173 results about "Probability vector" patented technology

The invention discloses a fault diagnosis method and device, computer equipment and a computer readable storage medium, relates to the technical field of power systems and automation thereof, and solves the problem of fault misjudgment caused by important fault characterization of a method which is easy to lose when single modal information is used for fault diagnosis at present. The method comprises the following steps: converting a bus voltage signal into a two-dimensional time-frequency diagram based on a continuous wavelet transform model, and converting the bus voltage signal into a one-dimensional frequency spectrum sequence based on a fast Fourier transform model; performing feature extraction on the two-dimensional time-frequency graph based on a time-frequency image feature extraction model to obtain a first feature vector, and performing feature extraction on the one-dimensional frequency spectrum sequence based on a frequency spectrum sequence feature extraction model to obtain a second feature vector; performing feature fusion on the first feature vector and the second feature vector based on a feature fusion model, and obtaining a category probability vector by using a fault diagnosis model based on the fused feature vector; and determining the highest probability value in the category probability vector, and taking the corresponding fault type as a target fault type.

The invention discloses a long-tail SKU sales prediction method and system, and belongs to the technical field of sales prediction, and the method comprises the following steps: obtaining SKU sales feature data, and carrying out the preprocessing of the sales feature data; obtaining a multi-dimensional feature table based on the standardized training data set; a rule engine and a lightweight classification model are used for joint judgment to classify the multi-dimensional feature table, and four SKU category labels and probability vectors thereof are output; differentiated training models of the four SKU category labels are established, different types of SKUs are routed to the corresponding training models, and SKU sales volume point predicted values are output; probability calibration is carried out, and a plurality of quantiles are calculated to form interval prediction; establishing an SKU multi-hierarchy relationship, and performing hierarchy consistency solution on prediction results among multiple hierarchies; in the online operation process of the prediction model, input data distribution drift is monitored, and model updating is triggered when the input data distribution drift exceeds a threshold value. According to the method, the prediction precision and stability of the long-tail SKU are improved, and rapid adaptive prediction of sudden and off-peak demands is realized.

The invention relates to the technical field of on-line monitoring, in particular to a flue gas cooler leakage monitoring system and method.The method comprises the steps that a multi-source real-time sensing layer module continuously collects tube plate weld joint area temperature data, cooling water flow characteristics and flue gas sulfur-containing gas concentration, and the sampling period is 250 milliseconds; the intelligent feature fusion module performs noise reduction processing and dynamic weight fusion on the multi-source data and outputs a leakage probability vector; the dual-system collaborative early warning module maps the probability vectors into confidence levels, a 65% confidence threshold triggers an acoustic verification system to start, an acoustic emission sensor array captures fractured sound wave features and then compares the fractured sound wave features with a voiceprint library through a convolutional neural network, and an early warning signal is output when the fractured sound wave features exceed a 0.85 similarity threshold; and the dynamic knowledge base updates the case characteristics online according to the early warning result. According to the system, second-level identification and continuously optimized closed-loop monitoring at the initial stage of leakage germination are realized, and the problem of response delay caused by intermittent detection in the prior art is solved.

The invention belongs to the technical field of spectral analysis, and discloses a Raman spectrum characteristic peak segmentation method based on distance vector and probability vector output. Spectral detail features are extracted through multi-scale wavelet decomposition, and a spectral multi-layer structure representation matrix is constructed; identifying a potential peak site and a hierarchical affiliation relationship thereof based on the local curvature change rate; generating an adaptive distance calculation kernel function in combination with the asymmetry index and the peak shape complexity coefficient; and calculating a distance vector of hierarchical perception and a probability vector extracted by deep learning, and constructing a joint segmentation decision function of a multi-layer peak structure. According to the method, the hierarchical relationship among the main peak, the shoulder peak and the sub-peak can be accurately distinguished, the complex conditions of peak overlapping, asymmetric peak shapes, low signal-to-noise ratio and the like are effectively processed, and the accuracy and the reliability of Raman spectrum analysis are improved.

