451 results about "Low Confidence" patented technology

The invention belongs to the technical field of target detection and recognition, and particularly relates to a thinking chain-based large and small model collaborative target detection and recognition method. According to the method, the small model is responsible for most of easy-to-detect targets, the calculation pressure of the large model is reduced, the large model is responsible for suspected samples, vision and language multi-mode reasoning is combined, the overall false detection rate and the omission ratio are both reduced, confidence evaluation is conducted through the joint probability, automatic screening and manual rechecking of uncertain results are achieved, the reliability of key results is guaranteed, and the method is suitable for large-scale popularization and application. According to the'pseudo thinking chain + pseudo label 'method, by means of reasoning and labels generated by the model, data dependence on manual labeling is reduced, only low-confidence samples are manually confirmed, the human intervention range is narrowed, the human cost is remarkably saved, and semantic information with finer granularity is provided for the model by introducing phrase-level feature descriptors. And the identification capability of complex target attributes and states is improved.

The invention provides an under-mine video stream dynamic denoising method based on multi-modal fusion, which comprises the following steps: constructing a time sequence synchronous fusion mechanism of visible light, infrared and laser radar data, and realizing time-space alignment of multi-source heterogeneous data; a dynamic noise model is established by introducing a fractional calculus optical flow field concept and combining a Gaussian mixture model, so that a dynamic noise region is accurately identified; an improved self-adaptive wavelet threshold function is constructed, a function threshold parameter can be linked with a dust concentration sensor in real time, and the de-noising intensity is dynamically adjusted according to the actual dust concentration; designing a dual-path feature enhancement neural network to effectively separate and enhance structural features and texture features in the video image; a cascaded detection decision system is created, a lightweight network is used as a primary detector, a high-confidence detection result is directly output, and a low-confidence detection result is input into a Transform correction module for secondary reasoning. According to the invention, dynamic denoising, feature enhancement and target intelligent monitoring of the video stream under the mine can be realized.

The invention discloses an intention recognition method based on cross attention and multi-scale uncertainty. The intention recognition method comprises the following steps: preprocessing multi-modal data; parallel multi-modal feature coding oriented to intention recognition; the invention relates to multi-scale uncertainty perception decoding. According to the method, a parallelized multi-modal feature extraction path is constructed, and a hierarchical fusion mechanism based on cross attention is designed, so that deep semantic alignment and complementary enhancement of four types of heterogeneous information including the posture, the motion track, the global scene and the local vision of a rider are realized; the problems of incomplete feature representation and insufficient cross-modal correlation modeling caused by dependence on a single information source or adoption of a shallow fusion strategy in a traditional method are solved, so that the accuracy and robustness of intention recognition in a complex traffic scene are remarkably improved. According to the method, a multi-scale uncertainty perception decoding framework is introduced, risk early warning or context auxiliary verification is carried out on a low-confidence identification result, and the reliability of an automatic driving system in a safety critical scene is improved.

The invention provides a semiconductor chip automatic test optimization system, and belongs to the technical field of electrical variable measurement. Comprising an acquisition module used for establishing an electrical variable time sequence arranged according to a time sequence, a drift analysis module used for calculating a short-time fluctuation amplitude value, a long-term drift slope and a high-frequency noise amplitude value based on the impedance time sequence arranged according to the time sequence, and a contact judgment module. The evaluation module is used for constructing a contact impedance coefficient based on a short-time fluctuation amplitude value, a long-term drift slope and a high-frequency noise amplitude value in a jth sliding window, and evaluating and optimizing the contact impedance coefficient, and the data marking module is used for marking electrical variable measurement data corresponding to a test channel which is evaluated to be instable in contact as low confidence. According to the system, the ith test channel of the temperature sensor chip can be tested more accurately, the short-time fluctuation amplitude value, the long-term drift slope and the high-frequency noise amplitude value are calculated at the same time, multi-source data participates, and the test accuracy is improved.

