Railway maintenance field operation protection card control method, equipment and medium

Abstract

The invention discloses a railway maintenance field operation protection card control method and device and a medium, and relates to the technical field of rail transit equipment, and the method comprises the steps: carrying out the spatial diffusion of a comprehensive risk index value, generating a dynamic risk field, carrying out the field intensity gradient vectorization and path integration of the dynamic risk field, and generating a safety path coordinate sequence; performing spatio-temporal trajectory matching on the safe path coordinate sequence to obtain a path execution deviation degree, performing spatio-temporal causal analysis on the path execution deviation degree, and constructing an event causal diagram; and obtaining the violation contribution degree of the event causal graph through anti-fact query, and carrying out strategic compiling on the basic permission rule base according to the violation contribution degree to generate an incremental control strategy and a visual decision report. According to the method, the dynamic risk field is generated and the event causal graph is constructed, so that the risk root is accurately positioned, and the rule dynamic adaptability and the decision interpretability are enhanced.

Description

Technical Field

[0001] This invention relates to the field of rail transit equipment technology, and in particular to a method, equipment and medium for on-site protection and control during railway maintenance. Background Technology

[0002] As a critical national infrastructure, railway transportation relies heavily on maintenance safety as a core line of defense for train operation safety and the protection of personnel lives. With the expansion of the high-speed rail network and the increase in the capacity of conventional railways, maintenance sites present complex scenarios involving multi-trade collaboration, multi-equipment linkage, and a complex interplay of environmental variables. The traditional protection model, dominated by human experience and constrained by static rules, is no longer sufficient to address dynamic risks. In recent years, the industry has gradually adopted IoT, big data, and AI technologies to optimize protection systems. Some solutions achieve risk warnings by collecting operational and train running data, or use machine learning models to predict risk levels. Other research attempts to generate safe work paths through path planning algorithms or trace the root causes of risks based on causal analysis.

[0003] However, existing technologies still have some shortcomings. Most solutions only extract isolated features from single types of data and fail to achieve cross-domain correlation of job responsibility and authority topology, operational behavior patterns, train operation dynamics, and environmental disturbance time-frequency characteristics. This makes it difficult to capture coupled risks in complex scenarios, resulting in a large deviation between risk quantification results and actual scenarios. Traditional permission rule bases are built based on preset thresholds or empirical logic and cannot dynamically adjust permission strategies according to the real-time calculated contribution of violations. This makes it difficult to cope with the sudden evolution of risks at the maintenance site and also makes it impossible to accurately locate the driving factors of violations for targeted control. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for protecting and controlling railway maintenance site operations to solve the problems of difficulty in capturing coupled risks in complex scenarios and the inability to cope with the evolution of sudden risks and accurately locate driving factors.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for protection and control at railway maintenance sites, which includes collecting historical accident data, job responsibility attributes and multi-source real-time working condition data and performing preprocessing, and extracting the permission mapping features of the preprocessed job responsibility attributes and the working condition risk features of the preprocessed multi-source real-time working condition data.

[0008] The preprocessed historical accident data and permission mapping features are matched and instantiated to generate a basic permission rule base.

[0009] The risk characteristics of the working conditions are input into the risk prediction model. The risk coding layer performs feature mapping and risk association construction, the risk evolution layer performs global optimization search, and outputs a comprehensive risk index value.

[0010] Spatial diffusion is performed on the comprehensive risk index value to generate a dynamic risk field. The dynamic risk field is then vectorized by field strength gradient and path integral to generate a safe path coordinate sequence.

[0011] Spatiotemporal trajectory matching is performed on the safe path coordinate sequence to obtain the path execution deviation. Spatiotemporal causal analysis is then performed on the path execution deviation to construct an event causal graph.

[0012] By obtaining the violation contribution degree of the event cause-effect graph through reverse fact query, the basic permission rule base is compiled into a strategy based on the violation contribution degree, and incremental control strategies and visual decision reports are generated.

[0013] As a preferred embodiment of the railway maintenance site operation protection and control method of the present invention, the multi-source real-time working condition data includes job operation data, train operation data and environmental parameters;

[0014] The preprocessing includes data cleaning, format standardization, and noise filtering.

[0015] As a preferred embodiment of the railway maintenance site operation protection and control method of the present invention, the specific steps for extracting the permission mapping features of the preprocessed job responsibility attributes and the work condition risk features of the preprocessed multi-source real-time work condition data are as follows.

[0016] Perform permission topology encoding on the preprocessed job responsibility attributes to generate permission mapping features;

[0017] Analyze the behavioral patterns of preprocessed job operation data to generate quantitative operational features;

[0018] The preprocessed train operation data is vectorized to obtain operation status characteristics;

[0019] The preprocessed environmental parameters are transformed in the time-frequency domain to generate environmental disturbance characteristics.

[0020] By integrating operational quantitative characteristics, operational status characteristics, and environmental disturbance characteristics, operating condition risk characteristics are generated.

[0021] As a preferred embodiment of the railway maintenance site operation protection and control method of the present invention, the specific steps for performing rule matching and instantiation on preprocessed historical accident data and permission mapping features to generate a basic permission rule base are as follows:

[0022] Perform statistical analysis and rule mining on the preprocessed historical accident data to generate a pre-set rule template library;

[0023] By leveraging permission mapping features, logical conditions are matched against a pre-defined rule template library to generate a set of matching rules.