The invention discloses a context poisoning attack defense method oriented to an RAG application, and relates to the technical field of RAG. the method comprises the following steps: inputting a target query statement, and retrieving the target query statement to obtain multiple pieces of context information; taking representative sentences in the retrieved context information, and identifying and filtering potential malicious template clusters; the big language model gives all candidate answers according to existing context information, the logarithmic probability of all contexts to different candidate answers is calculated, and after the influence of parameter knowledge of the big language model is removed from the logarithmic probability, the support degree of all contexts to different candidate answers is obtained; the whole logarithmic probability vector is used as a support degree distribution condition of the context to the candidate answers; identifying a single piece of harmful information from the support degree distribution condition of the context to the candidate answers through a logistic regression model so as to filter wrong answers; according to the attack defense method provided by the invention, centralized injection of multiple malicious texts and sparse injection of a small number of malicious texts can be defended.

The invention discloses a fault self-adaptive diagnosis method for a rolling bearing for a rotating machine, and the method comprises the steps: firstly collecting a vibration signal in the operation process of the rolling bearing, and carrying out the preprocessing of the vibration signal, and obtaining the preprocessed vibration signal data; performing time sequence modeling on the preprocessed vibration signal data, extracting time sequence features, and generating a stage probability vector according to the time sequence features; mapping the stage probability vector into a stage embedded vector, introducing the stage embedded vector into a stage adaptive attention mechanism, and outputting an optimized feature vector; and finally, constructing a fault diagnosis network based on deep learning to perform fault diagnosis on the rolling bearing, outputting probability distribution of various faults of the rolling bearing, selecting the fault type with the maximum probability as a diagnosis result, and determining the severity of the fault type. According to the method, the fault stage can be automatically sensed, the evaluation standard can be dynamically adjusted, the method adapts to industrial data characteristics, the initial detection rate and the late false alarm rate are balanced, and therefore more accurate predictive maintenance is achieved.

The invention discloses a power electronic transformer working mode recognition method and device, equipment, a medium and a product, and relates to the field of artificial intelligence, and the method comprises the steps: collecting original inductive current data during the operation of a power electronic transformer; performing adaptive segmentation normalization processing on the original inductive current data to generate a normalized inductive current sequence; calculating a wavelet packet energy entropy and a time domain differential entropy of the normalized inductive current sequence, and splicing the wavelet packet energy entropy and the time domain differential entropy into a two-dimensional fusion feature vector; inputting the normalized inductive current sequence and the two-dimensional fusion feature vector into a trained deep learning model to obtain a prediction probability vector; based on the prediction probability vector, the working mode category with the maximum probability value is selected as the recognition result, and the recognition precision of the working modes of the power electronic transformer can be guaranteed under the working conditions of high noise interference or rapid mode switching.

The invention relates to an Internet of Things data man-machine interaction visualization system based on artificial intelligence, which belongs to the field of man-machine interaction technology, industrial Internet of Things and artificial intelligence technology and comprises a multi-modal data acquisition unit, a cognitive load state real-time inference module, a man-machine interaction interface self-adaptive generation module and a closed-loop regulation control module. The multi-modal data acquisition unit is used for acquiring multi-modal data related to the state of an operator in real time; the cognitive load state real-time inference module is used for inferring and generating a probability vector representing the current cognitive load state of the operator based on the acquired multi-modal data; according to the scheme, a complete feedback control loop is constructed by arranging the multi-modal data acquisition unit, the cognitive load state real-time inference module, the human-computer interaction interface self-adaptive generation module and the closed-loop regulation control module.

The invention relates to the technical field of cable defect identification, in particular to a cable fault identification method based on a convolutional neural network. The method comprises the following specific steps: establishing a one-dimensional convolutional neural network model based on a transmission line model to predict a cable fault type and a fault point location; building a test platform to collect data of cables of different models and with a certain length, wherein the data is used for building a data set for model training; training the prediction model by using the data set, and iterating for multiple times until the target accuracy is met; reflected wave data of a to-be-detected cable are collected and input into the trained prediction model, and a fault type probability vector and the distance between the fault position and the initial end of the cable are output through the prediction model. The invention provides a one-dimensional convolutional neural network prediction model fusing multi-scale feature extraction and a time sequence attention mechanism, the model has good effects on cable fault identification and fault position prediction, and a cable fault identification model for different types and different defect positions is realized.