The invention provides an image text recognition method and system based on a mask diffusion model, a storage medium and equipment, and belongs to the technical field of image or video recognition or understanding. According to the method, multi-scale visual features of an image are extracted through a visual encoder, a mask diffusion decoder is combined, a diversified mask strategy and random character replacement disturbance are adopted in a training stage, denoising loss and auto-reflection loss are calculated respectively, and a model is optimized in a combined mode; in the reasoning stage, starting from a full mask state, a complete text sequence is recovered through multi-round iterative denoising. According to the method, the one-way modeling limitation of a traditional autoregression model is broken through, all-around context-dependent modeling is achieved, an autoreversion error correction mechanism and a block low-confidence mask strategy are introduced, and the recognition accuracy and reasoning efficiency in complex scenes such as shielding and fuzzy scenes are remarkably improved. The method provided by the invention reaches a leading level on a plurality of public data sets, and has the advantages of high precision and high speed.

The invention relates to an archive credible question and answer method based on graph context retrieval and knowledge graph enhancement, and belongs to the field of computer software and artificial intelligence. According to the method, the relation between key entities and entities in a historical file is extracted by using a multi-modal large language model MLLM, and low-confidence fields are automatically identified to generate marks needing to be manually rechecked, so that the correctness of information extraction is ensured while the workload of manual rechecking is reduced, and then based on a given knowledge base, the information extraction efficiency is improved. Previous related questions and answers are retrieved according to query questions to achieve context enhancement, and finally, credible answers are generated through a large language model (LLM) in combination with extracted information in historical archives. On the premise of keeping the value of the archive voucher, the processing efficiency and the knowledge service capability are remarkably improved.

The invention provides a self-adaptive information retrieval and optimization system based on retrieval enhancement generation, which comprises a prospective answer module used for receiving a query statement and context information and inputting the query statement and the context information into a large language model so as to generate a prospective preliminary answer; the confidence coefficient detection module is used for setting a confidence coefficient score threshold value to carry out keyword-level confidence coefficient evaluation on the content in the prospective preliminary answer, correcting a low-confidence-coefficient part in the answer according to an evaluation result, and outputting a complete answer content as a confidence coefficient answer; the knowledge graph extension query module is used for executing named entity recognition operation on the confidence answer and obtaining an extension query answer in combination with a pre-constructed small knowledge graph; and the context memory pool module is used for storing the context information input for the first time and related information fragments obtained in each round of retrieval process, preferentially performing information retrieval from the memory pool in the subsequent retrieval process, and accessing the external index database only when no matched content exists in the memory pool.

The invention discloses a data scarcity industrial vertical field-oriented knowledge-guided retrieval enhancement generation method, belongs to the technical field of natural language processing and industrial intelligence crossing, and can improve the retrieval accuracy and generation reliability of a large language model in an industrial scene. According to the method, a'sparse vector + dense vector 'mixed knowledge base is constructed, and general knowledge and industrial field texts are fused; after the model generates an initial text, judging whether external retrieval is needed or not through multi-dimensional evaluation token confidence; extracting attention weights for the low-confidence tokens, and screening key tokens to generate a retrieval query; dynamically adjusting the weight of a retriever based on a BGE-M3 model, and optimizing a retrieval result through reordering; the retrieval knowledge is converted into a context with an index, and a prompt template is constructed to generate a correction value iteration calibration text; and finally, optimizing the text format, and generating a structured response meeting industrial requirements. The method solves the problems of lack of professional knowledge of large language models in the industrial field, shallow retrieval and generation fusion, lack of knowledge calibration mechanisms and the like.