[0024] The matching rule set is compiled into rules and parameters are instantiated to generate a basic permission rule library.

[0025] As a preferred embodiment of the railway maintenance site operation protection and control method described in this invention, the steps of inputting the working condition risk characteristics into the risk prediction model, constructing feature mapping and risk association in the risk coding layer, performing global optimization search in the risk evolution layer, and outputting a comprehensive risk index value are as follows.

[0026] A risk encoding layer is built using a deep neural network, and a risk evolution layer is built using a random forest algorithm.

[0027] A risk prediction model is constructed by fusing and associating features between the risk encoding layer and the risk evolution layer through an attention mechanism.

[0028] The working condition risk characteristics are input into the risk prediction model. The risk encoding layer uses multi-scale convolutional kernels to perform feature mapping and risk association parsing, generating high-dimensional feature vectors and risk weight matrices.

[0029] The risk evolution layer uses a gradient boosting decision tree to perform a global optimization search on high-dimensional feature vectors to form a comprehensive risk score;

[0030] The high-dimensional feature vector, risk weight matrix, and comprehensive risk score are fused by tensor product and then weighted by attention to output a comprehensive risk index value.

[0031] As a preferred embodiment of the railway maintenance site operation protection and control method described in this invention, the steps of spatially diffusing the comprehensive risk index value to generate a dynamic risk field, and then performing field strength gradient vectorization and path integration on the dynamic risk field to generate a safe path coordinate sequence are as follows:

[0032] The comprehensive risk index value is subjected to nonlinear anisotropic diffusion to generate a dynamic risk field.

[0033] Gradient field analysis and geodesic manifold integration are performed on the dynamic risk field to generate a safe path coordinate sequence.

[0034] As a preferred embodiment of the railway maintenance site operation protection and control method of the present invention, the steps of performing spatiotemporal trajectory matching on the safety path coordinate sequence, obtaining the path execution deviation, performing spatiotemporal causal analysis on the path execution deviation, and constructing an event causal graph are as follows.

[0035] Dynamic time warping and phase synchronization are performed on the safe path coordinate sequence to generate a trajectory alignment mapping relationship;

[0036] Multimodal deviation measurement is performed on the trajectory alignment mapping relationship to extract the path execution deviation.

[0037] Spatiotemporal causal analysis and causal path verification are performed on the path execution deviation, and an event causal graph is constructed.

[0038] As a preferred embodiment of the railway maintenance site operation protection and control method of the present invention, the steps of obtaining the violation contribution degree of the event causal graph through reverse fact query, compiling the basic permission rule base into a strategy based on the violation contribution degree, and generating incremental control strategies and visual decision reports are as follows.

[0039] The intervention effect of the event causal graph is evaluated by using the Do-operator counterfactual query to obtain the contribution of violations.

[0040] Based on the contribution of violations, the basic permission rule base is used to perform policy logic reasoning and parameter instantiation to generate incremental control policies;

[0041] The decision data is encapsulated and rendered in a multimodal manner based on the contribution of violations and the cause-and-effect graph of events, generating a visual decision report.

[0042] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the railway maintenance site operation protection and control method as described in the first aspect of the present invention.

[0043] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the railway maintenance site operation protection and control method as described in the first aspect of the present invention.

[0044] The beneficial effects of this invention are as follows: by generating a dynamic risk field and performing safe path planning, the abstract comprehensive risk index value is transformed into an executable physical space safe path, thereby improving the accuracy of path planning; by simultaneously constructing an event cause-effect graph and obtaining the contribution of violations, the root cause of risks is accurately located, and the dynamic adaptability of rules and the interpretability of decisions are enhanced. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 A flowchart for the protection and control methods at railway maintenance sites.

[0047] Figure 2 A flowchart for generating comprehensive risk index values.

[0048] Figure 3 A flowchart generated for a safe path.

[0049] Figure 4 A flowchart generated for the incremental control strategy. Detailed Implementation

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0052] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0053] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for protective control at railway maintenance sites, comprising the following steps:

[0054] S1. Collect historical accident data, job responsibility attributes and multi-source real-time working condition data and preprocess them, extract the permission mapping features of the preprocessed job responsibility attributes and the working condition risk features of the preprocessed multi-source real-time working condition data.

[0055] S1.1 Multi-source real-time operating data includes job operation data, train operation data, and environmental parameters;

[0056] It should be noted that the operational data is captured in real time by the IMU (Inertial Measurement Unit), camera, and UWB positioning built into the AR glasses worn by the workers; the train operation data is directly obtained through the real-time data interface of the transponders, radar speed measuring devices, and train control terminals deployed along the track; and the environmental parameters are continuously collected by independent meteorological sensors such as temperature and humidity sensors, wind speed sensors, and visibility sensors deployed at the work site, as well as track surface monitoring probes.

[0057] It should also be noted that historical accident data is extracted from historical accident reports; job responsibility attributes are obtained by accessing the API interface of the company's existing human resources management terminal.

