High confidence decision system for track intelligent flaw detection

Abstract

This invention discloses a high-confidence decision-making system for intelligent rail flaw detection, comprising: a data acquisition module for acquiring multi-source sensor data; a damage database for storing historical rail damage information; an input protection module for evaluating the confidence of multi-source sensor data based on historical damage data, providing input confidence reinforcement; an intelligent decision-making module for fusing reliable data and outputting intelligent decisions based on an intelligent decision-making model; a decision database for storing historical decision sets; an output protection module for evaluating the confidence of intelligent decisions based on historical decisions, providing output confidence reinforcement; and an execution module for maintaining and managing the rails according to the system decisions. This invention effectively prevents attacks, information loss, or decision-making errors during the data input and model decision-making output processes of the rail flaw detection system, reducing the probability of missed or false detections of rail damage, thereby improving railway operation safety and accelerating the digital and intelligent transformation of the railway industry.

Description

Technical Field

[0001] This invention relates to the field of intelligent rail flaw detection technology, and more specifically, to a high-confidence decision system for intelligent rail flaw detection. Background Technology

[0002] As a crucial piece of railway infrastructure, the safety of rails is paramount for improving transport capacity and ensuring passenger safety. Rail flaw detection technology aims to use multi-source sensors to periodically and rapidly detect internal and external damage to rails, enabling proactive maintenance and management before accidents occur. Currently, flaw detection technology is widely used in conventional and high-speed railways and urban rail transit.

[0003] Sensors must collect information on internal and external damage to rails without compromising their performance. Acoustic sensing, visual sensing, and eddy current testing are commonly used in non-destructive testing of rails. Among these, acoustic sensing, such as ultrasonic sensing, possesses excellent directivity and penetrability; ultrasonic waves are reflected and refracted at the damaged interface within the inspected object, thus responding to internal rail damage. Visual sensing, based on optical imaging technology, directly and effectively acquires image information of the external environment and is primarily used to detect surface damage on rails.

[0004] However, the distortion of sensor mapping of damage sources is affected by equipment parameters, environmental conditions, sensitivity, and calibration. In rail flaw detection, the variable operating environment of sensors necessitates adaptive adjustment of equipment configuration parameters, sensor malfunctions cause information loss, and the complexity of rail damage mechanisms makes damage identification difficult. These issues lead to missed and false alarms, severely impacting flaw detection efficiency, railway traffic safety, and consuming significant resources. Therefore, considering the information input of the flaw detection system, comparing and evaluating the reliability of information collected by multi-source sensors to improve the reliability of sensor characterization of damage sources is of great significance.

[0005] Currently, most rail transit operation and maintenance tasks, including data collection, data analysis, and identification and decision-making, are still performed manually, inevitably introducing subjective factors into the decision-making results. Therefore, intelligent identification and decision-making based on Artificial Intelligence (AI) technology is gradually being applied. AI technology has evolved from initial artificial neural networks to machine learning, and then rapidly to deep learning, represented by convolutional neural networks, aiming to solve problems such as target classification, detection, regression, and generation, thus bringing about changes in industries such as industry, transportation, and medicine.

[0006] However, while AI models have facilitated rapid industry development, many application cases lack assessments of the models' robustness, interpretability, and reliability. Furthermore, the "Trusted Artificial Intelligence White Paper" emphasizes the robustness and safety of AI models, as well as the need for human agency and oversight. Therefore, in the safety-oriented railway industry, AI-based intelligent rail flaw detection systems require the design of relevant protective mechanisms to constrain and regulate the system's reliability and safety, thereby further improving the accuracy of rail maintenance and reducing the risk of railway safety accidents. Summary of the Invention

[0007] To address the issues of missed or false detections of rail damage caused by missing sensor target source information and model decision-making errors, this invention aims to design a protection mechanism on the data input side and decision output side to manage and control the intelligent rail flaw detection system, thereby achieving reliable hardening of the intelligent rail flaw detection system.

[0008] To achieve the above technical objectives, this invention provides a high-confidence decision system for intelligent track flaw detection, characterized by comprising a data acquisition module (1), a damage database (2), an input protection module (3), an intelligent decision module (4), a decision database (5), an output protection module (6), and an execution module (7), wherein: the data acquisition module (1) is used to acquire multi-source sensor data of the rail in a specific spatiotemporal domain in real time; the damage database (2) is used to store historical damage information of rails in various sections of the railway; and the input protection module (3) is used to match and compare the multi-source sensor data with the relevant data in the damage database. The intelligent decision-making module (4) is used to process and integrate reliable data, and make intelligent decisions based on the intelligent decision-making model set and the constraints and operation and maintenance strategies of the railway industry. The decision database (5) is used to store the historical decision set obtained by the intelligent decision-making model set based on the historical damage dataset. The output protection module (6) is used to match and compare the decision inferred by the intelligent decision-making model set with the historical decision set that may be related. The execution module (7) performs maintenance management of the rails according to the railway rail operation and maintenance requirements and the reliable decision output by the reliable system and the corresponding reliability level.

