Intelligent railway automation inspection system and method based on multi-modal fusion

Abstract

The application discloses an intelligent railway automation inspection system and method based on multi-modal fusion, which synchronously collects track structure vibration signals and line environment visual data by deploying high-sensitivity three-axis vibration sensors and high-definition industrial cameras on track locomotives; uses an edge computing unit to perform VMD variational modal decomposition and mixed domain feature extraction on original vibration data, and performs illumination adaptive preprocessing on image data; then uploads feature data to the cloud through 5G / Ethernet, and a multi-modal large model with a cross-modal attention fusion mechanism is deployed to perform adaptive weight distribution and complementary alignment on vibration features and visual features, realizes accurate identification of track cracks, fastener loss, foreign matter intrusion, ballast disease and other abnormalities, and outputs track structure integrity scores. The application solves the technical problems of high false alarm rate of traditional single modal detection and poor compatibility in harsh environments, and significantly improves the automation degree and detection accuracy of railway inspection.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for rail transit, specifically to an intelligent railway automated inspection system and method based on multimodal fusion. Background Technology

[0002] With the rapid growth of my country's railway transportation mileage and the continuous improvement of operating speed, the safety inspection and maintenance of railway lines has become a key link in ensuring operational safety.

[0003] Traditional railway inspections mainly rely on manual foot patrols or visual inspections from track vehicles, which suffers from high labor intensity, low inspection efficiency, strong subjectivity, and difficulties in operation at night or in inclement weather. The efficiency of manual inspections is typically only 5-10 km / h, and it is highly susceptible to the influence of personnel experience and fatigue, resulting in a high rate of missed inspections, making it difficult to meet the safety requirements of high-density, high-speed operations.

[0004] While some existing track inspection equipment can collect vibration or image data, they often suffer from the following technical deficiencies:

[0005] 1. Limited functionality and isolated modes: Most devices rely solely on single vibration detection or single visual inspection. While single vibration detection can reflect the overall health of the structure, it cannot accurately pinpoint specific fault types and is easily affected by the train's own vibrations, leading to false alarms. While single visual inspection can visually identify surface defects, the image quality drops sharply in adverse weather conditions such as nighttime, heavy fog, rain, and snow, and it cannot detect hidden damage inside the rails, resulting in a high rate of missed detections.

[0006] 2. Simple data fusion, lacking deep complementarity: A few existing technologies attempt to combine vibration and image data, but only at the level of simple data layer splicing or decision layer voting fusion. They fail to fully explore the deep complementary relationship between the two modes and cannot adaptively adjust the weights of each mode according to environmental changes, resulting in insufficient compatibility under complex working conditions.

[0007] 3. Offline analysis and poor real-time performance: Existing inspection vehicles often need to operate during "maintenance windows," and the collected data needs to be brought back to the laboratory for offline analysis. This makes it impossible to detect potential hazards and issue immediate warnings in real time, resulting in long fault response times and increasing the risk of safety accidents.

[0008] Therefore, there is an urgent need for a railway automated inspection system that can achieve deep fusion of multimodal data, real-time intelligent analysis, automatic hierarchical early warning, and can be mounted on existing operating locomotives to address the shortcomings of the aforementioned existing technologies. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent railway automated inspection system and method based on multimodal fusion, in order to solve the technical problems of high false alarm rate, poor compatibility in harsh environments, and inability to provide real-time early warning in traditional single-modal detection.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A multimodal fusion-based intelligent automated railway inspection system includes:

[0012] The vehicle-mounted sensing unit includes an industrial camera, a triaxial vibration sensor, and a positioning module. The industrial camera is used to continuously acquire image data of track fasteners, rail surfaces, track bed, and surrounding environment at a resolution of 2048×1080 and a frame rate of 30fps. The triaxial vibration sensor is used to acquire triaxial vibration acceleration signals of the track structure during train operation at a sampling rate of 1kHz. The positioning module is used to obtain train position information with decimeter-level accuracy in real time.

[0013] An edge processing unit, integrated on the main control board, is used to perform VMD variational mode decomposition and time-domain and frequency-domain hybrid feature extraction on the vibration acceleration signal, perform Retinex illumination adaptive enhancement preprocessing on the image data, and encrypt and package the processed feature data.