[0058] S1.2 Preprocessing includes data cleaning, format standardization, and noise filtering;

[0059] It should be noted that the collected real-time operating data, train operation data, environmental parameters, historical accident data, and job responsibility attributes from multiple sources are cleaned. Abnormal values ​​of sensors (such as sudden changes in the positioning coordinates of AR glasses) are identified and removed using the box plot method. Linear interpolation is used to fill in the data packet loss caused by API interface transmission. All timestamps, speed units, and spatial coordinates are unified and noise filtered. Kalman filtering and wavelet transform are used to filter out power frequency interference and smooth jitter.

[0060] S1.3 Perform permission topology encoding on the preprocessed job responsibility attributes to generate permission mapping features;

[0061] It should be noted that the job ID, role type, permission level, and equipment operation list, etc., of the preprocessed job responsibility attributes are extracted. The role type is used as a graph node, and the numerical comparison relationship between permission levels (higher permission level points to lower permission level) is used as a directed edge to construct a permission hierarchy directed graph. A topological sorting algorithm is performed on the permission hierarchy directed graph, and the graph nodes with an in-degree of zero are removed iteratively to generate an inheritance relationship sequence. The equipment operation list is mapped to a binary permission bitmask, and feature fusion is performed with the inheritance relationship sequence to generate a permission mapping feature vector.

[0062] It should also be noted that "in-degree" refers to the number of edges pointing to graph nodes.

[0063] S1.4 Analyze the behavioral patterns of the preprocessed job operation data and generate quantitative operational features;

[0064] It should be noted that the preprocessed job operation data is matched with the preset standard operation template trajectory using trajectory morphology to generate a trajectory consistency metric coefficient. Vibration signal modal resonance analysis is performed on the trajectory consistency metric coefficient to generate abnormal vibration resonance features. The three-axis acceleration (including uniform speed, acceleration, and jitter) of the abnormal vibration resonance features is kinematically characterized to generate a motion stability index. The trajectory consistency metric coefficient, abnormal vibration resonance features, and kinematic state characterization are fused to form operation quantitative features.

[0065] It should also be noted that the trajectory dataset is extracted from historical standard operation records. K operation trajectories (the value of k is determined according to the statistical distribution of the trajectory dataset) are randomly selected from the trajectory dataset as the original center points. The dynamic time warping distance from all operation trajectories to the original center points is calculated. Each operation trajectory is assigned to the category of the original center point with the closest dynamic time warping distance, forming k initial clusters. Each initial cluster is traversed, and the operation trajectories that are not the center points of the initial clusters are taken as the new center points. The total dynamic time warping distance from all operation trajectories after replacement to the new center points is obtained. If the total dynamic time warping distance is smaller than that of the original center point, the current operation trajectory is taken as the original center point. This process continues until the total dynamic time warping distance no longer changes, resulting in stable clusters. The standard operation template trajectory is generated by obtaining the average coordinates of the operation trajectories within the stable clusters.

[0066] The expression for calculating the dynamic time-warped distance is:

[0067] ;

[0068] in, Indicates the operation trajectory and operation trajectory Dynamic time-warped distance between , Represents a trajectory dataset. Indicates the operation trajectory The A spatial coordinate point, Indicates the operation trajectory The One spatial coordinate point;

[0069] The expression for calculating Euclidean distance is:

[0070] ;

[0071] in, Indicates the operation trajectory The Spatial coordinates and operation trajectory The Euclidean distance between spatial coordinate points , Indicates the operation trajectory The The first spatial coordinate point 3D coordinate values, Indicates the operation trajectory The The first spatial coordinate point 3D coordinate values;

[0072] S1.5. Vectorize the preprocessed train operation data to obtain operation status characteristics;

[0073] It should be noted that the phase space reconstruction method is used to reconstruct the dynamic characteristics of the preprocessed train operation data (i.e., the coupling relationship between speed, acceleration, and line resistance) to generate a phase space trajectory. An initial state point is randomly selected on the phase space trajectory, and a disturbance point is randomly generated in the nearest neighbor (e.g., one millionth of a distance) of the initial state point. The motion trajectories of the initial state point and the disturbance point are synchronously tracked at a fixed time step, and the Euclidean distance between the initial state point and the disturbance point at each time step is recorded to form a distance sequence. The distance sequence is linearly transformed to obtain the Lyapunov exponent. The frequency of the zero-crossing point of the second derivative of the phase space trajectory is analyzed to generate the braking emergency stop risk characteristics. The root mean square deviation between the real-time speed and the planned speed of the phase space trajectory is simultaneously calculated to obtain the scheduling deviation. The Lyapunov exponent, braking emergency stop risk characteristics, and scheduling deviation are structurally integrated to generate the operation status characteristics.

[0074] It should also be noted that the planned speed was obtained by analyzing the dynamic relationships of relevant route topology data, such as gradient, curvature, and speed limit point information.

[0075] S1.6 Perform time-frequency domain transformation on the preprocessed environmental parameters to generate environmental disturbance characteristics;

[0076] It should be noted that continuous wavelet transform analysis is used to analyze the energy distribution characteristics of preprocessed environmental parameters (including wind speed, temperature, humidity, and vibration signals) at different time scales, and the wavelet energy spectrum of a specific frequency band (such as 0.1-10Hz) is extracted as a frequency domain perturbation index. At the same time, Hilbert-Huang transform is performed on the preprocessed environmental parameters to obtain instantaneous frequency and amplitude, generating time-domain non-stationary fluctuation characteristics. The frequency domain perturbation index and the time-domain non-stationary fluctuation characteristics are fused to form environmental perturbation characteristics.