[0009] Advantages and beneficial effects of the present invention:

[0010] The design provided by this invention adds dual-sided reliable assessment and reinforcement to a general-purpose intelligent rail flaw detection system, including a multi-source sensor data input side and an intelligent decision output side, so that the system can ultimately output a reliable decision, improve the safety and reliability of the flaw detection system, reduce the probability of missed or false detection of rail damage, and thus improve the level of railway operation safety.

[0011] The design provided by this invention is adapted to flaw detection operations in industries with high repetition and regularity, thus having high generalization ability. Its application fields include, but are not limited to, the railway industry, and therefore have immeasurable social benefits. Attached Figure Description

[0012] Figure 1 The present invention illustrates an implementation scheme based on cloud-edge-device collaborative computing.

[0013] Figure 2 A schematic diagram of the intelligent track flaw detection system described in this invention is shown;

[0014] Figure 3 This invention illustrates the reliability assessment process of the input protection module for multi-source sensor data.

[0015] Figure 4 The intelligent decision-making implementation process described in this invention is illustrated.

[0016] Figure 5 The present invention illustrates the reliable evaluation process of intelligent decision-making by the output protection module.

[0017] Figure 6 The reliable transmission process described in this invention is illustrated;

[0018] In the attached diagram, the meanings of each number are as follows:

[0019] 1-Data acquisition module; 2-Damage database; 3-Input protection module; 4-Intelligent decision-making module; 5-Decision database; 6-Output protection module; 7-Execution module. Specific implementation methods

[0020] To make the implementation schemes and advantages of the embodiments of this application clearer, the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the application, but merely represents selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort fall within the scope of this application for a reliable system for rail flaw detection.

[0021] Figure 2The specific structure and connection relationship of the intelligent track flaw detection system provided by the present invention are shown, including: data acquisition module (1), damage database (2), input protection module (3), intelligent decision module (4), decision database (5), output protection module (6) and execution module (7); Example

[0022] Figure 1 The implementation scheme of the intelligent track flaw detection system provided by the present invention based on cloud-edge-device collaborative computing is shown;

[0023] The data acquisition module (1) is equipped with multi-source sensors, antennas, GNSS (Global Navigation Satellite System) signal receivers, communication equipment, walking equipment, power supply, processor and memory, etc. It has clock synchronization function, spatial positioning function, data acquisition function, wireless communication function, data processing and data storage function, etc. It walks along the rail at 36km / h and collects multi-source sensor data X1 in a specific time and space domain, which belongs to the terminal side of this implementation scheme;

[0024] Furthermore, the multi-source sensor includes an ultrasonic sensor array, two eddy current flaw detection sensors, and two vision sensors (CCD high-definition cameras, etc.), which are used to collect internal rail damage data {x1}, near-surface damage data {x2,x3}, and external damage data {x4,x5}, respectively, i.e., X1={x1,x2,x3,x4,x5};

[0025] It should be noted that the ultrasonic sensor array consists of a 0° probe, a forward 70° probe, a reverse 70° probe, a forward 37° probe, and a reverse 37° probe. Echoes detected by different probes at the same location are represented by pixels of different colors. As the data acquisition module moves along the rail, the ultrasonic sensor array data {x1} is presented as continuous two-dimensional image data with a sampling frequency of 50Hz; {x2,x3} is one-dimensional data with a sampling frequency of 10kHz; and {x4,x5} is two-dimensional image data with a sampling frequency of 50Hz.

[0026] Furthermore, the data acquisition module provides the multi-source sensor with spatiotemporal information of the flaw detection location, i.e., the data acquisition location, based on the GNSS signal. This information includes: time coordinates T, longitude coordinates M, latitude coordinates L, and altitude coordinates H, i.e., (T, M, L, H). This adds a spatiotemporal stamp to the data. For any coordinate (T1, M1, L1, H1) in the spatiotemporal domain, the multi-source sensor data is represented as X1(T1, M1, L1, H1) = {x1(T1, M1, L1, H1), x2(T1, M1, L1, H1), ..., x5(T1, M1, L1, H1)}.

[0027] The damage database is used to store historical damage information of rails in various sections of the railway. After being judged and confirmed by experts, it serves as a reliable root of data to transmit credibility to the downstream of the system.