The invention relates to the technical field of geological disaster early warning, in particular to a rainfall classification-based landslide disaster early warning method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining the characteristics of each grid node through collection and grid processing of multi-source data such as geology and rainfall; segmenting rainfall events and extracting morphological and statistical features; machine learning rainfall typing is carried out based on the features, and probability vectors of all rainfall types are obtained; then, constructing a graph structure fusing a space and a geological relationship, and fusing static and dynamic characteristics of grid nodes; inputting the spatio-temporal characteristic graph into a graph neural network, taking a probability vector as a condition signal, adjusting attention weight through learnable mapping, and realizing adaptive spatial information aggregation guided by rainfall typing; and calculating a landslide probability and generating an early warning based on the updated grid node representation, so as to deeply couple rainfall typing and a graph neural network, realize the crossing from a static threshold value to dynamic feature modulation, and improve the early warning accuracy, timeliness and spatial perception capability.

The invention provides a brain wave epilepsy detection method based on Transform. The method comprises the following steps: acquiring EDF data of a multi-channel electroencephalogram from a database, processing the EDF data, generating a CSV file, and performing data preprocessing on the generated CSV file; defining a plurality of machine learning models, and independently training the models on the EEG feature data to obtain the classification performance of the models; predicting a probability vector by using a machine learning model to construct Transform model input data; the attention of the Transform model is used for dynamic training, and a trained Transform fusion model is obtained; and outputting a prediction result, and storing the trained Transform fusion model. According to the method, more efficient model fusion is realized through dynamic fusion, and the accuracy and generalization ability of epilepsy detection are improved.

The invention relates to the technical field of state monitoring, and discloses an actuator mechanical state evaluation method and system based on time sequence characteristics, and the method comprises the steps: carrying out the normalization processing of an original time sequence signal of an execution machine, and obtaining a time sequence signal of the execution machine; identifying a local extreme point in the time sequence signal, fitting an upper envelope line and a lower envelope line of the time sequence signal, and separating the time sequence signal into intrinsic mode components according to the upper envelope line, the lower envelope line and the local extreme point; extracting a mean value, a variance and an energy value of the intrinsic mode component to obtain an initial feature set of the execution machine; removing the redundant features of which the feature correlation is higher than a preset threshold value to obtain an optimized feature set of the execution machine; performing forward propagation on the optimized feature set to obtain a generation state probability vector of a mechanical state category; and generating the working state of the execution machine according to the generated state probability vector. According to the invention, the efficiency and reliability of actuator mechanical state evaluation can be improved.

This application provides a method, system, medium, and device for human-centered video scene reconstruction and separation. The method includes: for a first-person perspective video sequence, initializing a 3D Gaussian set covering the background, hands, and objects based on a priori knowledge of structure recovery from motion; assigning a learnable dynamic category probability vector to each Gaussian point in the 3D Gaussian set; constructing a dedicated deformation branch; according to the learnable dynamic category probability vector of each Gaussian point, assigning each Gaussian point to the dedicated deformation branch for processing through a preset soft-hard two-stage routing mechanism, determining the Gaussian points processed by the dedicated deformation branch; rendering the Gaussian points processed by the dedicated deformation branch to determine the 4D scene reconstruction image and the decomposed reconstruction images of the background, hands, and objects. This application achieves 4D scene reconstruction of human-centered video and explicit, fine-grained separation of the background, hands, and objects.

The invention provides an intelligent production line fault prediction method and system based on random forest enhancement. The method comprises the following steps: generating a production line equipment association map including equipment nodes, connection edges and operation state files; continuously receiving a real-time state signal flow packaged by a self-adaptive time window from the sensing array, and dynamically adjusting the collection granularity according to the equipment fault sensitivity; inputting the state flow into a random forest enhanced classifier into which a fault sample distribution weight is introduced, and performing step-by-step decision traversal to output a membership probability vector set of each equipment node belonging to a preset fault mode; traversing equipment nodes in the atlas, performing breadth-first search by taking a fault mode exceeding a confidence threshold as an activated node, and recording a fault propagation direction sequence and time delay distribution; and finally, coding the propagation sequence and the delay distribution into a fault evolution path diagram, calculating a fault arrival probability and an expected influence range of a propagation path, and generating a prediction and early warning instruction. According to the invention, accurate prediction and propagation traceability of production line faults are realized.