[0077] S1.7 Integrate operational quantitative characteristics, operating status characteristics, and environmental disturbance characteristics to generate operating condition risk characteristics;

[0078] It should be noted that the operational quantification features, operating status features, and environmental disturbance features are fused across modes to generate a high-dimensional feature matrix. Principal component analysis is then used to reduce the dimensionality of the merged high-dimensional feature matrix and extract the principal component with the highest variance contribution rate. The maximum correlation minimum redundancy feature selection algorithm is used to remove redundant information from the principal component and enhance the discriminative power between features. Finally, operating condition risk features are generated.

[0079] S2. Perform rule matching and instantiation on the preprocessed historical accident data and permission mapping features to generate a basic permission rule base;

[0080] S2.1 Perform statistical analysis and rule mining on the preprocessed historical accident data to generate a pre-set rule template library;

[0081] It should be noted that the preprocessed historical accident data is classified and statistically analyzed according to accident type (such as derailment, collision, and equipment failure). High-frequency associated operation sequences (such as continuous braking and overspeeding through curves), environmental condition combinations (such as rain and steep gradients), and equipment status characteristics (such as braking alarm frequency) are extracted for each type of accident. Causal rules (such as "when operation X and environment Y occur simultaneously, accident Z is likely to occur") are integrated. The co-occurrence patterns in the preprocessed historical accident data are analyzed using association rule mining algorithms. The high-frequency associated operation sequences, causal rules, and co-occurrence patterns are structured and integrated to form a pre-set rule template library.

[0082] It should also be noted that the association rule mining algorithm is used to analyze the co-occurrence patterns in the preprocessed historical accident data. Specifically, the number of occurrences of each accident feature item (such as "accident type = derailment") is counted, and the ratio of the number of occurrences to the total number of records in the preprocessed historical accident data is used as the support. Accident feature items with a support not lower than a preset support threshold are selected as frequent items, and a head table is generated in descending order of support. Each frequent item is sorted according to the order of the head table and inserted into an FP-tree structure (paths sharing the same prefix are merged). The conditional pattern base of each frequent item in the head table (containing all prefixes of the frequent item, such as "braking distance extended") is extracted, a conditional FP-tree is constructed, and frequent itemsets (such as "severe gradient + wet and slippery track → braking distance extended") are recursively mined to generate co-occurrence patterns in the preprocessed historical accident data.

[0083] The support threshold is set based on the distribution of preprocessed historical accident data. The range of the support threshold is (0.05, 0.15). The lower limit of 0.05 can ensure that the accident feature appears frequently enough in the preprocessed historical accident data (e.g., at least 500 times in 10,000 data points) to avoid random patterns. The upper limit of 0.15 can prevent the support threshold from being too high, which may lead to the missed detection of important but not very high frequency associations (e.g., the combination of a specific environment and equipment failure may have a moderate frequency but extremely high risk).

[0084] S2.2 Utilize permission mapping features to perform logical condition matching on the pre-set rule template library and generate a matching rule set;

[0085] It should be noted that for each rule in the pre-set rule template library, bidirectional matching is performed through permission mapping features. Specifically, the operation permission requirements (such as "requires advanced braking operation permissions") and environmental permission restrictions (such as "only applicable to non-rainy day permission areas") in the rule triggering conditions are compared one by one with the current permission mapping features. A subset of rules whose permission mapping features can completely cover the operation permission requirements of the rule triggering conditions are selected and integrated to generate a matching rule set.

[0086] S2.3. Compile the rule logic and instantiate the parameters of the matching rule set to generate the basic permission rule library;

[0087] It should be noted that each rule in the matching rule set is structured, and the logical conditions described in natural language (such as "when speed > 80km / h and slope > 5‰") are compiled into Boolean expressions (such as "speed parameter > 80 and slope parameter > 5"), and the abstract parameters in the rules (such as "high risk level" and "default speed limit") are instantiated into specific values ​​(such as "risk level = 3" and "speed limit = 60km / h"), forming a basic permission rule base containing explicit logical judgment statements and specific parameter values.

[0088] S3. Input the operating condition risk characteristics into the risk prediction model, the risk coding layer performs feature mapping and risk association construction, the risk evolution layer performs global optimization search, and outputs the comprehensive risk index value.

[0089] S3.1. Use a deep neural network to build a risk encoding layer and use a random forest algorithm to build a risk evolution layer;

[0090] It should be noted that in the PyTorch framework, a risk encoding layer is constructed using a temporal convolutional network. Specifically, multi-scale convolutional kernels (such as those with sizes of 3 and 5) are used to capture short-term fluctuations and long-term trends. The expansion factor is set to capture long-range dependencies with fewer layers (such as setting it to a power of 2 to exponentially expand the receptive field to cover the entire risk event cycle). The GELU activation function is used to enhance the nonlinear representation capability. The feature dimension is compressed through a global max pooling layer to extract risk patterns, and gradient stability is maintained through residual connections, thus completing the construction of the risk encoding layer.