[0028] Furthermore, the root of trust in the data provides a basis of trust for the multi-source sensor data in the intelligent track flaw detection system;

[0029] Furthermore, historical damage information corresponds one-to-one with historical damage in the spatiotemporal domain, including the temporal coordinates T, longitude coordinates M, latitude coordinates L, and altitude coordinates of historical damage, and the corresponding historical damage dataset X. ε = in With x i ∈X1,i=1,…,5, represents data from the same type of sensor.

[0030] Furthermore, the input protection module is connected to the data acquisition module, the damage database and the intelligent decision-making module. Based on the relevant historical damage data in the damage database, the reliability assessment of the multi-source sensor data X1 is performed, and the reliability level of the multi-source sensor data after assessment is used to determine whether to transmit it to the intelligent decision-making module.

[0031] Furthermore, before the multi-source sensor data X1 is transmitted to the intelligent decision-making module, it must first undergo a reliability assessment by the input protection module;

[0032] Furthermore, the damage database and input protection module (3) are deployed in the edge computing center of the railway station, and the edge computing resources complete the reliable assessment of multi-source sensor data;

[0033] It should be noted that the data acquisition module transmits multi-source sensor data to the input protection module based on any of the following communication technologies: WiFi, 4G, 5G, GSM (Global System for Mobile communication), and LTE (Long Term Evolution).

[0034] Furthermore, Figure 3 The following is an example of the reliability assessment process of the input protection module (3) for multi-source sensor data X1:

[0035] Step 10-1: The data acquisition module uses its own computing resources to preprocess the acquired multi-source sensor data X1, and transmits the preprocessed multi-source sensor data X1 to the input protection module based on any of the following communication technologies: WiFi, 4G, 5G, GSM (Global System for Mobile communication), and LTE (Long Term Evolution).

[0036] Specifically, the preprocessing process includes spatiotemporal alignment, outlier removal, median filtering, mean normalization, and truncation. Spatiotemporal alignment refers to aligning multi-source sensor data based on the spatiotemporal domain information provided by GNSS signals, ensuring that {x1}, {X2,x3}, and {x4,x5} describe the multi-source information of the rail in the same spatiotemporal domain. Truncation refers to truncating the multi-source sensor data according to the sampling frequency of the multi-source sensor data in the order of time domain information, truncating the multi-source sensor data once every 0.1s. Thus, the multi-source sensor data within a truncated window represents the damage information of the rail with a length of approximately 1m. That is, the damage information represented by the multi-source sensor data X1 transmitted to the input protection module has a temporal length of 0.1s and a spatial length of 1m.

[0037] Step 10-2, Historical Injury Dataset X ε Both the multi-source sensing data X1 and the shallow features obtained from the convolutional kernel set K = {k1,k2,…,k5} are used to obtain X. ε The corresponding shallow feature matrix F ε The shallow feature matrix F1 corresponding to X1, i.e. Where f i ε For k i convolution The obtained shallow features, Where f i For k i Convolution x i The obtained shallow features, K represents convolution; it extracts damage-related features from multi-source sensor data.

[0038] It should be noted that {k2,k3} is a (1×w1) one-dimensional convolution kernel used to convolve {x2,x3}, and {k1} and {k4,k5} are (w2×w2) two-dimensional convolution kernels used to convolve {x1} and {x4,x5} respectively. w1 is determined based on the sampling frequency of the eddy current flaw detector sensor, for example, w1 = 200. w2 is determined based on the size of the two-dimensional data, for example, when the two-dimensional data is 512×512, w2 = 3. The convolution kernel k... i The number of shallow features extracted from multi-source sensor data is 8, i.e., k i ={k i1 ,k i2 ,…,k i8},i=1,…,5, then

[0039] Step 10-3, in the feature matrix F ε Three initial cluster centers are set in the feature space, based on the Euclidean distance in n-dimensional space. Calculate the distance from each sample in the damaged dataset to its cluster center and minimize it. This process is repeated iteratively between cluster center c and cluster C until stability is achieved;

[0040] Step 10-4: Input the protection module to determine whether the multi-source sensor data X1 has a historical damage dataset X. ε The characteristics of the data, namely whether satisfy

[0041] Step 10-5: If it does not exist, then the multi-source sensor data does not belong to the historical damage dataset, that is... Furthermore, the absence of damage data characteristics indicates that the corresponding rail is undamaged in the spatiotemporal domain (T1, M1, L1, H1), and the input protection module clears the multi-source sensor data X1.

[0042] Step 10-6: If it exists, then the multi-source sensor data X1 may belong to the historical damage dataset X. ε The input protection module is in the damage dataset X ε In the middle, we attempted to search for relevant historical damage data. Matching with multi-source sensor data X1;

[0043] Step 10-7: Input protection module determines whether it is multi-source sensor data X1 and potentially related historical damage data. Whether a match is successful in the spatial domain (M1, L1), i.e., whether... satisfy

[0044] Step 10-8: If it does not exist, the match fails. The multi-source sensor data X1 may belong to the historical damage dataset X. ε This indicates that it may represent newly emerging damage. After damage data feature identification, a Level 1 trust stamp ε is added. 1 Level 1 trusted data was obtained.

