Urban rail transit maintenance tool counting method and system

Abstract

The invention relates to the technical field of urban rail transit maintenance, and discloses an urban rail transit maintenance tool counting method and system. According to the method, a multi-modal tool data set composed of a three-dimensional scanning image, a radio frequency identification signal and an inventory record table is obtained, geometric features of tools are extracted through a point cloud segmentation algorithm, a distribution thermodynamic diagram is generated according to the radio frequency identification signal, and inventory records are processed to obtain historical use frequency; and inputting the data into a multi-source fusion model to generate a counting vector, constructing an optimal counting path diagram through a dynamic programming algorithm, and outputting a counting result. The system comprises a multi-modal data acquisition module, a geometric feature extraction module, a distribution thermodynamic analysis module and the like. The tool checking efficiency and accuracy are improved, the management cost is reduced, informatization management and tracing are achieved, and efficient development of urban rail transit maintenance work is facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of urban rail transit maintenance, in particular to a method and system for checking maintenance tools and instruments of urban rail transit. BACKGROUND

[0002] In the efficient operation of urban rail transit, the smooth development of maintenance work is the key link to ensure the safe and stable operation of trains, and the effective management of tools and instruments is an important basis for the orderly progress of maintenance work. With the continuous expansion of urban rail transit, the number of lines increases, the number of stations increases, and the operating mileage continues to grow. Correspondingly, the types and quantities of maintenance tools and instruments are also increasing rapidly. The traditional manual counting of tools and instruments has gradually exposed many drawbacks and has been difficult to meet the needs of the rapid development of modern rail transit.

[0003] From the efficiency point of view, the manual counting process of tools and instruments is tedious and time-consuming. Maintenance personnel need to check the name, quantity, specification and other information of each tool and instrument one by one. When facing a large number of tools and instruments, the entire counting process is very long. For example, in some large vehicle depots, the number of maintenance tools and instruments may reach thousands of types, and manual counting often takes several hours or even a whole day, which seriously affects the overall progress of maintenance work, resulting in low efficiency of maintenance, and may delay the normal maintenance plan of the train, increase the downtime of the train, and cause great economic losses to the operation of urban rail transit.

[0004] In terms of accuracy, manual counting is prone to omissions and errors. On the one hand, long-term repetitive work can easily cause fatigue in maintenance personnel, and it is difficult to maintain high concentration at all times, so some tools and instruments may be missed or similar tools and instruments may be confused during the counting process, resulting in errors in the counting results. On the other hand, errors may occur when manually recording data, further reducing the accuracy of the counting data. Such inaccurate counting results will cause great trouble to the subsequent management and deployment of tools and instruments, such as when maintenance work requires the use of tools and instruments, the actual quantity may not match the record, and the required tools may not be found, delaying the maintenance work, or even because of the use of defective or incorrect tools and instruments, causing damage to the train and creating safety hazards.

[0005] In terms of management cost, the manual counting method requires a large amount of manpower and time cost. In order to complete the counting work of tools and instruments, the operating unit has to arrange special personnel to do this work, which not only increases the labor cost, but also occupies a large amount of time of the maintenance personnel due to the low efficiency of manual counting, so that they cannot devote more energy to more valuable maintenance technology improvement and equipment maintenance work. In addition, due to the difficulty of real-time sharing and integration of manual counting data, the operating unit is difficult to effectively monitor and analyze the use and inventory of tools and instruments, and cannot make reasonable procurement and deployment decisions in a timely manner, further increasing the management cost and operation risk.

[0006] In terms of traceability and information management, the traditional manual counting method has serious shortcomings. When tools and instruments are lost or damaged, it is difficult to accurately trace their use history and flow process through manual records, which is not conducive to finding out the causes and holding people accountable. At the same time, the data of manual counting is mostly recorded in paper documents or simple electronic spreadsheets, which is difficult to effectively interface with modern information management systems, cannot realize the informatization and intelligentization of tool and instrument management, and cannot meet the development trend of intelligent operation and management of urban rail transit.

[0007] With the continuous advancement of the intelligent development of urban rail transit, the efficiency, accuracy and intelligence of maintenance tool and instrument management are increasingly required, and the traditional manual counting method has become a bottleneck restricting the improvement of the quality and efficiency of urban rail transit maintenance work. Therefore, it is urgent to develop a method and system for quickly, accurately and intelligently counting urban rail transit maintenance tools and instruments SUMMARY

[0008] The purpose of the present application is to provide a method and system for counting urban rail transit maintenance tools and instruments to solve the problems raised in the background art.

