Sliding contact line fault maintenance and prevention method based on visual inspection

Abstract

The invention relates to the technical field of sliding contact line visual inspection, and discloses a sliding contact line fault maintenance and prevention method based on visual inspection. According to the invention, on the basis of collaborative acquisition of terahertz imaging, multispectral vision and optical fiber sensing, conductor internal crack, surface wear and stress distribution information is synchronously obtained, and multi-modal data acquisition is realized; by combining a 3D convolutional network and a physical information neural network, accurate prediction of crack propagation and wear evolution is realized, and a result is ensured to accord with a material mechanics law; the dynamic threshold adjustment mechanism optimizes early warning logic according to real-time working conditions, and enhances the environmental adaptability of the system; grading early warning strategies are linked with equipment control, gradual protection is formed from speed reduction operation to emergency shutdown, and power supply continuity is guaranteed to the maximum extent. According to the method, the omission ratio and the false alarm rate are effectively reduced, the service life of the sliding contact line is prolonged, and the non-planned shutdown loss is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of visual detection of slide wire, in particular to a slide wire fault maintenance and prevention method based on visual detection. BACKGROUND

[0002] Currently, slide wire fault detection mainly relies on manual inspection and contact sensing technology. Manual inspection needs to interrupt the operation of the equipment, and the conductor surface state is checked by visual inspection or simple tools, which is low in efficiency and cannot capture dynamic abnormalities (such as instantaneous sparking of the current collector shoe or expansion of micro-cracks). Contact sensors (such as resistance strain gauges and eddy current probes) indirectly infer the damage degree by measuring the changes of conductor resistance and inductance, but there are obvious defects: first, the sensor needs to directly contact the slide wire, and long-term friction can easily cause its own wear, and the false positive rate increases in a dusty and oily environment; second, only local point data can be obtained, and it is difficult to quantify key indicators such as wear area distribution and crack three-dimensional deformation; third, the mapping relationship between the detection signal and the real physical damage is ambiguous, and it is impossible to establish an accurate life prediction model.

[0003] The existing technical system has systematic deficiencies: the detection means is single, and there is a lack of collaborative analysis of internal defects and surface damage; the fault is judged by relying on empirical threshold, without considering the dynamic coupling effect of material performance degradation and environmental parameters (temperature, humidity); the early warning mechanism is disconnected with the equipment control, and it is impossible to realize the whole-chain closed-loop management from "detection-diagnosis-response". These defects result in high missed detection rate and maintenance cost of the traditional method, and restrict the reliability improvement of large-scale mobile power supply system, and therefore a slide wire fault maintenance and prevention method based on visual detection is proposed. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a slide wire fault maintenance and prevention method based on visual detection to solve the problems in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a slide wire fault maintenance and prevention method based on visual detection, comprising the following steps:

[0006] Step one: multi-modal data synchronous acquisition: deploy terahertz imaging module, multi-spectral array camera and distributed optical fiber sensor, wherein:

[0007] The terahertz imaging module scans the internal layer crack defects of the conductor with 0.1-1THz pulse;

[0008] The multi-spectral camera synchronously captures the current collector shoe motion trajectory in the visible light band (400-700nm) and the conductor surface wear in the near-infrared band (850-1700nm);

[0009] The optical fiber sensor monitors the conductor axial stress fluctuation with a 5cm interval;

[0010] Step two: multi-physics coupling analysis: input step one data into the hybrid physics-data driven model, execute:

[0011] Based on 3D convolutional neural network (3D-CNN) to extract the three-dimensional diffusion rate vector V of the conductor crack crack ; Fusion fiber stress data to calculate crack-stress coupling factor α =‖V crack ‖·K t (K t is the stress concentration coefficient);

[0012] Superpixel segmentation is performed on the near-infrared image, and the depth gradient Δh and area ratio β of the wear groove are extracted;

[0013] Through the synergistic evolution trend of physical information neural network (PINN) α and β, the loss function is embedded in Archard wear equation and Paris crack propagation law to predict the future T period of synergistic evolution trajectory;

[0014] Step three: trigger hierarchical early warning when any of the following conditions is met:

[0015] Crack-stress coupling factor α exceeds the material fatigue threshold;

[0016] One of the wear area ratio β and the wear depth gradient Δh exceeds the threshold;

[0017] The synergistic growth rate of α and β exceeds the preset synergistic rate threshold.