[0045] Steps 10-9: Update the damage database based on the spatiotemporal domain (T1, M1, L1, H1) and store the Level 1 trusted data. The data is then transmitted to the intelligent decision-making module for decision recognition.

[0046] Step 10-10: If a match exists, the match is successful, indicating that the multi-source sensor data X1 may belong to the historical damage dataset X. ε Match historical injury data Transmitted to the input protection module;

[0047] Steps 10-11: Input protection module to compare multi-source sensor data X1 with historical damage data. Whether the feature spaces are similar, i.e., whether... Satisfying 2 or more correlation coefficients Where i = 1, 2, ..., 5, cov(·) is the covariance, E(·) is the expected value, and μ i For feature f i The mean, σ i For feature f i standard deviation and Similarly, we can obtain;

[0048] Steps 10-12: If not found, it indicates that the multi-source sensor has a sampling error or malfunction, which needs to be checked and corrected. After correction, the multi-source sensor data should be collected again and matched.

[0049] Steps 10-13: If they exist, it indicates that the multi-source sensor data belongs to the historical damage dataset, X1∈X ε After damage data feature identification and spatiotemporal information matching, a level 2 trusted stamp ε is added. 2 Level 2 trusted data was obtained. The data is transmitted to the intelligent decision-making module for decision recognition.

[0050] It should be noted that the trusted data transmitted to the intelligent decision-making module α = 1 or 2 is the pre-processed data, i.e., the multi-source sensing data obtained in step 10-1.

[0051] Furthermore, the intelligent decision-making module (4) is connected to the input protection module and the output protection module, and makes intelligent decisions based on the reliable data from the input protection module and in accordance with the railway industry's flaw detection and maintenance specifications and strategies.

[0052] Furthermore, the intelligent decision-making module (4) has data fusion and decision-making functions, consisting of a data fusion unit and an intelligent decision-making model set U = {u1, u2, u3}. The data fusion unit fuses data from homogeneous sensors to reduce redundancy of multi-source sensor data and improve data accuracy. The intelligent decision-making model set U is based on reliable data. Output intelligent decision

[0053] Furthermore, the intelligent decision-making models u1, u2, and u3 are all artificial intelligence models. u1 takes internal damage data {x1} as input, u2 takes near-surface damage data {x2, x3} as input, and u3 takes external damage data {x4, x5} as input. They all output the type and size of the damage, Y1 = {y1, y2, y3}.

[0054] It should be noted that the artificial intelligence model uses neural networks as the basic unit. This invention adopts a general structure, which specifically includes a feature extraction unit, a feature fusion unit, a fully connected layer unit, and a softmax classifier or detection head. The softmax classifier is a classification model, and the detection head is a target detection model.

[0055] The feature extraction unit extracts shallow global texture information and deep local semantic information of the input data layer by layer based on CNN (Convolutional Neural Network) layer (an existing technology, which will not be described in detail), and inputs them into the fusion unit.

[0056] The fusion unit concatenates the shallow global texture information with the deep local semantic information to obtain the fused feature, which is then input into the fully connected layer unit;

[0057] Fully connected layer units and softmax classifiers or detectors perform calculations based on fused features, outputting target classification probabilities or target bounding boxes;

[0058] Furthermore, damage label sets are added based on the correspondence between historical damage and historical damage data. And the damage label set L ε Compared with historical damage dataset X ε A one-to-one correspondence in the space-time domain, X ε (T a M a ,L a H a ) corresponds to L ε (T a M a ,L a H a ),in correspond i = 1, 2, ..., 5;

[0059] It should be noted that the damage label set L ε Includes information on the type and size of the damage. In this embodiment, wear, repair, and peeling are selected as external damage, and the damage label is used. Clearly indicate the type and area of ​​external damage;

[0060] Furthermore, the intelligent decision-making module (4) is deployed in a cloud computing center, where cloud computing resources complete the training and decision output of the intelligent decision-making model set, and it does not accept multi-source sensor data without a credit rating as input.

[0061] Furthermore, Figure 4 The intelligent decision-making module (4) is shown, and its implementation process is as follows:

[0062] Step 20-1: The data fusion unit fuses trusted data. Data from homogeneous sensors is used to obtain fused data.