[0009] To achieve the above purpose, the present application provides the following technical solution: a method for counting urban rail transit maintenance tools and instruments, the method comprising:

[0010] Obtaining a set of multi-modal tool and instrument data; the multi-modal tool and instrument data includes three-dimensional scanning images, radio frequency identification signals and inventory record tables; the three-dimensional scanning images contain tool and instrument contour and surface texture information, and the radio frequency identification signals contain tool and instrument number and position labels;

[0011] Based on the three-dimensional scanning images, the tool and instrument geometric features are extracted by a point cloud segmentation algorithm, and the geometric features include size parameters, shape categories and surface wear marks;

[0012] According to the radio frequency identification signal, a tool distribution heat map is generated through signal clustering analysis, and the distribution heat map includes a tool dense area and an abnormal missing position.

[0013] The inventory record table is subjected to time sequence alignment processing to generate a tool historical use frequency.

[0014] The geometric features, the distribution heat map and the historical use frequency are input into a multi-source fusion model to generate a tool counting vector.

[0015] Based on the counting vector, an optimal counting path graph is constructed through a dynamic programming algorithm, and a tool counting result is output; the nodes of the optimal counting path graph represent tool storage positions, and the edges represent counting sequences and priority weight.

[0016] Preferably, the geometric features of the tools are extracted through a point cloud segmentation algorithm, including:

[0017] The three-dimensional scanning image is subjected to noise reduction filtering processing to generate denoised point cloud data.

[0018] Based on a pre-defined tool reference model library, a nearest neighbor matching algorithm is used to identify tool types and segment independent point cloud clusters.

[0019] A curvature calculation algorithm is used to extract the surface curvature of the point cloud clusters, and a size deviation value is generated in combination with the standard size parameters in the reference model library.

[0020] The size deviation value, the shape category and the surface wear mark are encoded as structured geometric features.

[0021] Preferably, the tool distribution heat map is generated through signal clustering analysis, including:

[0022] The radio frequency identification signal is subjected to intensity calibration, and noise data with a signal intensity lower than a threshold value is removed.

[0023] Based on a density clustering algorithm, tool aggregation areas are divided, and signal density values of the areas are calculated.

[0024] According to the signal density value and a pre-set tool quantity threshold value, a heat level is generated.

[0025] The heat level is associated with the abnormal missing position to form a two-dimensional distribution heat map.

[0026] Preferably, the multi-source fusion model includes a feature compression module and a space-time correlation module, and the feature compression module includes:

[0027] The size parameters in the geometric features are subjected to normalization processing to obtain a first compressed vector.

[0028] Discretely encode the heat level in the distribution heat map to generate a second compression vector;

[0029] Slide window statistics on the historical usage frequency to extract the frequency change trend to obtain a third compression vector;

[0030] Merge the first compression vector, the second compression vector and the third compression vector into a low-dimensional feature sequence through a feature splicing layer.

[0031] Preferably, the optimal inventory path graph is constructed by a dynamic programming algorithm, comprising:

[0032] Initialize node coordinates according to the storage positions of tools and instruments, and generate an edge weight matrix based on priority weights;

[0033] The inventory vector is taken as a node attribute, and the edge weight matrix is composed of time cost of inventory order and priority weights;

[0034] Iteratively calculate the shortest path value of each node through a state transition equation, and update the edge weight matrix;

[0035] Generate an optimal inventory path sequence covering all nodes according to the updated edge weight matrix.

[0036] Preferably, the construction method of the tool and instrument reference model library comprises:

[0037] Collect three-dimensional point cloud data of multiple standard tools and instruments, and extract reference dimensions and shape contour features;

[0038] Perform affine transformation on the shape contour features to generate multi-view reference models;

[0039] Classify the reference models according to tool and instrument types, and associate with a standard parameter database;

[0040] Store the classified reference models as a reference model library, and regularly update the models based on new scanning data.

[0041] Preferably, the parameter optimization method of the density clustering algorithm comprises:

[0042] Calculate the initial clustering radius and the minimum sample number according to the distribution of historical signal data;

[0043] Select the parameters with the highest matching degree of clustering results and artificial inventory data by traversing the parameter combinations through a grid search algorithm;

[0044] Dynamically adjust the clustering radius and the minimum sample number according to the matching degree to optimize the heat level division precision.