[0091] A risk evolution layer is constructed using the random forest algorithm. By integrating multiple decision trees (while controlling complexity to ensure prediction accuracy, such as setting the number of decision trees to 100), the prediction stability and generalization ability are improved. The maximum depth of the decision trees should be able to effectively capture the nonlinear relationship of feature interactions, such as setting the maximum depth to 10. The Gini coefficient splitting criterion is adopted. When splitting the decision tree, a nonlinear transformation is performed on each split point to generate the Gini coefficient. The split point with the largest reduction in Gini coefficient is selected to optimize the feature selection efficiency (the Gini coefficient ranges from 0 to 1, and the larger the Gini coefficient, the lower the purity). The average of the total reduction in Gini coefficient of all split points is used as the feature importance weight. The splitting strategy is dynamically adjusted using the feature importance weight to complete the construction of the risk evolution layer.

[0092] S3.2. A risk prediction model is constructed by fusing and associating features of the risk coding layer and the risk evolution layer through an attention mechanism.

[0093] It should be noted that the risk encoding layer and the risk evolution layer are dimensionally aligned and feature-stitched to obtain fused features containing encoded features and evolutionary information. The fused features are weighted by importance using a multi-head attention mechanism. A Q matrix (query), a K matrix (key), and a V matrix (value) are generated through independent linear layers. The dot product of Q and K is used as the correlation degree, and the correlation degree is normalized by Softmax to generate an attention weight matrix. The V matrix is ​​weighted and summed using the attention weight matrix to generate context-aware fused features. The context-aware fused features are input into a fully connected layer for feature compression and nonlinear mapping to reduce the dimensionality to the risk value range (e.g., [0,1]), thus completing the construction of the risk prediction model.

[0094] Historical operational risk features are used as training data and divided into training, validation, and test sets. A data loader is used to diversify the features of the sample sets, and Z-score normalization is applied for standardization to create enhanced samples. The Adam optimizer is used to perform gradient backpropagation on the enhanced samples, and a learning rate scheduler is simultaneously applied to dynamically adjust the learning rate to obtain the training loss. On the validation set, K-fold cross-validation is performed on the training loss to obtain the validation set error rate. Training terminates when the validation set error rate no longer decreases after several consecutive rounds (e.g., 10 times), and the trained risk prediction model is output synchronously.

[0095] S3.3 Input the working condition risk characteristics into the risk prediction model. The risk coding layer uses multi-scale convolution kernels to perform feature mapping and risk association parsing to generate high-dimensional feature vectors and risk weight matrices.

[0096] It should be noted that the operating condition risk characteristics are input into the risk prediction model. The short-term fluctuations and long-term trends of the operating condition risk characteristics are captured in parallel by multi-scale convolutional kernels. The receptive field is expanded exponentially by the expansion factor to cover the complete risk event cycle. At the same time, GELU activation is used to enhance the nonlinear feature representation. The output is a pooled feature with compressed dimensions after global max pooling layer. The pooled feature is then residually connected to maintain gradient stability and generate enhanced features. The enhanced features are linearly transformed and mapped to high-dimensional feature vectors. At the same time, the relevant weights of each channel of the convolutional layer are Softmax normalized through the channel attention mechanism to generate a risk weight matrix that reflects the importance of different feature dimensions.

[0097] S3.4 The risk evolution layer uses a gradient boosting decision tree to perform a global optimization search on the high-dimensional feature vectors to form a comprehensive risk score;

[0098] It should be noted that the high-dimensional feature vector output from the risk encoding layer is input into the risk evolution layer. A global optimization search framework is constructed based on the gradient boosting decision tree algorithm. In specific operations, GBDT is selected to capture nonlinear interaction relationships. The number and depth of decision trees are set according to the feature dimension of the high-dimensional feature vector. The number of decision trees must be sufficient to cover all feature combination patterns of the high-dimensional feature vector to avoid underfitting, and the depth of the decision trees must prevent memory noise. The average risk score of historical accident data is used as the initial prediction value. Multiple decision trees are iteratively trained to predict risk. The first decision tree fits the residual between the high-dimensional feature vector and the initial prediction value. Each decision tree focuses on the gradient direction of the residuals of all previous decision trees (i.e., the trend of residual change). By selecting the split node that contributes the most to the reduction of residuals, the risk prediction results are gradually corrected. The risk prediction results of all decision trees are weighted and summed to generate a comprehensive risk score.

[0099] It should also be noted that the average risk score and initial prediction value are the statistical average of the risk labels extracted from the preprocessed historical accident data (the risk labels are generated by the labeling operation in the preprocessing stage of historical accident data, such as the severity level of historical accidents, binary labels of whether a safety event has been triggered, etc.).

[0100] S3.5. Perform tensor product fusion and attention-weighted projection on the high-dimensional feature vector, risk weight matrix and comprehensive risk score to output the comprehensive risk index value.

[0101] It should be noted that the comprehensive risk score is converted to the same dimension as the high-dimensional feature vector through linear transformation; the high-dimensional feature vector, the risk weight matrix, and the converted risk score vector are multiplied element by element and superimposed to form a high-order fusion vector that retains the interaction of multi-source information. An attention mechanism is applied to the high-order fusion vector to obtain the correlation degree of each component within the high-order fusion vector, generating an attention weight matrix. The high-order fusion vector is then weighted and fused according to the attention weight matrix and normalized to output the comprehensive risk index value.

[0102] S4. Spatial diffusion of the comprehensive risk index value to generate a dynamic risk field, and field strength gradient vectorization and path integration of the dynamic risk field to generate a safe path coordinate sequence.