The invention provides a bridge structure vibration long-term monitoring algorithm and system based on computer vision, and relates to the technical field of beam monitoring. In the prior art, there is no vision-based bridge monitoring algorithm with high precision, strong robustness and shielding self-recovery capability. According to the method, matching point confidence is calculated, and matching points with low confidence scores are removed; calculating the displacement of the bridge at the plurality of feature points, and generating vibration time sequence data; the inherent frequency and the corresponding modal shape are extracted and compared with the reference frequency range and the reference modal shape, the monitoring result is obtained, shielding self-recovery is carried out according to the environmental interference influence, the precision and reliability of bridge monitoring are improved, and the robustness of bridge monitoring activities in a complex scene is improved.

The invention discloses a positioning calibration method of surface mounting equipment, equipment, a medium and a product, and relates to the technical field of printed circuit boards, a first position of a bonding pad on a circuit board and a second position of an element pin of an element to be mounted are obtained through recognition by using an image recognition model, the confidence coefficient is recorded, and when matching calculation is carried out, the positioning calibration accuracy is improved. And carrying out position transformation calculation on the positions of the bonding pad and the element pin matched in the current iteration, calculating a transformed position error in combination with the confidence coefficient, and outputting a final matching result of the bonding pad and the element pin and a corresponding final space mapping parameter after a first iterative calculation condition is met. According to the method, the influence of a low-confidence-coefficient position detection result on space mapping parameter calculation is effectively reduced, the stability and reliability of the final space mapping parameter are ensured, compared with traditional position error calculation, the method has higher anti-interference capability, and the accuracy of positioning matching of high surface mounting equipment is improved.

The invention relates to the technical field of background investigation, and particularly discloses a talent background investigation method based on multi-source data evaluation, which comprises the steps of constructing a dynamic talent knowledge graph, performing semantic alignment processing on multi-source cross-modal data, evaluating data quality based on a multi-dimensional confidence model, and constructing a confidence closed-loop correction mechanism. Performing intelligent correction on the low-confidence data, and reversely updating the credibility weight of the data source according to a correction result; an event-driven incremental update architecture is adopted, and dynamic events of multi-source data are captured in real time through a distributed event awareness engine. According to the method, the dynamic knowledge graph is used as a unified carrier, and the multi-source data is upgraded from fragmented storage to a semantic association network, so that the talent background survey data is converted from static piling to a dynamically evolved knowledge network, and the accuracy and timeliness of evaluation are remarkably improved.

The invention discloses a character recognition method and system based on a large model and an OCR technology, and relates to the technical field of character recognition, and the method comprises the steps: extracting picture information, carrying out the unified preprocessing, generating a text detection box of a character region through a DBNet lightweight text detection model, and obtaining a text position; quickly identifying characters in the textbox by using a lightweight OCR model to obtain text content, and generating an identification information group; carrying out average value calculation on the confidence coefficient of the lightweight OCR model result, and analyzing the overall confidence coefficient; and for the result with low confidence coefficient, inputting the corresponding identification information group into the multi-modal large model, and carrying out secondary identification. According to the method, the text position set and the text character string sequence are combined into the identification information group, so that the effect of structured storage of detection and identification results is achieved, subsequent information retrieval and multi-modal fusion analysis are facilitated, and the effect of optimizing the subsequent processing efficiency and precision is achieved through the combination of the confidence coefficient screening step.

The invention provides a blind area early warning method and device and a storage medium, and the method comprises the steps: carrying out the multi-dimensional recognition of a vehicle road scene, obtaining the data of a plurality of sensors, determining the environment type of the road scene where a vehicle is located, adjusting the original weight of each sensor based on the environment type, obtaining a first weight corresponding to each sensor, and carrying out the adjustment of the first weight. Through recognition of environment types, weights are dynamically distributed based on indexes related to environment scenes in a non-linear mode so as to adapt to complex scenes. The sensor data is fused according to the first weight, the influence of the data source with low confidence on the fusion processing is dynamically reduced, blind area early warning is performed based on the fusion result, and the false alarm rate is effectively reduced. In addition, according to dynamic weight adjustment of the indexes related to the environment, the response speed of system execution control is increased.