[0014] The communication transmission unit is used to transmit the packaged feature data in real time via 5G or Ethernet to the cloud intelligent analysis unit deployed on the cloud server based on the encrypted UDP protocol.

[0015] The cloud-based intelligent analysis unit is used to receive the feature data, perform adaptive weight allocation and complementary alignment of vibration features and visual features through a cross-modal attention fusion mechanism, identify track anomaly types using a trained multimodal recognition model, and output a track structure integrity score.

[0016] The application early warning unit is used to automatically trigger graded early warnings based on the track anomaly type identification results and integrity score, and send alarm notifications containing location, anomaly type and on-site images to maintenance personnel via SMS, email or mobile work order.

[0017] A method for automated intelligent railway inspection based on multimodal fusion includes the following steps:

[0018] S1: Data synchronous acquisition. During the operation of the rail locomotive, track image data, vibration acceleration data and position data are synchronously acquired through the on-board sensing unit, and timestamps are aligned.

[0019] S2: Edge preprocessing, performing VMD decomposition and hybrid domain feature extraction on vibration data at the edge, and performing illumination adaptive enhancement and preliminary target detection on image data;

[0020] S3: Encrypted transmission, which encrypts and packages the processed feature data and location data, and uploads them to the cloud in real time via the 5G network;

[0021] S4: Multimodal fusion analysis, which fuses vibration features and visual features in the cloud through a cross-modal attention mechanism, inputs them into the multimodal recognition model, identifies the anomaly type and calculates the integrity score;

[0022] S5: Tiered early warning and work order dispatch. Based on the detection results, the abnormality level is determined, an early warning is automatically triggered, and alarm information and maintenance work orders are pushed to relevant maintenance personnel.

[0023] The automated inspection system of this invention includes an onboard sensing layer, an edge processing layer, a cloud-based intelligent analysis layer, and an application early warning layer. By deploying high-sensitivity triaxial vibration sensors and high-definition industrial cameras on a rail locomotive, it simultaneously collects track structure vibration signals and visual data of the track environment. Using an edge computing unit, it performs VMD variational mode decomposition and hybrid domain feature extraction on the raw vibration data, and performs adaptive illumination preprocessing on the image data. Subsequently, the feature data is uploaded to the cloud via 5G / Ethernet, and a multimodal large model with a cross-modal attention fusion mechanism is deployed to adaptively weight and complementarily align vibration and visual features. This enables accurate identification of anomalies such as track cracks, missing fasteners, foreign object intrusion, and track bed defects, and outputs a track structure integrity score.

[0024] This invention solves the technical problems of high false alarm rate and poor compatibility in harsh environments of traditional single-mode detection. Experiments show that the system can achieve a detection accuracy of 99.2% for track defects and reduce the false alarm rate to 0.8%. It can complete real-time inspections at normal train operating speeds without interrupting line operations, significantly improving the automation level and detection accuracy of railway inspections and providing strong technical support for the safe operation of railway lines. Attached Figure Description

[0025] Figure 1 These are physical images of some of the devices used in this invention.

[0026] Figure 2 This is a schematic diagram of the system structure of the present invention.

[0027] Figure 3 This is a cloud platform railway data feedback diagram of the present invention.

[0028] Figure 4 This is a diagram of the cloud platform device status interface of the present invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to specific embodiments.

[0030] Combination Figure 1 , Figure 2 As shown, an intelligent railway automated inspection system based on multimodal fusion includes:

[0031] 1. Vehicle-mounted sensing unit

[0032] This unit serves as the system's data source entry point. It is securely mounted on the bottom or side of the railcar using a specially designed shock-absorbing base, allowing for rapid deployment without requiring any modifications to the locomotive. Specifically, it includes:

[0033] Industrial camera: It adopts a 2-megapixel high-sensitivity CMOS sensor, supports high-speed acquisition of 2048×1080@30fps, and has a built-in infrared fill light module, which can achieve clear imaging in low light conditions of 0.1Lux at night. It is specially designed for continuous acquisition of image data of track fasteners, rail surface, track bed and surrounding environment.

[0034] The triaxial vibration sensor uses a high-precision MEMS accelerometer with a range of ±16g and a sampling rate of up to 200Hz. It can accurately capture the high-frequency vibration response caused by track defects when a train passes by, providing quantitative data on track smoothness and structural health.