[0045] Preferably, the space-time correlation module comprises:

[0046] Timestamp alignment is performed on the low-dimensional feature sequence to generate a time correlation matrix;

[0047] A spatial correlation matrix is generated by extracting spatial correlation features through a convolution attention mechanism.

[0048] The time correlation matrix and the spatial correlation matrix are subjected to Hadamard product operation to generate spatiotemporal fusion features.

[0049] The spatiotemporal fusion features are superimposed on the original low-dimensional feature sequence through residual connection to output the inventory vector.

[0050] Preferably, the construction method of the state transition equation comprises:

[0051] The transition cost between nodes is defined as the weighted sum of the time cost and the priority weight.

[0052] The path value of each node is initialized to infinity, and the path value of the starting point is zero.

[0053] The minimum path value of each node based on the predecessor node is calculated through a recursive formula, and the optimal predecessor node is recorded.

[0054] The complete inventory path sequence is generated by backtracking according to the optimal predecessor node.

[0055] Preferably, the present application further comprises a city rail transit maintenance tool inventory system, which comprises:

[0056] A multi-modal data acquisition module is configured to acquire a multi-modal tool data set, wherein the multi-modal tool data comprises three-dimensional scanning images, radio frequency identification signals and inventory record tables; wherein the three-dimensional scanning images contain tool contour and surface texture information, and the radio frequency identification signals contain tool numbers and position labels.

[0057] A geometric feature extraction module is configured to extract tool geometric features based on the three-dimensional scanning images through a point cloud segmentation algorithm, wherein the geometric features include size parameters, shape categories and surface wear marks.

[0058] A distribution thermal analysis module is configured to generate a tool distribution thermal map based on the radio frequency identification signals through signal clustering analysis, wherein the distribution thermal map contains tool dense areas and abnormal missing positions.

[0059] An inventory record processing module is configured to perform time sequence alignment processing on the inventory record table to generate tool historical use frequency.

[0060] A multi-source fusion module is configured to input the geometric features, the distribution thermal map and the historical use frequency into a multi-source fusion model to generate a tool inventory vector; wherein the multi-source fusion model comprises a feature compression module and a spatiotemporal correlation module.

[0061] The inventory path generation module is used for constructing an optimal inventory path graph by a dynamic programming algorithm based on the inventory vector, and outputting a tool inventory result; the nodes of the optimal inventory path graph represent tool storage positions, and the edges represent inventory sequences and priority weights.

[0062] Compared with the prior art, the present application has the following advantages:

[0063] Firstly, the present application acquires a multi-modal tool data set, including three-dimensional scanning images, radio frequency identification signals and inventory record tables and other multi-source information. By using three-dimensional scanning images, the point cloud segmentation algorithm can quickly extract tool geometric features, including size parameters, shape categories and surface wear marks, etc. This kind of fast processing method based on images greatly saves time compared with manual identification and measurement. At the same time, according to the signal clustering analysis of the radio frequency identification signals, a tool distribution heat map is generated, which can intuitively present the tool dense area and abnormal missing position, so that the maintenance personnel can quickly locate the tool position and reduce the time waste of finding tools. In addition, the multi-source fusion model fuses the geometric features, distribution heat map and historical use frequency to generate an inventory vector, and then combines a dynamic programming algorithm to construct an optimal inventory path graph, which plans the best inventory sequence for the maintenance personnel, avoids unnecessary walking and repeated operation, and makes the entire inventory process more efficient and orderly. With these technical means, the tool inventory time is greatly shortened. For example, in a medium-sized vehicle depot, the tool inventory time can be shortened from several hours to tens of minutes by using the method, which effectively improves the overall progress of the maintenance work, ensures that the train can complete the maintenance task on time, reduces the train downtime, and provides a guarantee for the efficient operation of urban rail transit.