The invention relates to the technical field of sound recognition, in particular to an optical fiber sensing voiceprint feature analysis model construction method of a composite neural network, and the method comprises the steps: collecting a sound signal through distributed optical fiber sensing, setting a sound frequency amplitude threshold value, and extracting a sound signal of an abnormal interval; carrying out noise reduction processing on the collected signals and extracting voiceprint features; constructing an unsupervised model to calculate an outlier score value, and combining a threshold value to judge whether the signal is an abnormal signal; carrying out principal component analysis on the abnormal signal, and carrying out feature compression and mapping; a composite neural network model is trained based on the extracted features, and multi-class voiceprints are recognized; and an uncertainty evaluation mechanism is introduced, a classification result is dynamically adjusted according to confidence, low-confidence data is marked as unidentified and collected and classified again, and self-learning and iterative optimization of voiceprint recognition are realized. The method provided by the invention has high precision, high robustness and strong generalization ability, and is suitable for optical fiber voiceprint event recognition in a complex environment.

The invention relates to a multi-spectral diagnosis result credibility evaluation method. The method comprises the following steps: synchronously acquiring three-spectral images and environmental parameters of target equipment; the collected three-spectrum image is preprocessed; inputting the preprocessed image into the improved model to obtain a prediction probability distribution vector of fusion features and three-spectral-band branches; obtaining a prediction category and uncertainty entropy; carrying out confidence score calculation; judging whether to trigger a feedback condition; according to the method, the complementary characteristics of three spectrums of visible light, infrared light and ultraviolet light are utilized, and the CBAM attention module is introduced, so that key characteristics under different spectrums are enhanced, and the sensing ability of the model to an abnormal region is improved; the method achieves the prediction result confidence evaluation based on a Monte Carlo Dropout reasoning mechanism, effectively recognizes a low-confidence result, gives a diagnosis conclusion, can also evaluate the confidence, is suitable for deployment in key scenes such as a high-pressure valve hall, and facilitates the improvement of the overall safety and stability of a system.

The invention relates to the technical field of power equipment fault detection, in particular to a power equipment fault diagnosis method and system based on a large power model, and the method comprises the steps: converting the multi-modal data of a power cable in operation into a feature vector, and inputting the feature vector into a pre-trained fault diagnosis model to obtain a preliminary diagnosis result and confidence; if the confidence coefficient is not lower than the threshold value, the preliminary diagnosis result is reserved; and if the confidence coefficient is lower than a threshold value, taking the feature vector as a current potential fault feature vector to perform secondary discrimination, namely obtaining a significance index by calculating the similarity and volatility of the current potential fault feature vector and a historical potential fault feature vector, analyzing a time change trend to obtain a cumulative trend index, and performing secondary discrimination on the cumulative trend index. And comprehensively determining a potential fault index through the significance index and the cumulative trend index, and determining a final diagnosis result according to the potential fault index. According to the scheme, the diagnosis accuracy of the fault diagnosis model on low-confidence potential faults and unknown faults is improved.

The invention discloses a source domain irrelevant cross-domain cardiac beat recognition method and system for pseudo-label mining, and relates to the technical field of pseudo-label learning, and the method comprises the steps: obtaining electrocardiogram data with cardiac beat labels in a source domain, inputting a pre-established source domain model for pre-training, and obtaining a pre-trained source domain model; acquiring unlabeled electrocardiogram data of a target domain, and screening and classifying the data of the target domain based on a pre-trained source domain model and a preset category threshold to obtain data with high false label confidence and data with low false label confidence; using the source domain model to initialize target domain model parameters, revising a strategy based on a pseudo label of local and global semantic perception, updating the pseudo label of the low-confidence data, and combining the high-confidence data to obtain updated pseudo-labeled target domain data; target domain data subjected to data augmentation pseudo labeling are input into a target domain model, a target domain model optimization total loss function is calculated, and cross-domain cardiac beat intelligent recognition irrelevant to a source domain is achieved.