[0063] Specifically, the measurement errors e2 and e3 of the two eddy current flaw detection sensors are usually independent and both conform to a normal distribution, i.e. Let σ2 and σ3 be the standard deviations of x2 and x3, respectively. Then, {x2, x3} is fused using the Kalman filter algorithm, as shown below:

[0064]

[0065]

[0066]

[0067] Similarly, among them, It is the fused near-surface damage data, also known as the estimate, K k ∈[0,1] is the Kalman gain that minimizes the variance of the estimate, n = 1000, and x is the number of discrete data points within a truncation window. 2i For the i-th discrete data point within a truncated window, It represents the mean of discrete data points within a truncated window;

[0068] Specifically, as described in step 10-2, the external damage data {x4,x5} is extracted by convolution kernel {k4,k5} to obtain shallow features {f4,f5}, and the shallow features are concatenated along the channel dimension to obtain the fused external damage data.

[0069]

[0070] Step 20-2: Train the intelligent decision-making model and establish an intelligent decision-making model set U;

[0071] Specifically, based on the decision-making task at [X] ε |L ε In this embodiment, historical injury data and matching injury labels are selected for training to train the intelligent decision-making model. For external injuries, this embodiment selects... A target detection model is trained with a learning rate of 0.0001, a batch size of 100, and 1000 training epochs. The weight parameters of the target detection model are continuously updated through iterative training to reduce the loss function, ultimately resulting in the trained target detection model. The loss function includes classification loss. cls Loss with bounding box bbx The details are as follows:

[0072] Loss = Loss cls +Loss bbx (4)

[0073]

[0074] Where N is the number of samples in a batch_size, M is the number of injury categories contained in a sample, and y ij The value is 1 if the sample i is the true class, and 0 otherwise. iij The probability that sample i belongs to category j is output by the object detection model;

[0075]

[0076]

[0077] Where A(·) represents the area, Box represents the true bounding box of sample i. i The predicted bounding box output by the object detection model, d(·) represents the calculation of the Euclidean distance, o i w i and h i Box i The center coordinates, width, and height, and They are respectively The center coordinates, width, and height;

[0078] Step 20-3: The intelligent decision-making model set U is based on the fused data according to the decision-making task. The specific target detection model outputs {o} to perform reasoning and decision-making. i ,w i ,h i ,p ij}

[0079] Furthermore, a decision database is used to store a set of intelligent decision-making models U based on a historical injury dataset X. ε The historical decision set Y obtained through reasoning ε After being confirmed by experts, the credibility is transmitted to the system output side as a root of confidence for decision-making.

[0080] Furthermore, the historical decision set Y ε It includes the decision results of multiple intelligent decision-making models, and is compared with the historical injury dataset X. ε A one-to-one correspondence in the space-time domain, specifically, X ε (T a M a ,L a Ha )correspond

[0081] Furthermore, the output protection module connects to the intelligent decision-making module, the decision database, and the execution module. Based on the relevant historical decisions in the decision database, it performs a credibility assessment on the intelligent decision Y1 and transmits the credible intelligent decision to the execution module.

[0082] Furthermore, before the intelligent decision Y is transmitted to the execution module, it must first undergo a reliability assessment by the output protection module;

[0083] Furthermore, the decision database and output protection module are deployed in the railway station edge computing center. The intelligent decision Y is transmitted from the intelligent decision module in the cloud to the output protection module at the edge, and the edge computing resources complete the decision credibility assessment.

[0084] Furthermore, Figure 5 The output protection module's reliable evaluation process for intelligent decision Y includes the following:

[0085] Step 30-1: The intelligent decision-making module bases its decisions on reliable data. The intelligent decision Y1 obtained through reasoning is transmitted to the output protection module;

[0086] Step 30-2: The output protection module determines the trustworthy data. Does it have a Level 2 trusted stamp?

[0087] Step 30-3: If it is not included, it indicates reliable data. With a Level 1 trusted stamp, corresponding to the newly appearing damage, Y1 is given a Level 1 trusted stamp, resulting in a Level 1 trusted decision.

[0088] Step 30-4: Update the decision database based on the spatiotemporal domain (T1, M1, L1, H1) and store the Level 1 trusted decisions. And transmit it to the execution module;

[0089] Step 30-5: If it is present, it indicates reliable data. Corresponding to historical damage, and certain In the spatial domain (M1, L1, H1) and trusted data Matching and outputting the protection module in the historical decision set Y ε Searching for relevant historical decisions Match with decision Y1;

[0090] Step 30-6: Match historical decisions in the spatial domain Transmitted to the output protection module;

[0091] Step 30-7: The output protection module compares decision Y with historical decisions. Are they similar, that is, are they All meet Where t i ∈T Y ;

[0092] It should be noted that the threshold set T Y Each element t in {t1, t2, t3} i The value is determined by the corresponding intelligent decision-making model u. i Output decision y i The range of values ​​is determined;

[0093] Step 30-8: If the conditions are not met, it indicates that the intelligent decision-making model set U has insufficient training or overfitting, and needs to be self-corrected. After correction, the decision-making process is re-inferred and matched.