[0064] In the data acquisition stage, the three-dimensional scanning image accurately records the contour and surface texture information of the tool, and the radio frequency identification signal contains the accurate tool number and position label, which provides an accurate basis for subsequent analysis. In the geometric feature extraction process, through a series of operations such as noise filtering processing of the three-dimensional scanning image, nearest neighbor matching algorithm identification based on the pre-defined tool reference model library, the tool type can be accurately determined, the geometric features can be extracted, and the size deviation value can be generated, which effectively avoids the errors caused by the difficulty in distinguishing similar tools during manual identification. When generating the distribution heat map in the signal clustering analysis, the radio frequency identification signal is calibrated in strength, the noise data is removed, and the density clustering algorithm is used to accurately divide the tool aggregation area, so that the heat map can accurately reflect the actual distribution of the tool, and the possibility of missing points and wrong points is reduced. Moreover, the application of the multi-source fusion model and the dynamic programming algorithm is based on accurate data for calculation and analysis, which further ensures the accuracy of the inventory result, provides reliable data support for the management and allocation of tools, and reduces the operation risk caused by inaccurate tool quantity.

[0065] From the perspective of management cost, the present application has significant cost saving advantages. Due to the significant improvement in inventory efficiency, the tool inventory work that originally requires a large amount of manpower and time can now be completed with less personnel and shorter time, which directly reduces the investment in labor cost. At the same time, the accurate inventory result makes the management and allocation of tools more scientific and reasonable, and the operating unit can adjust the procurement plan in a timely manner according to the actual situation, avoiding the situation of over-purchasing or insufficient purchasing caused by unclear tool quantity, thereby saving the procurement cost. In addition, through the analysis of the historical use frequency of the tool, the wear and damage of the tool can be predicted in advance, the maintenance and replacement plan can be reasonably arranged, the equipment damage and maintenance cost caused by tool failure can be reduced, the whole life cycle cost control is realized, and the economic benefit of urban rail transit maintenance work is improved.

[0066] In terms of information management and traceability, the present application has outstanding advantages. The multi-modal data obtained by the system can be conveniently stored and integrated into the information management platform, realizing real-time sharing and analysis of data. When the tool is lost, damaged or other problems occur, the use history, flow process and responsible person of the tool can be quickly traced through the query and analysis of the inventory record table, tool number and historical use data, so that the problem can be found out in time and corresponding measures can be taken. Moreover, various data reports and analysis results generated by the system provide strong support for the decision-making of the operating unit, which helps to realize the intelligentization and scientization of urban rail transit maintenance tool management, conforms to the trend of intelligentization development of urban rail transit, and improves the overall operation and management level. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 The working principle diagram of the city rail transit maintenance tool inventory method of the present application;

[0068] Figure 2 The flowchart for tool geometry feature extraction;

[0069] Figure 3 The flowchart for tool reference model library construction and updating;

[0070] Figure 4 The working flowchart for generating the inventory path based on the state transition equation. DETAILED DESCRIPTION

[0071] 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.

[0072] Please refer to Figures 1-4 The present application provides a technical solution: a city rail transit maintenance tool inventory method, which comprises:

[0073] A specific device is used to collect three-dimensional scanning images, radio frequency identification signals and inventory record tables. The three-dimensional scanning images contain tool contour and surface texture information, which can intuitively show the appearance of the tool. The radio frequency identification signals contain tool numbers and position labels, which are used for accurate positioning. The inventory record tables record the historical use of the tool.

[0074] Point cloud segmentation algorithm is used to extract the geometric features of the tool from the three-dimensional scanning images, including size parameters, shape categories and surface wear marks. These features help to determine the specifications, types and wear degrees of the tool.

[0075] A tool distribution heat map is generated according to the radio frequency identification signals: through signal clustering analysis, the radio frequency identification signals are processed to generate a distribution heat map containing tool dense areas and abnormal missing positions, which facilitates quick understanding of the distribution of the tool.

[0076] The inventory record table is processed by time sequence alignment to generate tool historical use frequency, which provides data support for subsequent analysis.

[0077] The geometric features, distribution heat map and historical use frequency are input into a multi-source fusion model to generate a tool inventory vector, realizing the fusion of multi-source data.

[0078] Based on the inventory vector, an optimal inventory path graph is constructed by a dynamic programming algorithm, and the tool inventory result is output. The nodes of the optimal inventory path graph represent the tool storage locations, and the edges represent the inventory order and priority weight, thereby improving the inventory efficiency.

[0079] The application will be further described below in combination with Examples 1 to 5:

[0080] Example 1:

[0081] In this embodiment, the process of extracting the geometric features of the tools by the point cloud segmentation algorithm is elaborated. After obtaining the three-dimensional scanning image, since there may be noise interference in the image, which affects the accuracy of geometric feature extraction, the three-dimensional scanning image is first subjected to noise reduction filtering processing. A Gaussian filtering algorithm is adopted, the principle of which is to perform weighted average on the gray values of the surrounding pixel points of each pixel point in the image, so that the image becomes smooth and the noise is removed. After Gaussian filtering processing, denoising point cloud data is generated.