The application provides a kind of early warning method for monitoring abnormal state of gas pipe network, it is related to gas monitoring and early warning field, using the monitoring system including data preprocessing module, space-time feature module, ST-CGN network module, edge intelligent optimization module and result processing module;Among them, the ST-CGN network module outputs the multi-working condition probability vector corresponding to the sliding window slice through the input of three-dimensional feature tensor, at the same time, the space-time convolution layer in the ST-CGN network module simultaneously models the local dependence of space neighborhood relationship and time sliding window, maintains time causality;The edge intelligent optimization module realizes the edge optimization of ST-CGN network module through MAML fast adaptation, knowledge distillation, model pruning and quantization.This method effectively solves the problems existing in the existing gas pipe network monitoring method, such as data missing, easy to be disturbed by noise, insufficient multi-point coupling identification, poor cross-scene generalization ability, monitoring alarm with lagging nature and so on.

In a dialogue system outputting utterances in a dialogue manner, an utterance filtering device for preventing output of possibly problematic expression includes: a pre-trained context model trained, in response to an input of a word vector sequence representing an utterance, to output a probability vector comprising elements indicating probability of each of the words in a prescribed word group appearing in a context in which the utterance is placed; and a determining unit 456 configured to input a word vector sequence representing a subject utterance to the context model, and to determine whether the subject utterance is to be discarded or approved depending on whether a value determined as a prescribed function of a probability vector output by the context model in response to the input is equal to or larger than a threshold value.

The application provides a kind of intelligent production line fault prediction method and system based on random forest enhancement, by generating production line equipment association graph containing device node, connection edge and running state archive;Continuously receive real-time state signal stream encapsulated with adaptive time window from sensor array, and dynamically adjust the collection granularity according to the device fault sensitivity;The state stream input is introduced into the random forest enhanced classifier of fault sample distribution weight, and the membership degree probability vector set of each device node belonging to the preset fault mode is output by traversing output step by step decision;Traverse the device node in the graph, and take the fault mode exceeding the confidence threshold as the activated node to perform breadth-first search, record the fault propagation direction sequence and time delay distribution;Finally, the propagation sequence and delay distribution are encoded into fault evolution path graph, the fault arrival probability and expected impact range of propagation path are calculated, and the prediction warning instruction is generated.The application realizes the accurate prediction and propagation tracing of production line fault.

The present disclosure provides a model training method of a joint data set, an X sheet identification method and related equipment. Different categories of data sets are combined during model training, thereby expanding the sample categories. Then, the network parameters of the feature extraction network corresponding to each category of data set are updated using the common loss value of the combined data set, which is beneficial to considering other categories of data sets when the feature extraction network corresponding to different categories of data sets is used, thereby improving the generality and accuracy of the identification model obtained by training. And according to the preset demand set by the user, the probability vector is analyzed to obtain the to-be-identified result of the to-be-identified X sheet. Not only can the to-be-identified X sheet be accurately identified, but also the actual needs of the user can be met, and the flexibility is high.

The invention discloses an FPGA score prediction method, system and device based on a cascade forest model and a medium. The method comprises the following steps: acquiring original match data, and performing offline training on a cascade forest model by using the original match data to obtain model parameter information; burning the model parameter information on an FPGA (Field Programmable Gate Array) chip; and inputting the real-time match data into the FPGA chip for score prediction to obtain a score prediction result output by the FPGA chip. According to the method, the limitation of a traditional random forest model is broken through, and the probability vector output by each layer and the original input features are spliced and progressively constructed through a multi-layer cascade forest module. By means of the design, more complex high-order features can be extracted layer by layer, and the expression ability and prediction precision of the model are remarkably improved. On the aspect of hardware, through accurate scheduling of the reasoning process and effective utilization of hardware resources, efficient operation on edge equipment with limited resources can be achieved, and meanwhile the requirement for real-time performance is met; and the low-power-consumption and high-efficiency reasoning capability is realized.

The invention discloses an electric power inspection multi-mode large model fine tuning method and device, and the method comprises the steps: carrying out the time series data spectrum preprocessing of a one-dimensional current signal through a Grubrum angle field algorithm, and obtaining a two-dimensional current spectrum; combining the visible light / infrared image and the diagnosis text to construct an instruction fine tuning data set; on the basis of the universal multi-modal backbone model, designing a 1 + N LoRA fine tuning framework comprising a current map encoder and an electrical semantic optimization LoRA module, and obtaining an adjusted model; respectively outputting preliminary diagnosis probability vectors through the visual branch and the current expert branch, and calculating the cosine similarity of the two preliminary diagnosis probability vectors; if the cosine similarity is lower than a preset threshold value or the physical paradox exists, triggering reflection logic; and adjusting the weight of the electrical semantic optimization LoRA module, recalculating a feature fusion result, and outputting a final diagnosis conclusion. The method solves the problems that in the prior art, single-mode sensing is limited, and diagnosis precision is insufficient due to environmental interference.