[0018] Preferably, the hybrid physics-data driven model in step two comprises:

[0019] Data-driven module: adopt parallel double-branch network to process crack and wear features respectively;

[0020] Physical constraint module:

[0021] Apply J-integral criterion constraint to crack analysis branch;

[0022] Introduce Archard wear equation constraint to wear analysis branch.

[0023] Preferably, the superpixel segmentation method in step two comprises:

[0024] Generate adaptive grid based on SLIC algorithm, and the grid size is positively correlated with the width of the conductor;

[0025] Extract the gray level gradient histogram (HOG) features of each superpixel region, and input the SVM classifier to identify the wear area.

[0026] Preferably, the grid size of the superpixel segmentation in step two is dynamically adjusted:

[0027] The initial grid size W0= conductor width / N, where N is a preset division coefficient;

[0028] According to the local variance σ of the wear gradient Δh 2 Real-time adjustment.

[0029] Preferably, the specific method for predicting the synergistic evolution trend of α and β in step two through the physical information neural network (PINN) includes:

[0030] The α-β phase space state equation is constructed within the PINN framework:

[0031]

[0032] Where σ is the conductor stress; T is the temperature; And are the spatial gradients of α and β, respectively; θ is the trainable parameter of the neural network;

[0033] Embed the LSTM network as a time series prediction submodule in the PINN to predict the future 6-hour trajectory;

[0034] Jointly optimize the data fitting term and the physical conservation term in the loss function:

[0035]

[0036] Where y pred is the network prediction value; y true is the measured value; k is the material wear coefficient determined according to the Archard wear equation k=K / (Hσ), K is the dimensionless wear coefficient, H is the material hardness, and σ is the contact stress; λ is the weight coefficient of the physical constraint term, with a value range of 0.1-1.0.

[0037] Preferably, the training data of the 3D-CNN in step two includes:

[0038] Crack samples: FEM simulation data covering I / II / III type cracks and mixed type cracks;

[0039] Wear samples: ladder wear, groove wear, and pitting morphology generated based on a friction testing machine.

[0040] Preferably, the hierarchical early warning in step three includes:

[0041] Primary warning (single indicator exceeding limit): adjust equipment operating parameters and record abnormal positions;

[0042] Second-level warning (double-index overrun): dispatch unmanned aerial vehicle for re-inspection and generate maintenance work order;

[0043] Third-level warning (coordinated growth overspeed): trigger emergency load shedding and personnel evacuation instructions.

[0044] Preferably, the coordinated rate threshold in step three has a dynamic adjustment method:

[0045] Construct the alpha-beta phase space evolution equation:

[0046]

[0047] Where k is the material constant, E is the activation energy, R is the gas constant, and T is the environmental temperature;

[0048] When the real-time calculated value exceeds the theoretical safety boundary, a warning is triggered.

[0049] Compared with the prior art, the present application has the following beneficial effects:

[0050] Based on the coordinated collection of terahertz imaging, multispectral vision, and optical fiber sensing, the present application synchronously acquires information about internal cracks, surface wear, and stress distribution of the conductor, realizing multi-modal data collection. Combined with 3D convolution network and physical information neural network, the present application realizes accurate prediction of crack propagation and wear evolution, ensuring that the results conform to the laws of material mechanics. The dynamic threshold adjustment mechanism optimizes the warning logic according to real-time working conditions, enhancing the environmental adaptability of the system. The hierarchical warning strategy links equipment control, forming a gradual protection from speed reduction to emergency shutdown, and maximizes the continuity of power supply. This method effectively reduces the false negative rate and false positive rate, prolongs the service life of the slide wire, and reduces unplanned downtime losses.

[0051] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and attained by the structure particularly pointed out in the description, claims, and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 Flowchart of the slide wire fault maintenance prevention method based on visual detection of the present application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0054] Referring to Figure 1 A visual detection-based slide wire fault maintenance prevention method, comprising the following steps:

[0055] Step 1: Multi-modal data synchronous acquisition and calibration

[0056] Hardware cooperative control + standard sample calibration is adopted to realize high-precision synchronous acquisition of terahertz, multi-spectral and optical fiber three-modal data, and solve the problem of traditional single-sensor detection blind area.