[0082] Based on the pre-defined tool reference model library, the tool type is identified by a nearest neighbor matching algorithm, and independent point cloud clusters are segmented. The nearest neighbor matching algorithm calculates the distance between each point in the denoising point cloud data and the model points in the reference model library, and the tool type corresponding to the model with the shortest distance is taken as the identification result. Assuming that P is a point in the denoising point cloud data, and Q i , i = 1, 2, …, n, the distance formula is calculated as:

[0083]

[0084] where P x , P y , P z are the three-dimensional coordinates of point P, Q ix , Q iy , Q iz are the three-dimensional coordinates of model point Q i . Find the minimum Q j , then consider that the tool type to which the point belongs is consistent with the model corresponding to Q j , and then segment the independent point cloud cluster.

[0085] The curvature calculation algorithm is used to extract the surface curvature of the point cloud cluster. Curvature is a quantity that describes the degree of curvature of a curve or surface. By calculating the curvature of each point in the point cloud cluster, the concave-convex situation of the tool surface can be understood. The size deviation value is generated in combination with the standard size parameters in the reference model library. The size deviation value reflects the difference between the actual tool size and the standard size, which is of great significance for judging whether the tool meets the use standard.

[0086] Finally, the size deviation value, shape category and surface wear mark are encoded as structured geometric features. A binary encoding method is used to convert different feature values into binary digital strings, facilitating subsequent data processing and analysis. For example, for the shape category, set the circle as "001", the square as "010", etc.; for the size deviation value, according to the deviation range, different intervals are divided, and each interval corresponds to a binary code; for the surface wear mark, according to the wear degree, it is divided into light wear "100", medium wear "101", heavy wear "110", etc. In this way, the geometric features of the tool are represented in a structured form, providing basic data for subsequent multi-source fusion and inventory work.

[0087] Example 2

[0088] This embodiment describes in detail the specific process of generating a tool distribution heat map through signal clustering analysis. After obtaining the radio frequency identification signal, since the signal may be disturbed during transmission, resulting in unstable signal strength, the radio frequency identification signal is first calibrated for strength. The mean filtering method is used to average the signal strength values received within a period of time, removing abnormal fluctuation noise data. Assuming that n signal strength values S i , i = 1, 2, …, n are received within the time interval [t1, t2], the calibrated signal strength value S avg is calculated as follows:

[0089]

[0090] The noise data with signal strength lower than the threshold value S th is removed, and S th is a standard value determined according to experience and experiment, and the signal below this value is considered invalid or unreliable.

[0091] Based on the density clustering algorithm, the tool aggregation area is divided, and the signal density value of each area is calculated. The density clustering algorithm is based on the density of data points for clustering, and the data points with connected density are divided into the same class. Let the neighborhood of data point x be N ∈ (x), where ∈ is the neighborhood radius, and the number of data points in the neighborhood is n ∈ (x), then the signal density value ρ(x) is calculated as follows:

[0092]

[0093] where V is the volume of the neighborhood N ∈ (x). By calculating the signal density value of each data point, it can be judged which area has a higher data point density, i.e. the tool aggregation area.

[0094] According to the signal density value and the preset tool quantity threshold Nth Generate heat level. When the signal density value p(x) is greater than or equal to N th , divide into high heat level, indicating that the area is dense with tools; when p(x) is less than N th , divide into low heat level.

[0095] Finally, associate the heat level with the abnormal missing position to form a two-dimensional distribution heat map. In the heat map, different colors or gray levels are used to represent different heat levels, making it easy and intuitive to observe the distribution of tools. For abnormal missing positions, determine by comparing the expected tool distribution and the actual detected signal situation, mark these positions in the heat map, such as using a red cross to indicate, so that workers can quickly find and handle. The distribution heat map generated in this way can provide an important reference for the inventory and management of tools.

[0096] Example 3:

[0097] The feature compression module processes and fuses different types of feature data to generate low-dimensional feature sequences for subsequent analysis and processing.