This invention discloses a method for proactive defense and resilient reconfiguration of distribution networks integrating spatiotemporal graph neural networks, belonging to the field of distribution network cyber-physical security. This invention aims to address the limitations of traditional passive defense methods in responding to spatiotemporally coordinated FDIA attacks and their insufficient reconfiguration resilience. First, a spatiotemporal graph model of the distribution network's CPS is constructed, and a spatiotemporal graph neural network is used to predict the FDIA attack probability vector. Second, an incomplete information stochastic game model is constructed, using the probability vector as the observation input to solve for the optimal defense resource deployment strategy. Finally, a multi-objective resilient reconfiguration model is constructed with the objectives of maximizing source-load spatiotemporal matching and restoring critical loads. A deep reinforcement learning algorithm is used to solve for the optimal switching sequence based on the game situation or physical state. This invention can effectively predict and defend against spatiotemporally coordinated attacks, significantly improving the proactive defense capability and operational resilience of the distribution network.

The invention is suitable for the technical field of DNA site recognition, and particularly relates to a DNA site recognition method and device based on multi-model dynamic integration. The method comprises the following steps: acquiring a DNA sequence; running a first base model by taking the DNA sequence as input to obtain a class probability vector Pj; the first base model is a machine learning model based on a deep forest algorithm, and j is a cascade layer serial number of the first base model; the class probability vector Pj is used as an additional feature dimension, a second base model is operated in a dynamic updating mode, and a prediction result Qp is obtained; the second base model is a machine learning model based on a support vector machine; compared with a traditional support vector machine for dichotomy, the model provided by the embodiment of the invention can provide more accurate confidence information on the basis of a dichotomy result; compared with a traditional machine learning scheme, the scheme provided by the embodiment of the invention adds a certain performance adjustment space.

The invention relates to an interpretable CNN (Convolutional Neural Network) training method and device based on Bernoulli sampling and a medium, and the method comprises the following steps: inputting a picture into a CNN, and obtaining a response feature map of a filter; performing Bernoulli sampling on the response feature map to obtain a binary distribution matrix; calculating a filter average weight matrix of each picture category according to the binarization distribution matrix, and calculating the sum of pairwise differences; calculating the Hadamard product of the distribution vector of the binary distribution matrix and the response feature map to obtain a mask feature map; respectively inputting the response feature map and the mask feature map into a CNN full connection layer to respectively obtain classification prediction probability vectors, and respectively calculating cross entropy loss with a real label; and according to the sum of the pairwise differences and the cross entropy loss, using a stochastic gradient descent method to realize network training, and obtaining an interpretable CNN for image classification. Compared with the prior art, the method has the advantages of high adaptability, high interpretability and the like.

The invention provides a Guangxi action behavior analysis method fusing artificial intelligence and data modeling, and relates to the technical field of action analysis. According to the method, through multi-modal feature extraction and an advanced machine learning algorithm, the accuracy, the real-time performance and the automation level of Guangxi action analysis are remarkably improved. According to the method, video frames and sensor data can be acquired in real time when an embroidery worker executes a wide embroidery operation, and a uniform data stream is realized. Using unsupervised clustering to map the candidate action segments to a stitch atom library, and using a graph convolution-long and short-term memory network and a hidden Markov model to perform feature alignment to generate a stitch probability vector; in addition, according to a stitch recognition result and a scoring index, an embroider skill score is calculated, and an abnormal action is positioned. The method further comprises an incremental learning strategy to continuously optimize the model and provide a scientific and effective skill improvement scheme.

This application discloses a method and system for comprehensive prediction and grading assessment of food safety risks, mainly relating to the field of comprehensive risk prediction technology. It addresses the problems of conventional data preprocessing methods struggling to construct robust feature representations for multi-source heterogeneous data and the failure of most models to effectively integrate prior domain knowledge. The method includes: averaging the dimensionality-reduced feature matrix over the time dimension to aggregate it into a feature vector, thereby obtaining a Laplace score; using the Laplace score and the dimensionality-reduced feature matrix to obtain an enhanced feature matrix; inputting the enhanced feature matrix into a trained model, weighting and aggregating the temporal features in the enhanced feature matrix to obtain a risk-aware context vector; fusing the risk-aware context vector with a preset external condition feature vector through a modulation gating mechanism to obtain a condition-aware decision feature vector; and performing hierarchical risk classification on the condition-aware decision feature vector to obtain a risk level probability vector.