[0057] 1. Hardware deployment:

[0058] Terahertz module: emit 0.1-1THz broadband pulse, penetrate the insulating layer to detect internal cracks (penetration depth 5mm, resolution 200μm).

[0059] Multi-spectral camera:

[0060] Visible light channel: 400-700nm, frame rate 1000fps, exposure time 0.1ms, freeze current collector shoe motion blur.

[0061] Near-infrared channel: 850-1700nm, combined with 1070nm laser line scanning, enhance the contrast of wear groove edge.

[0062] Optical fiber sensor: implant the conductor at an interval of 5cm, and use OFDR technology to demodulate strain signal (accuracy ±5μm).

[0063] 2. Data synchronization:

[0064] Align the three-modal clocks through PTP protocol, time deviation ≤1ms;

[0065] Spatial alignment: based on calibration board coordinate system conversion, positioning error ≤0.5mm.

[0066] 3. Calibration process:

[0067] Preset standard defect sample (including artificial crack and wear groove) at the end of the slide wire;

[0068] Each time the system starts, compare the sample imaging data with the reference value to calibrate the detection parameter deviation of terahertz, multi-spectral and optical fiber sensors;

[0069] Calibration coefficient calculation:

[0070]

[0071] Wherein, D 基准 is the standard defect size; D 实测 is the measured value of the sensor.

[0072] Example: Detection of standard sample at system startup:

[0073] Terahertz imaging measured crack width 0.19mm (reference value 0.20mm -> calibration factor γ = 0.20 / 0.19 ≈ 1.05);

[0074] Multispectral camera measured wear groove depth 0.48mm (reference value 0.50mm -> γ = 0.50 / 0.48 ≈ 1.04);

[0075] Real-time correction: write γ into sensor firmware, dynamically compensate measurement deviation; after calibration, the crack width measurement error in subsequent detection is reduced from ±0.03mm to ±0.01mm.

[0076] Step two: Crack-stress coupling factor α calculation

[0077] Using 3D-CNN+J integral physical constraint, quantifying the synergistic effect of crack propagation and stress concentration, predicting the risk of conductor fracture;

[0078] 1. 3D-CNN training:

[0079] Input data: Terahertz three-dimensional body data (256x256x64 voxels), covering I / II / III type cracks;

[0080] Network structure:

[0081] Encoder: 3 layers of 3D convolution (kernel size 3x3x3, step 2), extracting multi-scale features;

[0082] Decoder: 3 layers of deconvolution, outputting crack propagation rate vector V crack =(v x ,v y ,v z ).

[0083] Training data:

[0084] Crack type: I type (opening), II type (slip), III type (tearing);

[0085] Sample size: 500 groups of FEM simulation data + 200 groups of measured data.

[0086] 2. Physical constraints: loss function embedded J integral criterion:

[0087] L = ||V pred -V true || 2 + λ J ||J pred -J FEM || 2

[0088] Among them, V pred is the crack propagation rate vector predicted by the network; V true is the true value; J pred is the J integral value predicted by the network; J FEM is the finite element simulation benchmark value; J =0.5 is the physical constraint weight;

[0089] 3. Stress concentration factor K t calculate:

[0090]

[0091] Example calculations, such as:

[0092] Maximum local stress σ max =120MPa;

[0093] Nominal stress σ nom =80MPa;

[0094] Calculated as:

[0095] 4. Coupling factor α synthesis:

[0096] α=||V crack ||·K t

[0097] Example: If V crack =(0.2, 0.1, 0.05) mm / h, then:

[0098] Step 3: Dynamic extraction and classification of wear parameters

[0099] Adaptive superpixel segmentation + SVM classification is used to accurately quantify the surface wear area (β) and depth gradient (Δh) to identify contact failure risks. Based on the multispectral data calibrated in step 1, the conductor surface wear area is accurately segmented, and the wear area ratio β and depth gradient Δh are quantified.

[0100] 1. Data input and preprocessing

[0101] Data source:

[0102] Calibrated wear image captured by the near-infrared channel (850-1700 nm) of the multispectral camera (from the calibration module in step 1);

[0103] Example data: A copper conductor busbar (50 mm wide), with an image resolution of 0.1 mm / pixel after calibration.

[0104] Image enhancement:

[0105] The CLAHE algorithm is used to enhance the contrast of wear grooves with the following parameters: grid size 8×8, contrast limit 2.0;

[0106] Effect verification: The signal-to-noise ratio (SNR) of the enhanced image is improved from 15dB to 28dB.

