CN-122020556-A - Modularized intelligent analysis method for suspension state of overhead contact system

CN122020556ACN 122020556 ACN122020556 ACN 122020556ACN-122020556-A

Abstract

The invention discloses a modularized intelligent analysis method for a suspension state of a contact net, and relates to the technical field of intelligent operation and maintenance of rail transit. The method comprises the steps of firstly, performing quality primary screening on an original image sequence based on illumination, ambiguity and coverage indexes through an image pre-screening module. And then, the key component accurate positioning module utilizes a customized two-stage cascade deep learning network to carry out high-precision analysis on the effective image, and outputs structural data such as component types, bounding boxes, central lines, scale factors and the like. The multi-mode state judging module fuses the visual judging result based on the support vector machine and the non-visual sensing data corrected by the physical model, and carries out comprehensive judgment according to the dynamic weighting of the line section type. The method solves the problems of poor reliability and insufficient adaptability caused by uncontrolled data quality, coarse granularity of feature extraction and lack of physical modeling in fusion in the prior art, and realizes intelligent analysis of the contact network state with high robustness and high accuracy.

Inventors

ZHANG DONGBO
WANG QI
Qian Pingshen
SONG XIAO
ZHENG JUN
HE SHIDONG
CHEN RUI
ZOU HONGLIANG
LI ZHUO
WEN ZHU

Assignees

中国铁路沈阳局集团有限公司长春高铁基础设施段

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (10)

1. A modularized intelligent analysis method for a suspension state of a contact net is characterized by comprising the following steps: the method comprises the steps of performing imaging quality primary screening and effective area judgment on an original image sequence acquired by a track inspection platform through an image pre-screening module, and outputting a binarized image effectiveness label; Receiving an image with an effective label through a key component accurate positioning module, performing high-precision target detection to separate a core suspension component in the overhead line system, and outputting structured data containing a component type identifier, two-dimensional bounding box coordinates, a center line multi-section line description and local scale factors; Receiving the structured data through a multi-mode state judging module, merging the synchronously acquired non-visual sensing data, comprehensively judging the physical state of each component, and outputting a state judging result comprising a component identifier, a state type and a merging confidence value; And carrying out uncertainty quantification on the state judgment result through a defect suspected degree evaluation and feedback module, and generating a feedback signal for closed loop optimization of the system based on the quantification result.
2. The modularized intelligent analysis method for the suspension state of the overhead line system according to claim 1, wherein the image pre-screening module calculates three quality indexes, namely an illuminance index, a high-frequency energy attenuation ratio and a space coverage integrity index; The illumination index is calculated by adopting a 5×5 pixel sliding window on the green channel of the image, and the illumination is judged to be insufficient when the average pixel value of the full-image effective area on the green channel is lower than 30; The high-frequency energy attenuation ratio is calculated by using discrete cosine transform to count DCT sub-band energy duty ratio with row-column index greater than 8, when the duty ratio is lower than 15%, the motion blur is judged to be out of standard, the space coverage integrity index is calculated by using characteristic point matching of the current image frame and the standard scene template, the intersection ratio and the mass center offset are calculated, when the intersection ratio is lower than 0.7 or the mass center offset is 5% higher than the image width, the region is judged to be missing, and when any index does not meet the preset condition, the invalid label is output.
3. The modularized intelligent analysis method for the suspension state of the overhead line system, which is disclosed in claim 1, is characterized in that the accurate positioning module of the key component adopts a two-stage cascade neural network architecture, the first stage is an improved regional suggestion network, the anchor frame of the improved regional suggestion network is customized according to the aspect ratio distribution of typical components of the overhead line system and draws attention, the second stage is a fine tuning positioning network, the backbone network of the improved regional suggestion network integrates three-stage stacked deformable convolution layers, each stage of offset is predicted by an independent lightweight sub-network, and the sub-network input is the splicing of a characteristic diagram of the previous stage and regional suggestion characteristics.
4. A modular intelligent analysis method for suspension status of overhead line system according to claim 3, wherein the output head of the fine-tuning positioning network comprises four parallel branches: A first branch output component class probability distribution; The second branch returns to four-corner coordinates of the two-dimensional surrounding frame; The third branch generates a central line multi-section line vertex sequence formed by 21 control points through point-by-point regression, the head and tail points are constrained to the geometric centers of the two ends of the component, the middle point is obtained through interpolation, and each point coordinate is represented by an image normalization coordinate system; The fourth branch estimates a local scale factor, which is defined as the ratio of the projected length of the target part in the image to its standard geometric model length, scaled with reference to the hanger standard length of 1.2 meters.
5. The method for modular intelligent analysis of a catenary suspension state according to claim 1, wherein the multi-modal state discrimination module synchronously acquires non-visual sensing data acquired by a mechanical sensor array, a temperature sensor and an environmental weather station deployed on a catenary strut and a catenary, and performs temperature compensation on the mechanical sensing data first.
6. The method for modular intelligent analysis of overhead line system hanging state according to claim 5, wherein in visual domain, calculating three geometric features of offset, curvature integral and end point interval deviation based on centerline multi-segment line description, inputting a support vector machine classifier adopting radial basis function kernel, and outputting visual confidence coefficient covering normal, offset, relaxation and fracture states; in a physical domain, comparing the corrected tension value with the lower limit of a failure threshold corresponding to the type of the component, and calculating a mechanical abnormality index, and judging that the mechanical abnormality index is abnormal when the index is smaller than 0.8; The lower limits of tension safety thresholds of the hanging string, the positioner and the wrist arm are 8kN, 12kN and 25kN respectively, and the thresholds are dynamically adjusted along with the ambient temperature at a slope of-0.1 kN/°C.
7. The modularized intelligent analysis method for the suspension state of the overhead line system according to claim 6 is characterized in that the fusion weight of the visual domain and the physical domain evidence is determined by line section types, the line section types comprise straight line sections, small-radius curve sections, turnout sections and bridge tunnel transition sections, each section type corresponds to a pre-trained weight mapping function, the function takes section geometric parameters and historical fault statistics as input, visual weight is output, physical weight is the complement value of the function, the weight mapping function is realized through an off-line trained multi-layer perceptron, the input dimension is 7, and the output is a single visual weight value.
8. The modularized intelligent analysis method for the suspension state of the overhead line system is characterized in that a Bayesian inference engine is built in the defect suspected degree evaluation and feedback module to define two states to be evaluated of a normal state and a defect, the priori distribution of the states is calculated by the occurrence frequency of various states in historical data, likelihood functions of the states represent the probability that the current fusion confidence value is observed in a given state, and the fusion confidence value in the similar historical state is obtained by modeling through a kernel density estimation method.
9. The method for modular intelligent analysis of catenary suspension state according to claim 8, wherein the entropy of the posterior probability distribution is calculated as an uncertainty measure by Bayesian updating, and the entropy is determined to be "data insufficient and unable to be judged" when the entropy is higher than 1.2 bits, and otherwise is determined to be "determine no defect" or "have defect".
10. The method for modular intelligent analysis of catenary suspension state according to claim 9, wherein the defect suspected level evaluation and feedback module performs correlation analysis on uncertainty metrics and quality indexes of the image pre-screening module, identifies dominant quality factors causing high uncertainty, and generates targeted front-end acquisition parameter adjustment instructions or sensor calibration instructions based on the analysis.