[0098] First, normalize the size parameters in geometric features. Since the size parameters of different tools may have different magnitudes and ranges, in order to make these parameters comparable, a normalization method is used to map them to the [0,1] interval. Assuming the size parameter is x, its minimum value is x min , and its maximum value is x max , then the first compressed vector x norm The calculation formula is:

[0099]

[0100] In this way, the size parameters of different tools are compared on the same scale, eliminating the influence of magnitude differences.

[0101] Next, discretize the heat levels in the distribution heat map to generate a second compressed vector. Heat levels are usually continuous values, in order to facilitate processing, they are divided into different discrete levels. For example, the heat level is divided into three levels: low, medium and high, represented by "0", "1" and "2" respectively. In this way, each heat level can be encoded by a discrete number to form a second compressed vector.

[0102] Then, the historical usage frequency is counted by a sliding window to extract the frequency change trend, and a third compression vector is obtained. The sliding window counting is to set a fixed size window on the historical usage frequency data, and the window slides along the time series to calculate the average value or other statistics of the data in the window to reflect the change trend of the data. By calculating the average value of the sliding window at different time points, the change trend of the historical usage frequency can be observed, and these trend information is taken as a part of the third compression vector.

[0103] Finally, the first compression vector, the second compression vector and the third compression vector are merged into a low-dimensional feature sequence by a feature splicing layer. The feature splicing layer connects the three compression vectors in a certain order to form a new vector, which contains the key information of the geometric features, the distribution heat map and the historical usage frequency, and provides processing data for the subsequent spatio-temporal correlation module, realizing the preliminary fusion of multi-source features.

[0104] Embodiment 4:

[0105] This embodiment details the process of constructing the optimal inventory path graph by dynamic programming algorithm. In constructing the optimal inventory path graph, the node coordinates are initialized according to the tool storage location, and the edge weight matrix is generated based on the priority weight. Assuming that the coordinates of the tool storage location are (x i ,y i ), i = 1, 2, …, n, these coordinates determine the position of the node in two-dimensional space. The priority weight is determined according to the importance, usage frequency and other factors of the tool, and the priority weight between node i and node j is w ij . The element w ij in the edge weight matrix W represents the weight of the edge from node i to node j, and the edge weight matrix is composed of the time cost of the inventory order and the priority weight. Assuming that the time cost from node i to node j is t ij , the calculation formula of the edge weight matrix element w ij is:

[0106] w ij =t ij +α×w ij

[0107] Wherein, α is the weight coefficient, used to adjust the relative importance of the time cost and the priority weight in the edge weight, 0≤α≤1. By adjusting the value of α, the inventory path can be optimized according to the actual demand.

[0108] After the inventory vector is taken as the node attribute and the edge weight matrix is determined, the shortest path value of each node is iteratively calculated through the state transition equation, and the edge weight matrix is updated. The transition cost between nodes is defined as the weighted sum of the time cost and the priority weight, and the path value of each node is initialized as infinity, and the path value of the starting point is zero. Let the path value of node i be d(i), then the recursive formula of the state transition equation is:

[0109]

[0110] where N(j) represents the predecessor node set of node j. Through continuous iteration calculation, the minimum path value of each node based on the predecessor node is found, and the optimal predecessor node is recorded.

[0111] The optimal inventory path sequence covering all nodes is generated according to the updated edge weight matrix. Starting from the terminal node, the predecessor nodes of each node are found in reverse according to the recorded optimal predecessor nodes, until the starting node, thereby generating a complete inventory path sequence. This path sequence considers the storage location of the tool, the priority weight and the inventory time cost, and is the optimal inventory path, which can improve the efficiency and accuracy of tool inventory.

[0112] Example 5:

[0113] This embodiment mainly describes the specific workflow of the space-time association module. In the multi-source fusion model, the space-time association module further mines the space-time information in the data to generate more valuable inventory vectors.

[0114] The low-dimensional feature sequence is time-stamped and aligned to generate a time-series association matrix. The low-dimensional feature sequence is output by the feature compression module and contains various feature information of the tool. Since these feature data may be collected at different times, time stamp alignment is needed. Assuming that the low-dimensional feature sequence is X(t), and the time stamp is t i , i = 1, 2, …, n. By aligning the data with the same time stamp, a time-series association matrix T is constructed. The element T ij in matrix T represents the association relationship between feature j and other features at time t i , and its calculation method can be based on data correlation analysis, such as Pearson correlation coefficient. Let the values of feature j and feature k at time t i be X ij and X ik respectively, then the Pearson correlation coefficient calculation formula is:

[0115]

[0116] wherein, and The average value of feature j and feature k at all time points, respectively. By calculating the Pearson correlation coefficient, a time correlation matrix T is obtained, which reflects the correlation of different features in the time dimension.