[0094] Step 30-9: If the conditions are met, it indicates that decision Y belongs to the historical decision set. n ∈Y ε It is inferred from Level 2 trusted data, and a Level 2 trusted stamp ε is added after decision similarity judgment. 1 A Level 2 credible decision is obtained. Transferred to the execution module.

[0095] Furthermore, the execution module is a railway intelligent operation and maintenance robot, which makes reliable decisions based on railway rail operation and maintenance requirements and the output of the reliable system. The corresponding trust level and spatial domain information (M1, L1, H1) are used to maintain and manage the damage at specific locations of the rail.

[0096] Furthermore, the execution module does not accept decisions without a credit rating as input;

[0097] Furthermore, Figure 6 The reliable transmission process is shown. The data reliability root starts from the damage database, is transmitted to multi-source sensor data, then to the intelligent decision model set, and then to intelligent decision. Combined with the decision reliability root provided by the decision database, the reliable intelligent decision is finally output.

[0098] This enables intelligent and reliable rail flaw detection operations, thereby improving the efficiency and accuracy of rail maintenance and reducing the risk of railway safety accidents.

[0099] Given the importance of rails to the railway industry, and based on the aforementioned cloud-edge-device collaborative computing case, the high-confidence decision-making method and system for intelligent rail flaw detection provided by this invention helps rail flaw detection operations transform from traditional passive detection to proactive intelligent detection, improves the efficiency and accuracy of rail maintenance, reduces the risk of railway safety accidents, and thus promotes the autonomy, intelligence, and reliability of rail maintenance operations.

Claims

1. A high-confidence decision system for intelligent track flaw detection, characterized in that: It includes a data acquisition module (1), a damage database (2), an input protection module (3), an intelligent decision-making module (4), a decision database (5), an output protection module (6), and an execution module (7), wherein: The data acquisition module (1) is used to collect multi-source sensor data of rails in a specific time and space domain in real time. The damage database (2) is used to store historical damage information of rails in various sections of the railway. The input protection module (3) is used to match and compare the multi-source sensor data with the historical damage data related to the damage database. The intelligent decision module (4) is used to process and integrate reliable data, and make intelligent decisions based on the intelligent decision model set according to railway industry constraints and operation and maintenance strategies. The decision database (5) is used to store the historical decision set obtained by the intelligent decision model set based on the historical damage dataset. The output protection module (6) is used to match and compare the decision inferred by the intelligent decision model set with the historical decision set that may be related. The execution module (7) performs maintenance management of rails according to the railway rail operation and maintenance requirements, the reliable decision output by the reliable system, and the corresponding reliability level. Input protection module (3) for multi-source sensor data The credibility assessment process is as follows: Step 10-1: The data acquisition module utilizes its own computing resources to process the acquired multi-source sensor data. The multi-source sensor data undergoes preprocessing. ,in, To collect data on internal damage to the rails, This data represents near-surface damage to the rail. Data on external damage to the rails; The preprocessing workflow includes spatiotemporal alignment, outlier removal, median filtering, mean normalization, and truncation. Spatiotemporal alignment refers to aligning multi-source sensor data based on the spatiotemporal information provided by GNSS signals to ensure... , , Describes the multi-source information of rails in the same spatiotemporal domain; truncation refers to truncating multi-source sensor data according to the sampling frequency of the multi-source sensor data and the time domain information sequence; Step 10-2, Historical Injury Dataset With multi-source sensor data All are composed of a set of convolution kernels Shallow features were extracted, and the following were obtained respectively. The corresponding shallow feature matrix and The corresponding shallow feature matrix ,Right now ,in for convolution The obtained shallow features, ,in for convolution The obtained shallow features, " represents convolution; Extract damage-related features from multi-source sensor data; Step 10-3, in the feature matrix Three initial cluster centers are set in the feature space, according to 3D space Euclidean distance Calculate the distance from each sample in the damaged dataset to its cluster center and minimize it. Iterate through the cluster centers. Clustering Until stable; Step 10-4: Input protection module to judge multi-source sensor data Does it have a historical damage dataset? The characteristics of the data, namely ,whether ,satisfy <threshold ; Step 10-5: If it does not exist, then the multi-source sensor data does not belong to the historical damage dataset, that is... Furthermore, the lack of damage data characteristics indicates that the corresponding rail is in the spatiotemporal domain. No damage was found; the input protection module cleared the multi-source sensor data. ;in, For time domain coordinates, Longitude coordinates Latitude coordinates These are height coordinates; Step 10-6: If it exists, then multi-source sensor data It may belong to a historical damage dataset. The input protection module is in the damage dataset. In the middle, we attempted to search for relevant historical damage data. With multi-source sensor data match; Step 10-7: Input protection module determines whether it is multi-source sensor data. Related historical injury data In the spatial domain Whether the match was successful, i.e., whether... ,satisfy < ; Step 10-8: If not found, the matching fails, and multi-source sensor data is returned. It may belong to a historical damage dataset. This indicates that it may represent newly emerging damage, and a Level 1 trusted stamp is added after damage data feature identification. Level 1 trusted data was obtained. ; Steps 10-9: Based on the spatiotemporal domain Update the damage database and store the Level 1 trusted data. The data is then transmitted to the intelligent decision-making module for decision recognition. Step 10-10: If a match exists, the match is successful, indicating multi-source sensor data. It may belong to a historical damage dataset. Match historical injury data Transmitted to the input protection module; Steps 10-11: Input protection module to compare multi-source sensor data Compared with historical damage data Whether the feature spaces are similar, i.e., whether... Satisfying 2 or more correlation coefficients ,in , For covariance, For mathematical expectation, Features The mean, Features standard deviation and Similarly, we can obtain; Steps 10-12: If not, it indicates that there is a sampling error or malfunction in the multi-source sensor, which needs to be checked and corrected. After correction, the multi-source sensor data should be collected again and matched. Steps 10-13: If they exist, it indicates that the multi-source sensor data belongs to the historical damage dataset. After damage data feature identification and spatiotemporal information matching, a Level 2 trusted stamp is added. Level 2 trusted data was obtained. The data is then transmitted to the intelligent decision-making module for decision recognition.