[0107] 2. Dynamic Superpixel Segmentation

[0108] Initial mesh generation (SLIC algorithm):

[0109] Mesh size calculation: Where N is the preset division coefficient;

[0110] Dynamic Mesh Adjustment:

[0111] Local variance calculation: Calculate the wear depth gradient variance for each 5×5 grid area:

[0112]

[0113] Where Δh i is the wear depth of the i-th pixel; μ Δh is the average depth of the local area.

[0114] According to the local wear gradient variance σ 2 (Δh) adjustment:

[0115]

[0116] Example: σ is measured in a certain area 2 (Δh)=0.12mm 2 σ, the grid is adjusted to W = 2.5 mm.

[0117] 3.HOG feature extraction and SVM classification:

[0118] 3.1HOG parameter settings:

[0119] Cell size: 2×2 superpixels (corresponding to the dynamically adjusted grid size);

[0120] Number of directions: 9 directions (0°-180° evenly distributed);

[0121] Block size: 2×2 cells, block stride 1 cell;

[0122] Feature dimension: 36-dimensional HOG features (9 directions × 4 cells) are extracted for each superpixel.

[0123] 3.2.SVM classifier training:

[0124] Training data:

[0125] Positive samples: 200 wear-out superpixels (grooves, scratches, etc.);

[0126] Negative samples: 200 non-wear-out superpixels (flat surface, oil stains, etc.);

[0127] Kernel function: RBF kernel (C = 1.0, γ = 0.1).

[0128] 3.3 Wear-out area identification:

[0129] Input: 100 superpixel HOG features after dynamic grid division;

[0130] Output: Binary mask (wear-out = 1, non-wear-out = 0);

[0131] Example results:

[0132] 35 wear-out superpixels detected → Wear-out area ratio:

[0133]

[0134] 4. Wear-out depth gradient Δh calculation:

[0135] 4.1 Three-dimensional morphology reconstruction:

[0136] Based on binocular near-infrared imaging (camera parameters calibrated in step one), reconstruct the surface elevation map; depth accuracy: ±0.02mm (after step one calibration).

[0137] 4.2 Gradient calculation:

[0138] Local area: single dynamic grid coverage area (example: 2.5mm x 2.5mm);

[0139] Calculation formula:

[0140]

[0141] Where, h i is the depth of the ith sampling point; is the average height of the non-wear-out area.

[0142] Example data:

[0143] Take a wear-out groove as the center, a 5x5 grid area (25 pixel points);

[0144] Depth gradient data: near-infrared three-dimensional reconstruction value:

[0145] Sampling point height h i= [0.5, 0.6, 0.4,..., 0.5] mm (n = 25)

[0146] Average height of un-worn area

[0147] is expressed as:

[0148] Step four: PINN prediction

[0149] The PINN framework is used to embed LSTM time series prediction, combining physical laws and data-driven models to predict the crack-wear collaborative evolution trend.

[0150] 1. PINN architecture:

[0151] Input layer: a, b, (Temperature), and are the spatial gradients of a and b, respectively;

[0152] State equation:

[0153]

[0154] where a is the crack-stress coupling factor obtained in step two, b is the wear area ratio extracted in step three, and q is the neural network trainable parameter;

[0155] 2. Embed the LSTM network as a time series prediction submodule in PINN to predict the future 6-hour trajectory;

[0156] Input: historical 60-minute a, b time series (time step 1 minute), format X = [(a1, b1), (a2, b2),..., (a 60 , b 60 )] ;

[0157] Network structure:

[0158] 2-layer LSTM, hidden layer dimension 64, Dropout rate 0.2;

[0159] Output layer: future 6-hour prediction value (t = 1, 2,..., 360 minutes).

[0160] Embedding PINN method:

[0161] LSTM output as the initial condition of PINN, involved in the state equation solution.

[0162] 3. Loss function

[0163] The data fitting term and the physical conservation term are jointly optimized in the loss function:

[0164]

[0165] where y pred is the network predicted value; y true = (a, b) is the measured value; k is the material wear coefficient determined according to the Archard wear equation k = K / (Hσ); K = 2.5 x 10 -5 is the dimensionless wear coefficient; H = 1.2 GPa is the material hardness; σ = 85 MPa is the contact stress; λ = 0.3 is the weight coefficient of the physical constraint term, with a value range of 0.1-1.0.