[0035] Positioning module: Integrates BeiDou-3 / GPS multi-frequency signals. It achieves decimeter-level positioning accuracy of ±10cm in horizontal direction and ±20cm in vertical direction. It can accurately label each set of detection data with mileage to ensure the accuracy of fault location.

[0036] 2. Edge processing unit

[0037] Integrated into a high-performance main control board, it serves as the edge computing core of the system, responsible for the initial processing of raw data to reduce the computing pressure and transmission bandwidth in the cloud.

[0038] Vibration signal processing: Addressing the low signal-to-noise ratio and non-stationary nature of the original vibration signal, this invention employs the VMD (Variational Mode Decomposition) algorithm to decompose the signal. This adaptively decomposes the complex vibration signal into several stationary intrinsic mode functions, effectively separating the train's own vibration noise from the impact signal caused by track defects. Based on this, time-domain features (peak-to-peak value, root mean square, kurtosis, waveform factor) and frequency-domain features (spectral entropy, centroid frequency) are extracted, achieving dimensionality reduction and characterization of the vibration signal.

[0039] Image preprocessing: To address the issue of large variations in outdoor lighting, this invention employs the Retinex algorithm to perform adaptive lighting enhancement on the acquired images, effectively eliminating the impact of environmental factors such as backlighting, shadows, and haze on image quality and improving the compatibility of subsequent recognition algorithms.

[0040] 3. Communication transmission unit

[0041] Employing 5G industrial-grade communication modules and an Ethernet backup mechanism, and supporting the UDP protocol stack, it achieves encrypted data packaging and real-time transmission. Addressing the issue of network fluctuations along railway lines, a breakpoint resumption and local caching mechanism was designed to ensure no data loss during network interruptions and automatic retransmission upon network recovery, guaranteeing data integrity.

[0042] 4. Cloud-based intelligent analysis unit

[0043] This is the core innovation of the invention: it is deployed on a cloud server to receive feature data uploaded from the edge and achieves deep fusion of vibration and visual data through a cross-modal attention fusion mechanism.

[0044] Traditional fusion methods often assign equal weights to features of both modalities, which works in ideal environments but becomes less reliable in harsh environments (such as when rain causes image blurring). If these modalities are still given high weights, it will lead to recognition errors.

[0045] The cross-modal attention fusion mechanism proposed in this invention can automatically calculate the confidence scores of two modalities in the current sample:

[0046] When the image is clear and the lighting is good, the confidence of the visual features is high, and the model automatically increases the weight of the visual features.

[0047] When an image is blurred due to weather interference, the confidence of visual features is low, and the model automatically reduces their weights and instead relies more on the highly reliable vibration features.

[0048] Conversely, when the vibration signal is disturbed by complex structures such as turnouts, the model will automatically increase the weight of visual features.

[0049] This adaptive weight allocation achieves complementary advantages between the two modal features, greatly improving the system's compatibility under complex operating conditions.

[0050] The fused features are input into the trained multimodal recognition model, which can automatically identify a variety of common defects such as track cracks, missing / loose fasteners, foreign object intrusion, water accumulation / caking of the track bed, and abnormal welds, and output a track structure integrity score of 0-10, with the lower the score, the worse the condition.

[0051] 5. Application Early Warning Unit

[0052] This unit is designed for end users and provides integrated operation and maintenance management functions.

[0053] Data visualization dashboard: Real-time display of the entire line's inspection progress, vibration trends, and heat map of defect distribution, allowing managers to intuitively grasp the overall health status of the line.

[0054] Tiered early warning engine: Based on the severity and completeness of the anomaly, it automatically classifies the anomaly into three levels of early warning and triggers different response mechanisms.

[0055] Automatic work order dispatch: Once an anomaly is detected, the system automatically generates a maintenance work order, which includes precise location coordinates, anomaly type, on-site photos, and other information. This order is then directly pushed to the maintenance personnel's mobile devices, achieving full automation from the discovery of potential hazards to the dispatch of work orders, reducing fault response time from hours to minutes.

[0056] Trend prediction: By comparing historical data, the system can analyze the deterioration trend of track condition, predict potential faults in advance, and realize the transformation from "fault repair" to "condition repair" and "predictive repair".