[0103] S4.1. Perform nonlinear anisotropic diffusion on the comprehensive risk index value to generate a dynamic risk field;

[0104] It should be noted that the comprehensive risk index value is spatially gridded to generate a discrete spatial domain, where each risk grid point in the discrete spatial domain corresponds to a comprehensive risk index value. Based on the definition of the discrete spatial domain, a nonlinear diffusion coefficient (positively correlated with the comprehensive risk index value of the risk grid point, i.e., the diffusion rate is faster in high-risk areas) and anisotropic diffusion tensor (e.g., the diffusion rate along the equipment running direction is higher than that in the vertical direction, reflecting the directional dependence of risk propagation) is used to perform nonlinear anisotropic diffusion. This allows the comprehensive risk index value to propagate in the discrete spatial domain according to the nonlinear law of "accelerated diffusion in high-risk areas and slow penetration in low-risk areas". During the diffusion process, the comprehensive risk index value of each grid point is dynamically updated, generating a dynamic risk field that evolves with time and space.

[0105] S4.2 Perform gradient field analysis and geodesic manifold integration on the dynamic risk field to generate a safe path coordinate sequence;

[0106] It should be noted that by obtaining the gradient vector field of the dynamic risk field, the dynamic risk field is analyzed using gradient field analysis. The gradient direction of the gradient vector field points to the direction of the fastest increase in the comprehensive risk index value, and the gradient magnitude of the gradient vector field reflects the rate of change of the comprehensive risk index value. By statistically analyzing the high-risk gradient band (the region where the comprehensive risk index value changes rapidly from low to high) and the low-gradient safe region (the region where the comprehensive risk index value changes slowly) in the dynamic risk field, the dynamic risk field is regarded as a continuous geometric space. A path energy function (the path integral includes the sum of risk values) is constructed with the goal of "minimizing the comprehensive risk index value". The gradient mutation penalty term is used to divide the continuous space covered by the dynamic risk field into a uniform grid (e.g., a 1m×1m grid) using a numerical integration method, generating a grid node dataset. With the goal of minimizing the path energy function, the path point coordinates of the grid node dataset are iteratively optimized to generate an optimized path point sequence. The optimized path point sequence is then subjected to spline interpolation and discretization to generate a safe path for the path energy function. The safe path will preferentially extend along the negative gradient direction while avoiding high gradient magnitude risk zones. The integrated safe path is discretized into a series of spatial coordinate points to generate a safe path coordinate sequence from the start point to the end point of the operation.

[0107] S5. Perform spatiotemporal trajectory matching on the safe path coordinate sequence, obtain the path execution deviation, perform spatiotemporal causal analysis on the path execution deviation, and construct an event causal graph;

[0108] S5.1 Perform dynamic time warping and phase synchronization on the safe path coordinate sequence to generate trajectory alignment mapping relationship;

[0109] It should be noted that the timestamps of the safety path coordinate sequences are standardized to unify the time scale and form a standard trajectory. A dynamic time warping algorithm is applied to find the optimal nonlinear alignment path by obtaining the Euclidean distance between different safety path coordinate sequences. This optimal nonlinear alignment path achieves length matching between two safety path coordinate sequences on the time axis through stretching or compression operations, while preserving the spatial correspondence of coordinate points within the safety path coordinate sequences (e.g., aligning the "acceleration phase" of safety path coordinate sequence A with the "uniform speed phase" of safety path coordinate sequence B). Based on the standard trajectory, phase nodes (including path start points, turning points, and deceleration points) are identified. The phase nodes of each standard trajectory are adjusted through interpolation to ensure that different standard trajectories have a spatial coordinate correspondence at the same phase node, generating a trajectory alignment mapping relationship (e.g., "the first point of the safety path coordinate sequence corresponds to the second point of the standard trajectory").

[0110] S5.2 Perform multimodal deviation measurement on the trajectory alignment mapping relationship and extract the path execution deviation.

[0111] It should be noted that, based on the trajectory alignment mapping relationship, the corresponding point coordinate pairs are extracted, and the spatial position deviation is generated by obtaining the Euclidean distance between each pair of point coordinate pairs; the time synchronization deviation is generated by comparing the timestamp difference between the original point (coordinate point in the safe path coordinate sequence) and the corresponding standard point (coordinate point in the standard trajectory); the direction change rate of the point coordinate pairs is obtained to generate the path shape deviation; the spatial position deviation, time synchronization deviation and path shape deviation are weighted and summed to integrate into the path execution deviation degree.

[0112] S5.3 Perform spatiotemporal causal analysis and causal path verification on the path execution deviation, and construct an event causal graph;

[0113] It should be noted that the partial path execution deviation is mapped to a unified spatiotemporal grid by timestamp and spatial coordinates to form a three-dimensional data matrix of "time-space-deviation". At the same time, the external event labels of the corresponding grid points are associated (such as whether equipment failure occurs in a certain spatial region at a certain time point) to generate spatiotemporal correlation data. The temporal autocorrelation of the path execution deviation (the source of the persistence and periodicity of the path execution deviation) is obtained through a sliding time window of fixed time length. Simultaneously, the spatial dependence of the path execution deviation is obtained through spatial neighborhood aggregation (such as whether the deviation in a certain area is affected by environmental disturbances in the neighboring area) to generate potential causal relationship candidates (such as "equipment failure event → subsequent increase in deviation"). The spatiotemporal correlation data and potential causal relationship candidates are integrated to generate an event causal graph. The causal nodes of the event causal graph include "external events" (equipment failure and environmental disturbances, etc.), "spatiotemporal factors" (time window and spatial region), and "path execution deviation". Causal edges represent potential causal relationship candidates.