[0166] 4. Implementation steps and examples

[0167] 4.1 Data preparation

[0168] Training data: 300 sets of time series data of a certain aluminum conductor trolley line (model AL-6061), including:

[0169] a sequence: 0.34→0.38→0.41→... (interval 1 minute);

[0170] b sequence: 17.5%→18.2%→19.0%→...;

[0171] Environmental parameters: temperature T = 40℃, contact stress σ = 85 MPa.

[0172] 4.2 Model training

[0173] 4.21 Forward propagation:

[0174] Calculate the theoretical wear rate according to the Archard equation:

[0175] Input 60 minutes of historical data and predict the next 6 hours

[0176] 4.22 Physical constraint calculation:

[0177] Archard term:

[0178] 4.23 Loss calculation:

[0179] Data term: ||0.53-0.50|| 2 +||0.23-0.22|| 2 = 0.0009 + 0.0001 = 0.0010

[0180] Physical term: ||0.02% / min-2.91 x 10 -7×3600 2 = 0.0011

[0181] Total loss:

[0182] 4.3 Forecast output

[0183] Future 6-hour trajectory:

[0184] 0.53→0.57→0.61→0.64 (error ±0.03);

[0185] 23.0%→24.8%→26.5%→28.3% (error ±1.2%).

[0186] Step five: Dynamic threshold adjustment and hierarchical control

[0187] Using phase space evolution equation + multi-level linkage strategy, the warning threshold is optimized according to the real-time working condition, and the hierarchical control instruction is triggered;

[0188] 1. Synergistic rate threshold calculation

[0189] Evolution equation and parameter definition:

[0190]

[0191] Where k is the material constant, E is the activation energy, R is the gas constant, and T is the environmental temperature;

[0192] Example calculation:

[0193] For example, k = 1.2 × 10 -4 , E = 85.6 kJ / mol, T = 313 K (40°C);

[0194]

[0195] When the real-time calculated value exceeds the theoretical safety boundary, a warning is triggered;

[0196] 2. Hierarchical warning trigger logic

[0197] First-level warning (single indicator overrun):

[0198] Trigger condition: For example, α > 0.7;

[0199] Control instruction:

[0200] Limit the acceleration of mobile devices: a ≤ 0.3 m / s 2 (original value 0.5 m / s 2 );

[0201] HMI interface highlights the crack location and records the coordinates (X=12.5m, Y=0.3m);

[0202] Example effect: After acceleration limit, the crack propagation rate is reduced by 18%.

[0203] Secondary warning (double index overrun)

[0204] Trigger condition: For example, β>15% and Δh>0.3mm:

[0205] Control instruction:

[0206] Schedule unmanned aerial vehicle re-inspection: carry high-precision infrared thermal imager (resolution 0.1℃) to scan the wear area.

[0207] If the local temperature rise ΔT>15K ΔT>15K (for example: the temperature of a certain point is detected to be 65℃65℃, and the environmental temperature is 50℃50℃), activate the standby current collector shoe to share 50% of the current load.

[0208] Example data: unmanned aerial vehicle re-inspection confirms that the wear area error is less than 3%, and the temperature rise area is accurately positioned.

[0209] Schedule unmanned aerial vehicle re-inspection and generate maintenance work order, and activate the standby current collector shoe;

[0210] Third level warning (cooperative rate overrun):

[0211] Trigger condition: For example

[0212] Control instruction:

[0213] Emergency cut off power supply circuit, activate standby slide wire channel (switching time ≤50ms);

[0214] Based on Dijkstra algorithm to plan the optimal switching path, the path length is shortened by 12%;

[0215] Acoustic and optical alarm and push maintenance work order to management system (including defect three-dimensional model and trend prediction);

[0216] Example effect: After a certain warning trigger, the system switching time is 48ms, which avoids arc discharge accident.

[0217] This scheme collects internal and external defect information of the slide wire through multi-modal data fusion, analyzes the crack propagation and wear evolution trend combined with the mixed model of physical constraints, dynamically optimizes the detection threshold and triggers graded warning, realizes the closed-loop management from accurate detection to intelligent decision, significantly improves the fault recognition rate and maintenance timeliness, and ensures the safe and stable operation of the mobile power supply system.

