchical fusion strategy; constructing a feature memory library by using a clustering algorithm, and generating a clustering center; generating patch features for the test image through a consistent fusion process; through comparative supervision, normal patches are close to a clustering center, and abnormal patches are pushed away; calculating the distance between the patch and the clustering center, and judging abnormity; and mapping the position of the abnormal patch, and positioning scratches, cracks, pollution or recesses. According to the method, through an unsupervised mode, a large amount of defect labeling data is not needed, the data preparation cost is effectively reduced, and the accuracy and robustness of silicon wafer surface defect detection are remarkably improved.