[0057] A method for automated intelligent railway inspection based on multimodal fusion includes the following steps:

[0058] S1: Data synchronous acquisition. During the operation of the rail locomotive, track image data, vibration acceleration data and position data are synchronously acquired through the on-board sensing unit, and timestamp alignment is performed.

[0059] (1) Obtain data via UARTx (x=1,2,3...), and add a timestamp upon completion.

[0060] typedef struct {

[0061] uint64_t timestamp; / / Millisecond-level timestamp

[0062] uint8_t data

[1024] ; / / General data buffer

[0063] uint32_t data_len; / / Actual data length

[0064] TimestampedData;

[0065] (2) Periodically process the data buffer pool (T=1s)

[0066] S2: Edge preprocessing, performing VMD decomposition and hybrid domain feature extraction on vibration data at the edge, and performing illumination adaptive enhancement and preliminary target detection on image data.

[0067]

[0068] Optimize the meaning of target symbols

[0069] Meaning of constraint symbols

[0070] st "Subject to" means "restricted by" constraints. f(t) The original real signal to be decomposed ∑k=1Kuk(t) The sum of all modal components must equal the original signal to ensure no information loss during decomposition.

[0071] S3: Encrypted transmission, which encrypts and packages the processed feature data and location data, and uploads them to the cloud in real time via the 5G network.

[0072] The transmission key is exchanged via SSL algorithm, the data is transmitted in JSON format, and AT commands are sent from the UART port to the 5G module.

[0073] S4: Multimodal fusion analysis, which fuses vibration features and visual features in the cloud through a cross-modal attention mechanism, inputs them into a multimodal recognition model, identifies anomaly types, and calculates integrity scores.

[0074] The training process of the multimodal recognition model includes:

[0075] Construct a multimodal training dataset containing vibration signal samples, image samples, and corresponding annotations; train vibration feature extraction subnetwork and visual feature extraction subnetwork respectively;

[0076] A cross-modal attention module is introduced to minimize the distribution differences between modalities and learn complementary representations of features from the two modalities.

[0077] The entire network is fine-tuned end-to-end using anomaly classification loss and regression loss as a joint loss function.

[0078] Single-modal feature extraction: Fv=Ev(xv), Fi=Ei(xi);

[0079] Cross-modal alignment fusion: Ffus=Att(Fv,Fi);

[0080] Total combined loss: L = λ1L cls +λ2L reg +λ3L dis ;

[0081] Where: xv represents vibration samples, xi represents image samples, Ffus represents a dual-branch feature extraction network, Att represents cross-modal attention fusion, and L... cls For anomaly classification loss, Lreg For regression loss, L dis λ1, λ2, and λ3 are the modal distribution difference loss and the balancing weight coefficients.

[0082] S5: Tiered early warning and work order dispatch. Based on the detection results, the abnormality level is determined, an early warning is automatically triggered, and alarm information and maintenance work orders are pushed to relevant maintenance personnel.

[0083] The tiered early warning system specifically includes:

[0084] When the integrity score is below 6 or a serious anomaly is detected, such as a broken rail or a large-scale foreign object intrusion, a Level 1 warning is triggered, which is immediately sent to the dispatch center and the nearest maintenance personnel, and the train dispatch system is linked to issue a deceleration warning.

[0085] When the integrity score is between 6 and 8 or a general anomaly is detected, such as loose fasteners or minor cracks, a level 2 warning is triggered and sent to the maintenance personnel in the work area to arrange the planned repairs for the day.

[0086] When the integrity score is between 8 and 9, a Level 3 warning is triggered, the deterioration trend is recorded, and preventative maintenance is scheduled for the following week.

[0087] Example 1: System Deployment and Hardware Parameters

[0088] This system is installed at the bottom of the front of a certain operating passenger train using a high-strength aluminum alloy shock-absorbing base. The specific hardware configuration is as follows:

[0089] 1. Industrial camera: Model OV5640, resolution 2048×1080, frame rate 30fps, equipped with an 8mm fixed focus lens, facing the rail area, with a field of view covering the left and right rails and the middle track bed.

[0090] 2. Vibration sensor: Model B600, a three-axis digital accelerometer with a noise density as low as 20μg / √Hz and a sampling rate set to 200Hz. It is mounted on a rigid structure next to the camera to ensure accurate transmission of track vibration signals.