[0114] S6. Obtain the violation contribution degree of the event cause-effect graph through reverse fact query, compile the basic permission rule base into a strategy based on the violation contribution degree, and generate incremental control strategies and visual decision reports.

[0115] S6.1. Evaluate the intervention effect of the event causal graph through Do-operator counterfactual query to obtain the contribution of violations;

[0116] It should be noted that, based on the causal nodes in the event causal graph, counterfactual conditions (such as "if the equipment failure had not occurred") are set, and the state of the causal node is fixed to the counterfactual condition value (such as "equipment failure = not occurred"). Starting from the causal node, the simulated value of the path execution deviation under the counterfactual conditions is obtained layer by layer along the causal path. The difference between the simulated value of the path execution deviation after intervention and the actual path execution deviation before intervention is taken as the intervention effect quantity. The intervention effect quantity is then normalized to generate the violation contribution.

[0117] S6.2. Based on the contribution of violations, perform policy logic reasoning and parameter instantiation on the basic permission rule base to generate incremental control policies;

[0118] It should be noted that a three-tiered threshold system is used to hierarchically classify and instantiate parameters for violation contribution. The first-tier threshold range is set to [0.8, 1.0], the second-tier threshold range is (0.5, 0.8), and the third-tier threshold range is [0, 0.5). When the violation contribution score is within the first-tier threshold range, it is identified as a high-risk violation scenario, triggering the "emergency access control + full-link traceability" strategy logic and marking it as a red alert-level strategy group. When the violation contribution score is within the second-tier threshold range, it is identified as a medium-risk violation scenario, triggering the "flexible access restriction + targeted risk interception" strategy logic and marking it as a yellow alert-level strategy group. When the violation contribution score is within the third-tier threshold range, it is identified as a low-risk violation scenario, triggering the "flexible access warning + compliance guidance" strategy logic and marking it as a green prompt-level strategy group. The strategy logic for different threshold ranges is injected into the basic access rule base to generate incremental control strategies.

[0119] It should also be noted that the three thresholds are defined based on the quantitative score range of the violation contribution. The first threshold [0.8, 1.0] corresponds to violation scenarios that "cause business interruption, major compliance penalties, and irreversible loss of data and permissions"; the second threshold (0.5, 0.8) corresponds to scenarios that "affect business availability but do not trigger explicit prohibitions and involve moderate abuse of data and permissions"; and the third threshold [0, 0.5) corresponds to behaviors that "only trigger basic operation procedure warnings and have a very minor impact on business security and compliance". The basis for using 0.8 as the lower limit of the first threshold and the upper limit of the second threshold is that when the violation contribution reaches or exceeds 0.8, the violation will cause high-harm consequences, such as business interruption (maintenance work stoppage exceeding 2 hours), major compliance penalties (such as regulatory fines), or irreversible loss of core data and permissions. The basis for using 0.5 as the lower limit of the second threshold and the upper limit of the third threshold is that when the violation contribution is below 0.5, the violation only triggers low-harm consequences, such as basic operation warnings (single boundary violation warnings), and has no substantial impact on business availability and compliance (operational efficiency decrease < 5%).

[0120] S6.3. Encapsulate decision data and perform multimodal visualization rendering on the contribution of violations and the cause-and-effect graph of events to generate a visual decision report.

[0121] It should be noted that the decision data is encapsulated by the violation contribution and the event causal graph to construct a "contribution-causal graph association dataset". "Contribution value" and "level label" attributes are added to the causal nodes of the causal graph (e.g., the violation contribution of the "equipment failure" node is 0.7, and the risk level is "high"). A "total associated contribution" is added to each causal edge (e.g., the total violation contribution of the causal edge "equipment failure → increased deviation" is 0.7). The contribution-causal graph association dataset is then subjected to multimodal visualization rendering. Centered on the event causal graph, the size of the causal nodes is mapped to the violation contribution (e.g., a causal node with a violation contribution of 0.8 has a diameter of 15px). The colors of the causal nodes are filled according to the corresponding colors in the incremental control strategy, and a visual decision report is generated.

[0122] This embodiment also provides a computer device applicable to the railway maintenance site operation protection and control method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the railway maintenance site operation protection and control method proposed in the above embodiment.

[0123] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0124] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for implementing on-site protection and control of railway maintenance operations as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0125] In summary, this invention improves the accuracy of path planning by generating a dynamic risk field and planning safe paths, transforming abstract comprehensive risk index values ​​into executable physical space safe paths; and by simultaneously constructing an event cause-effect graph and obtaining the contribution of violations, it achieves precise location of risk root causes and enhances the dynamic adaptability of rules and the interpretability of decisions.