[0117] The spatial correlation feature is extracted by the convolution attention mechanism, and the spatial correlation matrix is generated. The convolution attention mechanism is a method that can automatically learn important features in data. It extracts features from low-dimensional feature sequences through convolution operations, and then assigns attention weights according to the importance of features. Assuming that the convolution kernel is K, the low-dimensional feature sequence X is convolved to obtain the feature map F, and the calculation formula is:

[0118] F = X-K

[0119] Where "*" represents convolution operation. Then, by calculating the attention weight, the feature map F is weighted to obtain the spatial correlation feature. These spatial correlation features are combined to generate the spatial correlation matrix S, and the elements S ij in the matrix S represent the correlation between features at different spatial positions.

[0120] The time correlation matrix T and the spatial correlation matrix S are subjected to Hadamard product operation to generate spatio-temporal fusion features. Hadamard product is an operation of multiplying corresponding elements of two matrices. Let the spatio-temporal fusion feature matrix be H, then H = T ⊙ S, where "⊙" represents Hadamard product. Through Hadamard product operation, the correlation features in time dimension and space dimension are fused together to obtain more comprehensive spatio-temporal fusion features.

[0121] Finally, the spatio-temporal fusion features are superimposed with the original low-dimensional feature sequence through residual connection to output the counting vector. The purpose of residual connection is to retain the information in the original data and avoid losing important information in the feature fusion process. Let the original low-dimensional feature sequence be X and the spatio-temporal fusion feature be H, then the calculation formula of the output counting vector Y is:

[0122] Y = X + H

[0123] Thus, through the processing of the space-time correlation module, the generated inventory vector not only contains various feature information of the tool, but also fully excavates the correlation information of the space-time dimension, providing more accurate and valuable data support for subsequent tool inventory work. It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or action from another, and do not necessarily require or imply that these entities or actions have any such actual relationship or order. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0124] It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or action from another, and do not necessarily require or imply that these entities or actions have any such actual relationship or order. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0125] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made hereto without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for counting tools for urban rail transit maintenance, characterized in that: include: Acquire a multimodal tool data set; the multimodal tool data includes a three-dimensional scanned image, a radio frequency identification signal, and an inventory record; the three-dimensional scanned image includes tool contour and surface texture information, and the radio frequency identification signal includes a tool number and location tag; Based on the three-dimensional scanned image, extracting geometric features of the tool through a point cloud segmentation algorithm, wherein the geometric features include size parameters, shape categories, and surface wear marks; Generate a tool distribution heat map based on the radio frequency identification signal through signal cluster analysis, wherein the distribution heat map includes tool-intensive areas and abnormal missing locations; Performing time-series alignment processing on the inventory record table to generate historical usage frequencies of tools; Input the geometric features, distribution heat map and historical usage frequency into a multi-source fusion model to generate a tool inventory vector; Based on the inventory vector, an optimal inventory path diagram is constructed through a dynamic programming algorithm, and the tool inventory results are output; the nodes of the optimal inventory path diagram represent the storage locations of tools, and the edges represent the inventory order and priority weights.

2. A method for counting tools for urban rail transit maintenance according to claim 1, characterized in that: The method of extracting geometric features of tools by using a point cloud segmentation algorithm includes: Performing noise reduction filtering on the three-dimensional scanned image to generate denoised point cloud data; Based on a predefined tool benchmark model library, the tool type is identified through the nearest neighbor matching algorithm and independent point cloud clusters are segmented. The curvature calculation algorithm is used to extract the surface curvature of the point cloud cluster, and the size deviation value is generated by combining the standard size parameters in the benchmark model library; The size deviation value, shape category and surface wear mark are encoded as structured geometric features.

3. The method for counting tools for urban rail transit maintenance according to claim 1, characterized in that: Generating a tool distribution heat map through signal cluster analysis includes: Performing intensity calibration on the radio frequency identification signal to remove noise data with signal intensity below a threshold; Divide the tool clustering areas based on the density clustering algorithm and calculate the signal density value of each area; Generate thermal levels based on signal density values ​​and preset tool quantity thresholds; The thermal levels are associated with the locations of anomaly absences to form a two-dimensional distribution thermal map.