Description

Modularized intelligent analysis method for suspension state of overhead contact system Technical Field The invention relates to the technical field of intelligent operation and maintenance of rail transit, in particular to a modularized intelligent analysis method for a suspension state of a contact net. Background The state of the overhead contact system of the electrified railway directly determines the current-carrying quality and the driving safety of the overhead contact system. Intelligent monitoring technology based on computer vision and multi-sensor fusion has become the development direction of operation and maintenance systems. However, the current mainstream technical scheme has deep limitations on the architecture design and physical modeling level, so that the reliability, accuracy and adaptability of the current mainstream technical scheme in a complex real scene are difficult to meet the operation and maintenance requirements of a high-grade railway. The concrete steps are as follows: The closed loop loss of data quality, existing systems generally lack effective quality assessment and screening mechanisms for raw perceptual data (especially images). In the process of inspection, a large amount of low-quality and even invalid data is necessarily generated under the conditions of motion blur, intense illumination change, short-term deviation of a target area from the field of view and the like. In the prior art (such as CN120431513 a), although the defocus frame is removed in the later stage, the low-quality data cannot be prevented from entering the core analysis flow, so that the calculation resource is wasted and the analysis result is systematically polluted, and the inherent problems of 'garbage in and garbage out' are formed. Coarse granularity of state sensing information, and the state core of a catenary suspension component (such as a dropper and a positioner) is embodied as the deformation (such as deflection, bending and relaxation) of the geometric form of the central line. The existing visual detection method mostly uses a general target detection framework, and only outputs a two-dimensional bounding box (such as CN 120431513A) or a macroscopic integral deformation function (such as a displacement function of CN 118797532A). The bounding box cannot describe continuous deformation of a one-dimensional structure, and a macroscopic function cannot locate the microscopic state of a specific part, so that fundamental information loss exists in the geometric features on which subsequent discrimination depends. Multimodal fusion is physically disjointed. In order to improve the confidence, it has become a trend to introduce sensors such as mechanics, temperature and the like for multi-mode fusion. The existing fusion strategy is mostly shallow data stacking or feature splicing. For example, CN118797532a maps heterogeneous data directly to health index through complex mathematical transformation, but its transformation process lacks explicit physical semantics, CN120431513a couples visually deduced physical quantity with image features simply. Neither can explicitly model and compensate for known physical disturbances such as ambient temperature that significantly changes contact line tension through thermal expansion effects. If such deterministic physical relationships are not corrected, the original sensing data are directly fused, which necessarily results in misjudgment of the state, and the fusion effect may be even worse than that of a single mode. The existing schemes are all open-loop systems, and output a state judgment result, but the reliability of the result cannot be estimated. The system cannot distinguish between "determining no defect" and "failing to judge due to poor quality of data", and is difficult to support high-confidence operation and maintenance decisions. Meanwhile, the system does not have the capability of reversely tracing the problem source according to the analysis result and dynamically optimizing the front-end sensing strategy, and cannot adapt to the continuously-changing complex environment. Disclosure of Invention The invention aims to provide a modularized intelligent analysis method for a suspension state of a contact net, which solves the fundamental technical problems that in the existing intelligent monitoring system of the contact net, the analysis result is unreliable and cannot be optimized in a self-adaptive manner under a complex real scene due to the defects of architecture coupling and physical modeling. In order to solve the technical problems, the invention adopts the following technical scheme: a modularized intelligent analysis method for a suspension state of a contact net comprises the following steps: the method comprises the steps of performing imaging quality primary screening and effective area judgment on an original image sequence acquired by a track inspection platform through an image pre-screening module, and outputting a