2. The high-confidence decision system for intelligent track flaw detection as described in claim 1, characterized in that: The data acquisition module (1) is equipped with multi-source sensors, an antenna, a GNSS signal receiver, communication equipment, a walking device, a power supply, a processor, and a memory. It has clock synchronization, spatial positioning, data acquisition, wireless communication, data processing, and data storage functions. It moves along the rail and collects multi-source sensor data in a specific spatiotemporal domain. ; The multi-source sensor system includes an ultrasonic sensor array, two eddy current flaw detection sensors, and two vision sensors, which are used to collect data on internal rail damage. Near-surface damage data External damage data ,Right now ; The ultrasonic sensor array consists of a 0° probe, a forward 70° probe, a reverse 70° probe, a forward 37° probe, and a reverse 37° probe. Echoes detected by different probes at the same location are represented by pixels of different colors. The array moves along the rail with the data acquisition module, and the ultrasonic sensor array data... Presented as continuous two-dimensional image data; One-dimensional data; It is two-dimensional image data; The data acquisition module provides spatiotemporal information of the flaw detection location (i.e., the data acquisition location) to the multi-source sensors based on GNSS signals, including: time domain coordinates. Longitude coordinates Latitude coordinates Height coordinates ,Right now This adds a spatiotemporal stamp to the data, that is, for any coordinate in the spatiotemporal domain... Multi-source sensor data is represented as .

3. The high-confidence decision system for intelligent track flaw detection as described in claim 1, characterized in that: The damage database (2) contains historical damage information that corresponds one-to-one with historical damage in the spatiotemporal domain, including the temporal coordinates of historical damage. Longitude coordinates Latitude coordinates Height coordinates With the corresponding historical damage dataset ,in and Data from the same type of sensor.

4. The high-confidence decision system for intelligent track flaw detection as described in claim 1, characterized in that: The intelligent decision-making module (4) has data fusion and decision-making functions, consisting of a data fusion unit and an intelligent decision-making model set. The data fusion unit combines data from homogeneous sensors to reduce redundancy and improve data accuracy from multi-source sensor data, forming an intelligent decision-making model set. Based on trusted data Output intelligent decision ; Intelligent decision-making model , , All are artificial intelligence models, among which Internal damage data For input, Near-surface damage data For input, External damage data For each input, the output will be the type and size of the damage. ; Artificial intelligence models use neural networks as basic units and adopt a general structure, which specifically includes feature extraction units, feature fusion units, fully connected layer units, and softmax classifiers or detection heads. The softmax classifier is a classification model, and the detection head is a target detection model. The feature extraction unit extracts shallow global texture information and deep local semantic information of the input data layer by layer based on CNNlayer, and inputs them into the feature fusion unit. The feature fusion unit concatenates the shallow global texture information with the deep local semantic information to obtain fused features, which are then input into the fully connected layer unit. Fully connected layer units and softmax classifiers or detectors perform calculations based on fused features, outputting target classification probabilities or target bounding boxes; Add a damage tag set based on the correspondence between historical damage and historical damage data. And damage tag set Compared with historical damage datasets A one-to-one correspondence in the space-time domain correspond ,in correspond , ; Damage Tag Set Includes information on the type and size of the damage. Select wear, repair, and peeling as external damage, and add a damage label. Clearly indicate the type and area of ​​external damage.