[0091] 3. Positioning Module: The AT6668 module is used, which supports BeiDou. In areas covered by reference stations, the positioning accuracy can reach the centimeter level.

[0092] 4. Main control board: It adopts STM32H743, which can complete the preprocessing algorithm locally in real time.

[0093] Example 2: Data Acquisition and Edge Processing

[0094] During normal train operation (speed 120km / h), the system initiates data collection:

[0095] 1. Timestamp Alignment: All data from the camera, sensors, and positioning module are synchronized using the unified PTP Precision Time Protocol with an error of less than 1ms, ensuring that images and vibration data from the same location can be accurately matched.

[0096] 2. Vibration Data Processing: Vibration data at 1kHz is read in real time at the edge, with each processing window lasting 100ms. VMD decomposition is performed on the 100 sampling points within the window, resulting in four intrinsic mode functions (IMFs). Subsequently, 12-dimensional features, including peak-to-peak value, kurtosis, and spectral entropy, are extracted from each mode, compressing the original 100-dimensional data into a 12-dimensional feature vector, significantly reducing the amount of data transmitted.

[0097] 3. Image data processing: For each frame of the image, Retinex illumination enhancement is first performed to eliminate overly dark areas caused by backlighting.

[0098] Example 3: Cloud-based Multimodal Fusion Analysis

[0099] The feature data processed at the edge is uploaded to the cloud server via the 5G network, where the multimodal model performs the following processing:

[0100] 1. Feature Input: The input includes a 12-dimensional vibration feature vector and a 512-dimensional visual feature vector extracted from the image by the ResNet network.

[0101] 2. Cross-modal fusion:

[0102] Figure 3 The image shows the feedback diagram of railway data from the cloud platform, displaying the data feedback from vibration and image data acquisition.

[0103] First, project both features into the same dimensional space. Then calculate the attention weights:

[0104] Where vi is the visual feature and uj is the vibration feature.

[0105] It can be obtained by the following calculations:

[0106] OutPut = vi * visual feature weight + uj * vibration feature weight

[0107] Using this weighting, on sunny days, the weight of visual features is 0.7 and the weight of vibration is 0.3; on rainy days, the weight of visual features automatically decreases to 0.2, while the weight of vibration features increases to 0.8.

[0108] 3. Anomaly Identification: The fused features are input into the classifier, which outputs the anomaly category. For example, when a vibration signal with a significant impact peak is detected, and a tiny crack is identified on the surface of the rail in the image, the model classifies it as a "rail crack" with a confidence level of 99.8% and gives an integrity score of 5.

[0109] Example 4: Early Warning and Work Order Dispatch

[0110] Based on the above test results, if the integrity score is below 6 out of 5, the system automatically issues a Level 1 warning.

[0111] 1. Immediately send an alarm via SMS and email to the dispatcher and work area manager of the line: "A rail crack has been detected. Please check immediately!", along with on-site photos and precise GPS coordinates.

[0112] 2. Automatically generate emergency work orders in the operation and maintenance management system and assign them to the nearest inspection personnel.

[0113] 3. At the same time, the system will synchronize this information to the train dispatching system to remind subsequent trains passing through this section to pay attention to the speed limit.

[0114] Figure 4 The cloud platform device status interface diagram of the present invention is shown, which is used to implement the intelligent railway automated inspection method based on multimodal fusion of the present invention.

[0115] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural modifications made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A smart railway automated inspection system based on multimodal fusion, characterized in that, include: The vehicle-mounted sensing unit includes an industrial camera, a triaxial vibration sensor, and a positioning module. The industrial camera is used to continuously acquire image data of track fasteners, rail surfaces, track bed, and surrounding environment at a resolution of 2048×1080 and a frame rate of 30fps. The triaxial vibration sensor is used to acquire triaxial vibration acceleration signals of the track structure during train operation at a sampling rate of 1kHz. The positioning module is used to obtain train position information with decimeter-level accuracy in real time. An edge processing unit, integrated on the main control board, is used to perform VMD variational mode decomposition and time-domain and frequency-domain hybrid feature extraction on the vibration acceleration signal, perform Retinex illumination adaptive enhancement preprocessing on the image data, and encrypt and package the processed feature data. The communication transmission unit is used to transmit the packaged feature data in real time via 5G or Ethernet to the cloud intelligent analysis unit deployed on the cloud server based on the encrypted UDP protocol. The cloud-based intelligent analysis unit is used to receive the feature data, perform adaptive weight allocation and complementary alignment of vibration features and visual features through a cross-modal attention fusion mechanism, identify track anomaly types using a trained multimodal recognition model, and output a track structure integrity score. The application early warning unit is used to automatically trigger graded early warnings based on the track anomaly type identification results and integrity score, and send alarm notifications containing location, anomaly type and on-site images to maintenance personnel via SMS, email or mobile work order.