The invention relates to the field of combinatorial optimization and artificial intelligence, and discloses a group logistics transportation scheduling method and system based on a role interaction graph neural network, and the method comprises the following steps: S1, dividing agent nodes and position nodes, and generating initial features; s2, iteratively updating node embedding by using a multi-channel attention mechanism of a graph neural network; s3, generating a delivery point distribution probability based on node embedding, and determining an initial distribution scheme; s4, local redistribution optimization is performed on the delivery points with low confidence distribution; and S5, performing parallel path planning on the optimal scheme, and outputting a result for reinforcement learning feedback. In the invention, through modeling of a graph neural network multi-channel attention mechanism on a complex interaction relationship and a synergistic effect of local redistribution optimization and parallel path planning, a second-level generation of a high-quality scheduling scheme is realized, cross-scale scene migration of the model is achieved, and group logistics transportation scheduling efficiency and robustness are improved.

The invention provides a vehicle target automatic labeling method and system based on deep learning. The method comprises the steps of obtaining video stream monitoring data collected by a road camera; based on the video stream monitoring data, utilizing a pre-trained vehicle identification model to automatically label a vehicle target in the video stream monitoring data to obtain a preliminary labeling result; performing uncertainty evaluation on the preliminary labeling result through an adversarial sample generation strategy to obtain an uncertainty score; when the uncertainty score does not exceed a preset threshold value, taking the preliminary labeling result as a final labeling result of the vehicle target; according to the invention, by using the pre-trained convolutional neural network model, large-scale video stream data can be quickly processed and preliminary annotation can be completed, so that the dependence on manual annotation is reduced; uncertainty evaluation is carried out on the preliminary labeling result through an adversarial sample generation strategy, errors or low-confidence-coefficient areas possibly existing in the labeling result can be effectively recognized, and the accuracy of the final labeling result can be guaranteed.

The invention belongs to the technical field of radar signal processing, and particularly relates to a low-confidence small target detection method in radar echoes based on a hybrid architecture, and the method comprises the steps: firstly carrying out the spectrum symmetric movement and dimension recombination of radar echo data, and generating five-dimensional tensors [B, T, C, H, W] containing time sequence features; then the tensor is input into a detection model formed by cascading a Hurglass 3D module and a YOLOv8 network, the Hurglass 3D module extracts multi-scale spatial-temporal features through a structure of three-dimensional convolution down-sampling, bottleneck layer and three-dimensional transposition convolution up-sampling, and feature fusion is achieved through jump connection; and finally, target detection is completed through a backbone network, a neck network and a decoupling detection head of the YOLOv8 network. According to the invention, through spatio-temporal feature combined extraction and small target feature enhancement, the detection accuracy and the positioning precision of the low signal-to-noise ratio small target in radar echoes are effectively improved.

The invention discloses a self-adaptive text extraction method and system based on artificial intelligence, and the method comprises the steps: carrying out the analysis of the document structure entropy of an example document set, quantifying the noise density, geometric distortion degree and background complexity of the example document set, and carrying out the self-adaptive selection of a preprocessing assembly line intensity grade according to the above; dynamically configuring image preprocessing parameters and AI recognition model parameters, and generating a recognition engine instance to output a preliminary recognition text; after regularized coarse screening extraction is carried out based on key field description, a multi-candidate generation strategy is started for low-confidence-coefficient candidate text fragments, a multi-person cooperative verification process is triggered for lower-confidence-coefficient fragments, finally all the fragments are processed through a text standardization module, and structured text extraction information is output. According to the method, accurate adaptation of processing intensity is achieved through document quality quantitative evaluation, the extraction accuracy and system robustness of complex heterogeneous documents are effectively improved through a multi-level confidence coefficient verification mechanism, and the identification error risk caused by image quality fluctuation or rule solidification is reduced.