4. The method for counting tools for urban rail transit maintenance according to claim 1, characterized in that: The multi-source fusion model includes a feature compression module and a spatiotemporal correlation module, and the feature compression module includes: Normalizing the size parameters in the geometric feature to obtain a first compression vector; Discretize and encode the thermal level in the distribution thermal map to generate a second compressed vector; Performing sliding window statistics on the historical usage frequency, extracting the frequency change trend, and obtaining a third compression vector; The first compressed vector, the second compressed vector and the third compressed vector are merged into a low-dimensional feature sequence through a feature concatenation layer.

5. The method for counting tools for urban rail transit maintenance according to claim 1, characterized in that: The method of constructing the optimal inventory path diagram by a dynamic programming algorithm includes: Initialize node coordinates according to the tool storage location and generate edge weight matrix based on priority weights; The counting vector is used as a node attribute, and the edge weight matrix is ​​composed of the time cost and priority weight of the counting order; Iteratively calculate the shortest path value of each node through the state transfer equation and update the edge weight matrix; Generate the optimal inventory path sequence covering all nodes based on the updated edge weight matrix.

6. The method for counting tools for urban rail transit maintenance according to claim 2, characterized in that: The method for constructing the tool benchmark model library includes: Collect 3D point cloud data of various standard tools and extract benchmark dimensions and shape contour features; Perform affine transformation on shape contour features to generate a multi-view reference model; Classify benchmark models according to tool type and associate them with a standard parameter database; The classified benchmark models are stored as a benchmark model library and are regularly updated based on new scan data.

7. The method for counting tools for urban rail transit maintenance according to claim 3, characterized in that: The parameter optimization method of the density clustering algorithm includes: Calculate the initial cluster radius and minimum number of samples based on the historical signal data distribution; The parameter combinations are traversed by the grid search algorithm, and the parameters with the highest matching degree between the clustering results and the manual inventory data are selected; The clustering radius and the minimum number of samples are dynamically adjusted according to the matching degree to optimize the accuracy of thermal grade division.

8. The method for counting tools for urban rail transit maintenance according to claim 4, characterized in that: The spatiotemporal correlation module includes: Perform timestamp alignment on low-dimensional feature sequences to generate a temporal correlation matrix; Extract spatial correlation features through convolutional attention mechanism and generate spatial correlation matrix; Perform Hadamard product operation on the temporal correlation matrix and the spatial correlation matrix to generate spatiotemporal fusion features; The spatiotemporal fusion features are superimposed on the original low-dimensional feature sequence through residual connection to output the clearing vector.

9. The method for counting tools for urban rail transit maintenance according to claim 5, characterized in that: The method for constructing the state transfer equation includes: The transfer cost between nodes is defined as the weighted sum of time cost and priority weight; Initialize the path value of each node to infinity and the starting point path value to zero; Calculate the minimum path value of each node based on the predecessor node through a recursive formula and record the optimal predecessor node; Generate a complete inventory path sequence by tracing back based on the optimal predecessor node.

10. An urban rail transit maintenance tool inventory system, characterized in that: include: Multimodal data acquisition module: used to acquire a multimodal tool data set, wherein the multimodal tool data includes a 3D scan image, a radio frequency identification signal, and an inventory record; wherein the 3D scan image contains tool outline and surface texture information, and the radio frequency identification signal contains the tool number and location tag; A geometric feature extraction module is configured to extract geometric features of the tool based on the three-dimensional scan image using a point cloud segmentation algorithm, wherein the geometric features include size parameters, shape categories, and surface wear marks; Distribution thermal analysis module: used to generate a tool distribution thermal map based on the RFID signal through signal cluster analysis, wherein the distribution thermal map includes tool-intensive areas and abnormal missing locations; Inventory record processing module: performs time sequence alignment processing on the inventory record table to generate the historical usage frequency of tools; Multi-source fusion module: inputs the geometric features, distribution heat map and historical usage frequency into a multi-source fusion model to generate a tool inventory vector; the multi-source fusion model includes a feature compression module and a spatiotemporal correlation module; Inventory path generation module: used to construct an optimal inventory path diagram based on the inventory vector through a dynamic programming algorithm and output the tool inventory results; the nodes of the optimal inventory path diagram represent the storage locations of tools, and the edges represent the inventory order and priority weights.

Images