5. The high-confidence decision system for intelligent track flaw detection as described in claim 4, characterized in that: The intelligent decision-making module (4) and its implementation process are as follows: Step 20-1: The data fusion unit fuses trusted data. Data from homogeneous sensors is used to obtain fused data. ; Two measurement errors and They are independent and all conform to a normal distribution, i.e. , , and They are respectively and If the standard deviation is , then The Kalman filter algorithm is used for fusion, as shown below: (1) (2) (3) The calculation formula and Similarly, among them, It is the fused near-surface damage data, also known as the estimate. It is the Kalman gain that minimizes the variance of the estimate. The number of discrete data points within a truncated window. For a truncated window, the first A discrete data point It represents the mean of discrete data points within a truncated window; External damage data By convolution kernel shallow features were extracted. The external damage data is obtained by stitching together the shallow features along the channel dimension. ; Step 20-2: Train the intelligent decision-making model and establish an intelligent decision-making model set. ; According to the decision-making task at [ Select historical injury data and matched injury labels for training to train the intelligent decision-making model. For external injuries, select [ The object detection model is trained by iteratively updating its weight parameters to reduce the loss function, ultimately resulting in the trained object detection model. The loss function is... Including classification loss With bounding box loss The details are as follows: (4) (5) in, This represents the number of samples in a batch_size. This represents the number of injury categories contained in a single sample. For the sample The value is 1 if it represents the true category, and 0 otherwise. Samples output by the object detection model Category The probability of; (6) (7) in, Indicates area, For the sample The true bounding box, The predicted bounding box output by the object detection model. This indicates the calculation of Euclidean distance. , and They are respectively The center coordinates, width, and height, , and They are respectively The center coordinates, width, and height; Step 20-3, Intelligent Decision Model Set Based on decision-making tasks and fused data The specific target detection model outputs the reasoning and decision-making process. .

6. The high-confidence decision system for intelligent track flaw detection as described in claim 1, characterized in that: Decision database (5) is used to store a set of intelligent decision models. Based on historical injury datasets Historical decision set obtained through reasoning After being confirmed by experts, the credibility is transmitted to the system output side as a root of confidence for decision-making. Historical Decision Collection It includes decision results from multiple intelligent decision-making models and is compared with historical injury datasets. A one-to-one correspondence in the space-time domain, specifically... correspond .

7. The high-confidence decision system for intelligent track flaw detection as described in claim 1, characterized in that: The output protection module (6) is for intelligent decision-making. Credibility assessment process: Step 30-1: The intelligent decision-making module bases its decisions on reliable data. Intelligent decision-making based on reasoning Transmitted to the output protection module; Step 30-2: The output protection module determines the trustworthy data. Does it have a Level 2 trusted stamp? Step 30-3: If it is not included, it indicates reliable data. It has a Level 1 trust stamp, corresponding to newly appearing damage. Add a Level 1 trusted stamp to obtain a Level 1 trusted decision. ; Step 30-4, Based on the spatiotemporal domain Update the decision database to store the Level 1 trusted decisions. And transmit it to the execution module; Step 30-5: If it is present, it indicates reliable data. Corresponding to historical damage, and certain In the spatial domain Above and trusted data Matching and outputting the protection module in the historical decision set Searching for relevant historical decisions With decision match; Step 30-6: Match historical decisions in the spatial domain Transmitted to the output protection module; Step 30-7: Output protection module comparison and decision With historical decision Are they similar, that is, are they All satisfy ,in ; It should be noted that the threshold set ={ Each element in} The value is determined by the corresponding intelligent decision-making model. Output decision The range of values ​​is determined; Step 30-8: If the conditions are not met, it indicates that the set of intelligent decision-making models is incomplete. If there is undertraining or overfitting, self-correction is required. After correction, reasoning and decision-making should be re-initiated and matching comparisons should be performed. Step 30-9: If satisfied, it indicates a decision. Belongs to the historical decision set, It is based on inference from Level 2 trusted data, and a Level 2 trusted stamp is added after decision similarity judgment. A Level 2 credible decision is obtained. It is then transmitted to the execution module.

8. The high-confidence decision system for intelligent track flaw detection as described in claim 1, characterized in that: The execution module (7) is a railway intelligent operation and maintenance robot, which makes reliable decisions based on the railway rail operation and maintenance requirements and the reliable system output. With corresponding trust level and spatial domain information Maintenance and management of damage to specific locations on the rails.

Images