The invention provides a two-stage metal surface defect detection method based on computing power perception, which is characterized in that improved U-Net and YOLOv8 are cascaded, the method is mainly used for metal surface defect detection, and aims to solve the problems of low precision, insufficient computing resource utilization rate, high detection efficiency and the like in small target defect detection in the prior art. And the detection efficiency and the accuracy are difficult to balance in a complex environment. The operation process of the method comprises the following steps: firstly, carrying out rough detection on a target by utilizing a YOLOv8 model; and then, according to the confidence of the coarse detection result and the current computing power of the equipment, designing a self-adaptive high-low confidence threshold mechanism, and grading the coarse detection result according to the mechanism. Experimental results show that the method remarkably improves the detection precision of the small target defect sample, optimizes the allocation of computing resources, effectively balances the detection efficiency and accuracy, and can be well suitable for a complex industrial detection environment.

The invention discloses a multi-round cross validation and weak supervision noise cleaning method based on a large language model. According to the method, weak supervised learning and large language model reasoning capability are fused, and multi-round cleaning and label optimization are performed on a low-confidence sample by introducing a small amount of high-quality label seeds and combining automatic rule construction and cross validation processes. According to the method, a verification and feedback system with a large language model as a core is constructed, efficient purification and enhancement of weak tags in large-scale text data are achieved, and high-quality training data and intelligent tag optimization support are provided for natural language processing tasks such as relation extraction, text classification and entity recognition.

The invention relates to the technical field of AI vision, in particular to a sesame seed candy forming defect real-time detection method and device based on AI vision. The method comprises the following steps: respectively collecting multimode images of qualified sesame seed candies, generating a qualified characteristic fingerprint set, and calculating sugar body light transmission uniformity and sesame adhesion density as a domain parameter set; constructing a probability distribution model, and setting an anomaly judgment threshold value and a domain parameter early warning threshold value; collecting a multi-mode image of the to-be-detected sesame seed candy in real time, extracting a to-be-detected feature fingerprint, and calculating the light-transmitting uniformity of a to-be-detected candy body and the sesame adhesion density; calculating a comprehensive abnormal score, and judging a defect; and capturing a low-confidence sample based on the comprehensive anomaly score, obtaining an artificial correction feedback sample, and updating a probability distribution model, an anomaly judgment threshold value and a domain parameter early warning threshold value by utilizing the feedback sample through online incremental learning. According to the invention, the detection cost of a high-yield production line can be reduced, and the model deployment period is shortened.

The invention belongs to the technical field of intelligent navigation situation sensing, and particularly relates to an intelligent navigation situation multi-dimensional sensing terminal, which generates a forward sensing domain covering a water surface and an underwater area in a preset space-time range in front of a route based on a target ship route trajectory, identifies candidate obstacles in the domain through multi-sensor fusion, and obtains a target navigation situation. Constructing an obstacle original data set containing position, size and medium type, performing confidence cross validation with collaborative perception data of ships in an overlapped perception domain, removing low-confidence candidate obstacles, generating a verified obstacle list, and combining real-time draught, attitude and tide data of the ships to obtain an obstacle data set; a three-dimensional electronic fence is constructed for each obstacle in the list, whether the collision risk exists in the navigation relative to the target ship or not is judged, finally, the electronic fences and the risk state are superposed to an electronic chart for visual display, high-reliability environment sensing data are provided for the autonomous navigation of the intelligent ship, and the obstacle avoidance efficiency in the navigation environment is remarkably improved.

The invention relates to the technical field of transformers, and particularly discloses a transformer fault diagnosis method and system.The transformer fault diagnosis method comprises the steps that firstly, concentration data of gas dissolved in transformer oil is collected, the potential fault type of a transformer is preliminarily determined based on a three-ratio method, and a diagnosis result confidence evaluation mechanism is introduced; a confidence score for the potential fault type is calculated. And when the confidence coefficient is lower than a preset threshold value, further combining the potential fault type and the confidence coefficient thereof with basic attribute data of the transformer to form fault diagnosis priori data. And then, deep conjoint analysis is performed on the concentration data of the gas dissolved in the oil and the fault diagnosis prior data by using the trained fault diagnosis model, so that more comprehensive fault features can be captured, and an accurate transformer fault diagnosis result can be obtained. According to the method, a deep learning algorithm is introduced during low-confidence diagnosis of a traditional three-ratio method, so that the accuracy and reliability of transformer fault diagnosis can be remarkably improved.