[0126] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for protective control at railway maintenance sites, characterized in that: include, Collect historical accident data, job responsibility attributes, and multi-source real-time working condition data and preprocess them. Extract the permission mapping features of the preprocessed job responsibility attributes and the working condition risk features of the preprocessed multi-source real-time working condition data. The preprocessed historical accident data and permission mapping features are matched and instantiated to generate a basic permission rule base. The risk characteristics of the working conditions are input into the risk prediction model. The risk coding layer performs feature mapping and risk association construction, the risk evolution layer performs global optimization search, and outputs a comprehensive risk index value. Spatial diffusion is performed on the comprehensive risk index value to generate a dynamic risk field. The dynamic risk field is then vectorized by field strength gradient and path integral to generate a safe path coordinate sequence. Spatiotemporal trajectory matching is performed on the safe path coordinate sequence to obtain the path execution deviation. Spatiotemporal causal analysis is then performed on the path execution deviation to construct an event causal graph. By obtaining the violation contribution degree of the event cause-effect graph through reverse fact query, the basic permission rule base is compiled into a strategy based on the violation contribution degree, and incremental control strategies and visual decision reports are generated.

2. The railway maintenance site operation protection and control method as described in claim 1, characterized in that: The multi-source real-time operating data includes job operation data, train operation data, and environmental parameters; The preprocessing includes data cleaning, format standardization, and noise filtering.

3. The railway maintenance site operation protection and control method as described in claim 2, characterized in that: The specific steps for extracting the permission mapping features of the preprocessed job responsibility attributes and the work condition risk features of the preprocessed multi-source real-time work condition data are as follows. Perform permission topology encoding on the preprocessed job responsibility attributes to generate permission mapping features; Analyze the behavioral patterns of preprocessed job operation data to generate quantitative operational features; The preprocessed train operation data is vectorized to obtain operation status characteristics; The preprocessed environmental parameters are transformed in the time-frequency domain to generate environmental disturbance characteristics. By integrating operational quantitative characteristics, operational status characteristics, and environmental disturbance characteristics, operating condition risk characteristics are generated.

4. The railway maintenance site operation protection and control method as described in claim 3, characterized in that: The steps for performing rule matching and instantiation on the preprocessed historical accident data and permission mapping features to generate a basic permission rule base are as follows: Perform statistical analysis and rule mining on the preprocessed historical accident data to generate a pre-set rule template library; By leveraging permission mapping features, logical conditions are matched against a pre-defined rule template library to generate a set of matching rules. The matching rule set is compiled into rules and parameters are instantiated to generate a basic permission rule library.

5. The railway maintenance site operation protection and control method as described in claim 4, characterized in that: The steps involve inputting the operating condition risk characteristics into the risk prediction model, constructing feature mapping and risk associations in the risk encoding layer, performing a global optimization search in the risk evolution layer, and outputting a comprehensive risk index value. The specific steps are as follows: A risk encoding layer is built using a deep neural network, and a risk evolution layer is built using a random forest algorithm. A risk prediction model is constructed by fusing and associating features between the risk encoding layer and the risk evolution layer through an attention mechanism. The working condition risk characteristics are input into the risk prediction model. The risk encoding layer uses multi-scale convolutional kernels to perform feature mapping and risk association parsing, generating high-dimensional feature vectors and risk weight matrices. The risk evolution layer uses a gradient boosting decision tree to perform a global optimization search on high-dimensional feature vectors to form a comprehensive risk score; The high-dimensional feature vector, risk weight matrix, and comprehensive risk score are fused by tensor product and then weighted by attention to output a comprehensive risk index value.

6. The railway maintenance site operation protection and control method as described in claim 5, characterized in that: The process involves spatial diffusion of the comprehensive risk index value to generate a dynamic risk field, followed by field strength gradient vectorization and path integration of the dynamic risk field to generate a safe path coordinate sequence. The specific steps are as follows: The comprehensive risk index value is subjected to nonlinear anisotropic diffusion to generate a dynamic risk field. Gradient field analysis and geodesic manifold integration are performed on the dynamic risk field to generate a safe path coordinate sequence.

7. The railway maintenance site operation protection and control method as described in claim 6, characterized in that: The process involves performing spatiotemporal trajectory matching on the safe path coordinate sequence to obtain the path execution deviation, conducting spatiotemporal causal analysis on the path execution deviation, and constructing an event causal graph. The specific steps are as follows: Dynamic time warping and phase synchronization are performed on the safe path coordinate sequence to generate a trajectory alignment mapping relationship; Multimodal deviation measurement is performed on the trajectory alignment mapping relationship to extract the path execution deviation. Spatiotemporal causal analysis and causal path verification are performed on the path execution deviation, and an event causal graph is constructed.

8. The railway maintenance site operation protection and control method as described in claim 7, characterized in that: The steps involve obtaining the violation contribution degree of the event causal graph through reverse fact query, compiling the basic permission rule base into a strategy based on the violation contribution degree, and generating incremental control strategies and visual decision reports. The specific steps are as follows: The intervention effect of the event causal graph is evaluated by using the Do-operator counterfactual query to obtain the contribution of violations. Based on the contribution of violations, the basic permission rule base is used to perform policy logic reasoning and parameter instantiation to generate incremental control policies; The decision data is encapsulated and rendered in a multimodal manner based on the contribution of violations and the cause-and-effect graph of events, generating a visual decision report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the railway maintenance site operation protection and control method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the railway maintenance site operation protection and control method according to any one of claims 1 to 8.

Images