Claims

1. A visual inspection-based trolley line failure maintenance prevention method, characterized by, The method comprises the following steps: Step 1: Multi-modal data synchronous acquisition: deploy a terahertz imaging module, a multi-spectral array camera and a distributed optical fiber sensor, wherein: The terahertz imaging module scans the internal layer crack defects of the conductor with 0.1-1THz pulse; The multi-spectral camera synchronously captures the current collector shoe motion trajectory in the visible light band (400-700nm) and the conductor surface wear in the near-infrared band (850-1700nm); The optical fiber sensor monitors the conductor axial stress fluctuation at an interval of 5cm; Step 2: Multi-physical field coupling analysis: input the data of step 1 into a hybrid physical-data driven model, and perform: Based on 3D convolutional neural network (3D-CNN) to extract the 3D diffusion rate vector V of conductor crack crack ; Fusion fiber stress data to calculate crack-stress coupling factor α =‖V crack ‖·K t (K t is the stress concentration coefficient); Superpixel segmentation is performed on the near-infrared image to extract the depth gradient Δh and area ratio β of the wear groove; Through the collaborative evolution trend of α and β, the loss function is embedded in the Archard wear equation and the Paris crack propagation law to predict the collaborative evolution trajectory in the future T period; Step 3: Trigger hierarchical early warning when any of the following conditions is met: The crack-stress coupling factor α exceeds the material fatigue threshold; One of the wear area ratio β and the wear depth gradient Δh exceeds the threshold; The collaborative growth rate of α and β exceeds the preset collaborative rate threshold.

2. The method for preventing the fault maintenance of the slide wire based on the visual detection, according to claim 1, characterized in that, The hybrid physical-data driven model in step 2 comprises: Data driven module: a parallel double branch network is used to process crack and wear features respectively; Physical constraint module: The J integral criterion constraint is applied to the crack analysis branch; The Archard wear equation constraint is introduced to the wear analysis branch.

3. The method of claim 1, wherein the method further comprises: The superpixel segmentation method in step 2 comprises: An adaptive grid is generated based on the SLIC algorithm, and the grid size is positively correlated with the conductor width; The gray level gradient histogram (HOG) features of each superpixel region are extracted and input into an SVM classifier to identify the wear area.

4. The method of claim 1, wherein the method further comprises: The grid size of the superpixel segmentation in step 2 is dynamically adjustable: The initial grid size W0 = conductor width / N, wherein N is a preset division coefficient; According to the local variance σ of the wear gradient Δh 2 (Δh) real-time adjustment.

5. The method of claim 1, wherein the method further comprises: The specific method for predicting the collaborative evolution trend of α and β by the physical information neural network (PINN) in step 2 comprises: A α-β phase space state equation is constructed within the PINN framework: where σ is the conductor stress; T is the temperature; and are spatial gradients of a and β, respectively; θ are trainable parameters of the neural network; An LSTM network is embedded as a time series prediction submodule in the PINN to predict the future 6-hour trajectory; The data fitting term and the physical conservation term are jointly optimized in the loss function: Wherein, y pred is a network prediction value; y true is a measured value; k is a material wear coefficient determined according to the Archard wear equation k = K / (Hσ), K is a dimensionless wear coefficient, H is material hardness, and sigma is contact stress; lambda is a weight coefficient of a physical constraint term, and the value range is 0.1-1.

0.

6. The method of claim 1, wherein the method further comprises: The training data of the 3D-CNN in step 2 comprises: Crack samples: FEM simulation data covering I / II / III type cracks and mixed type cracks; Wear samples: ladder wear, groove wear and pitting morphology generated based on a friction testing machine.

7. The method of claim 1, wherein the method further comprises: The hierarchical early warning in step 3 comprises: Primary warning (single indicator exceeding limit): adjust the equipment operation parameters and record the abnormal position; Secondary warning (double indicators exceeding limit): dispatch a drone for re-inspection and generate a maintenance work order; Tertiary warning (collaborative growth exceeding speed): trigger emergency load shedding and personnel evacuation instructions.

8. The method of claim 1, wherein the method further comprises: The collaborative rate threshold in step 3 has a dynamic adjustment method: An α-β phase space evolution equation is constructed: Wherein, k is a material constant, E is the activation energy, R is the gas constant, and T is the environmental temperature; When the real-time calculated value exceeds the theoretical safety margin, a warning is triggered.

Images