The invention discloses a semi-automatic labeling method and system for rail transit engineering construction video images, and the method comprises the steps: removing redundant frames from a video stream, extracting key frames, and forming a block index set; and driving the multi-modal large model to pre-annotate the image by using the constructed cue word, and carrying out binarization processing on an annotation result. Based on active learning, migrating unmarked samples to a marked set for multiple times and training a key sample classifier CD until the scale of the marked set reaches the standard; meanwhile, a confidence classifier CC is trained by comparing manual and pre-labeling results. And for residual samples in the unmarked set, after the residual samples reach the standard through a CC precision test, marking tasks are divided according to a threshold value theta: high-confidence samples are pre-marked, and low-confidence samples are manually marked. And finally, updating the set of the manually labeled samples, and determining whether to retrain the model or not according to category distribution. On the premise of ensuring the labeling quality, the blindness of manual labeling is effectively reduced, and the efficiency of the labeling process is improved.

The invention discloses a news industry classification method and device based on large language model active learning, relates to the technical field of text classification, and can remarkably reduce the manual annotation cost while ensuring the news text classification precision. According to the scheme, the method comprises the following steps: calling a large language model to classify each news text for multiple times for a plurality of news texts to obtain at least one corresponding tag; based on a majority voting mode, dividing the plurality of news texts and the corresponding labels into high-confidence samples and low-confidence samples; performing label labeling on the low-confidence sample to obtain a labeled sample; and taking the labeled sample and the high-confidence sample as a training set, carrying out iterative training on the classification model by using the training set until the model precision of the classification model meets a preset condition, obtaining a target classification model, and carrying out prediction classification on the news text by using the target classification model.

The invention discloses a side slope stability evaluation method based on side slope monitoring equipment, which comprises the following steps: acquiring side slope measurement data, and constructing a reconstruction error, a classification probability and a displacement trend based on the side slope measurement data; three indexes are mapped to the same two-dimensional plane, an equilateral triangle is constructed, and intuitive and explainable joint measurement of heterogeneous information is realized through a geometric fusion strategy of quantifying the overall risk by gravity center shift: normalization and confidence correction are carried out on three types of data including reconstruction errors, state classification probability and displacement trend; the interference of dimensional difference and low-confidence observation is suppressed, and meanwhile, the offset weights corresponding to the reconstruction error, the state classification probability and the displacement trend are calculated, so that the relative dominance of three abnormal factors, namely the reconstruction error, the state classification probability and the displacement trend, on the overall risk can be automatically reflected; and the stability score of the slope stability triangle is calculated by taking the health center as a reference, so that the slope stability evaluation precision is improved.

The invention provides an end-side adaptive document structure understanding method, which comprises the following steps of: uniformly rendering and normalizing a to-be-analyzed document, and outputting a page-level pixel grid and basic metadata; executing lightweight layout analysis and region classification to obtain a bounding box, a reading sequence and a region type label of each region in the page; each document area is routed to a corresponding special analysis channel for parallel analysis, and each analysis channel outputs a structured intermediate result and confidence; performing consistency verification and completion reasoning on intermediate results output by each channel, and generating a traceable verification evidence chain for low-confidence fragments; all channel results after verification are fused into a unified document-level structured output; and for a new document type or a continuous low-confidence mode, starting an adaptive process of a parameter efficient fine tuning technology to generate a channel-level increment weight packet, and updating model parameters of an analysis channel. The method has the beneficial effect that parallel accurate analysis of different elements such as tables, formulas, texts and the like can be realized.