The application discloses a link fault root cause positioning method and device, relates to the technical field of link fault root cause positioning, and comprises the following steps: acquiring a directed graph corresponding to a link topology structure and a state transition matrix corresponding to the directed graph, acquiring an error count increment, an error type and a reporting node; for each error in the error count increment, based on the error type and the directed graph, determining a responsible node set causing the reporting node to generate the error, allocating a responsibility proportion to each responsible node in the responsible node set, and updating elements in the state transition matrix corresponding to the directed graph according to the responsibility proportion of each responsible node; and based on the updated state transition matrix and a node root cause probability vector, obtaining a fault root cause, which can effectively improve the positioning efficiency and accuracy of the link fault root cause.

The invention discloses a softmax calculation method and device for an image classification model, and relates to the technical field of neural networks, and the method comprises the steps: obtaining a classification score vector, obtaining a classification score vector after numerical stabilization processing, carrying out the descending sorting, and obtaining a sorted classification score vector; respectively calculating an approximate index value of each element in the classification score vector, and respectively calculating a contribution degree corresponding to each element; screening according to the values of the contribution degrees to obtain a reserved item set and a perforated item set; respectively calculating the approximate normalization probability of each element, and obtaining a probability vector based on the approximate normalization probability of each element; according to the method, only the reserved items with remarkable contribution degree are subjected to approximate operation, and the perforated item elements with extremely low contribution degree are subjected to simplified processing, so that the number of elements needing to execute index operation is further greatly reduced, and the consumption of a large amount of index calculation is avoided.

Systems and methods are disclosed for predicting a next text. A method may include receiving one or more documents, such as a document associated with a healthcare provider. The document is then processed to generate one or more tokens which are representative of the document. The document is then processed with a machine-learning model, such as a topic model, and a topic vector is output for the document. Based at least in a part on this topic vector, the document is then processed by one or more expert machine-learning models, which each output a probability vector. The various probability vectors are then further processed to calculate a total probability vector for the document. Based at least in part on the total probability vector for the document, a text output is selected.

The invention provides an early warning method for monitoring an abnormal state of a gas pipe network, relates to the field of gas monitoring and early warning, and adopts a monitoring system comprising a data preprocessing module, a spatial-temporal characteristic module, an ST-CGN network module, an edge intelligent optimization module and a result processing module. Wherein the ST-CGN network module outputs a multi-working-condition probability vector corresponding to a sliding window fragment through input of a three-dimensional feature tensor, and meanwhile, a space-time convolution layer in the ST-CGN network module models local dependence of a space neighborhood relation and a time sliding window at the same time, and keeps time causality; and the edge intelligent optimization module realizes edge optimization of the ST-CGN network module through MAML rapid adaptation, knowledge distillation and model pruning and quantification. The method effectively solves the problems that in an existing gas pipe network monitoring method, data are prone to missing, noise interference is prone to occurring, multi-point coupling recognition is insufficient, the cross-scene generalization ability is poor, and monitoring alarm lags.

The invention discloses an exception handling system and method applied to mobile communication equipment, and relates to the technical field of data analysis, and the method comprises the steps: integrating hardware, software and the like, defining exception types, extracting numerical feature vectors, endowing severity coefficients, and constructing an exception feature system; the method comprises the following steps: collecting multi-source data, cleaning, constructing a classification model, strengthening high influence anomaly recognition through a weighted cross entropy loss function, and outputting an anomaly type probability vector; counting the single-exception maintenance duration and satisfaction of the station, and introducing a cooperative coefficient to construct a matrix to evaluate the multi-exception maintenance capability; when the equipment is abnormal, collecting data to generate a feature vector, outputting probability sorting through an SVM, and constructing a candidate abnormal set through threshold screening; and calculating a basic score based on the anomaly probability and the site single anomaly score, fusing the collaborative gain to generate a recommendation score, and outputting a recommended site according to the score. The method can effectively improve the situation that a single exception processing mode in the prior art cannot cope with the multi-fault collaborative repair requirement.