2. The intelligent railway automated inspection system according to claim 1, characterized in that, The feature extraction of the vibration signal by the edge processing unit specifically includes: The original vibration acceleration signal is subjected to VMD variational mode decomposition, which decomposes the non-stationary vibration signal into several intrinsic mode functions; Extract the time-domain features of each modal function, including peak-to-peak value, root mean square, kurtosis, and waveform factor; Extract the frequency domain features of each modal function, including spectral entropy, centroid frequency, and mean square frequency; After normalizing the above features, they are input into the subsequent fusion network.

3. The intelligent railway automated inspection system according to claim 1, characterized in that, The cross-modal attention fusion mechanism in the cloud-based intelligent analysis unit specifically includes: Linear projections are performed on the vibration feature vector and the visual feature vector respectively, mapping them to the same feature space; Calculate the cross-modal attention weight matrix to measure the correlation between vibrational features and visual features; The features of the two modalities are weighted and fused according to the attention weight. The modal features with high confidence are assigned higher weights, while the modal features affected by noise are automatically weighted lower. The fused high-dimensional features are input into a fully connected layer, which outputs anomaly category, confidence level, and a track integrity score of 0-10.

4. The intelligent railway automated inspection system according to claim 1, characterized in that, The application early warning unit also includes a historical data comparison engine, which is used to compare the current detection results with the historical inspection data of the same section to identify the deterioration trend of the track condition and realize preventive maintenance early warning.

5. An automated inspection method based on the multimodal fusion intelligent railway automated inspection system according to any one of claims 1-4, characterized in that, Includes the following steps: S1: Data synchronous acquisition. During the operation of the rail locomotive, track image data, vibration acceleration data and position data are synchronously acquired through the on-board sensing unit, and timestamps are aligned. S2: Edge preprocessing, performing VMD decomposition and hybrid domain feature extraction on vibration data at the edge, and performing illumination adaptive enhancement and preliminary target detection on image data; S3: Encrypted transmission, which encrypts and packages the processed feature data and location data, and uploads them to the cloud in real time via the 5G network; S4: Multimodal fusion analysis, which fuses vibration features and visual features in the cloud through a cross-modal attention mechanism, inputs them into the multimodal recognition model, identifies the anomaly type and calculates the integrity score; S5: Tiered early warning and work order dispatch. Based on the detection results, the abnormality level is determined, an early warning is automatically triggered, and alarm information and maintenance work orders are pushed to relevant maintenance personnel.

6. The automated inspection method according to claim 5, characterized in that, The training process of the multimodal recognition model in step S4 includes: Construct a multimodal training dataset containing vibration signal samples, image samples, and corresponding annotations; The vibration feature extraction subnetwork and the visual feature extraction subnetwork are trained separately. A cross-modal attention module is introduced to minimize the distribution differences between modalities and learn complementary representations of features from the two modalities. The entire network is fine-tuned end-to-end using anomaly classification loss and regression loss as a joint loss function.

7. The automated inspection method according to claim 5, characterized in that, The graded early warning in step S5 specifically includes: When the integrity score is below 6 or a serious anomaly is detected, such as a broken rail or a large-scale foreign object intrusion, a Level 1 warning is triggered, which is immediately sent to the dispatch center and the nearest maintenance personnel, and the train dispatch system is linked to issue a deceleration warning. When the integrity score is between 6 and 8 or a general anomaly is detected, such as loose fasteners or minor cracks, a level 2 warning is triggered and sent to the maintenance personnel in the work area to arrange the planned repairs for the day. When the integrity score is between 8 and 9, a Level 3 warning is triggered, the deterioration trend is recorded, and preventative maintenance is scheduled for the following week.

Images