Freight security monitoring method and system using tarp status image recognition

Abstract

The application discloses a tarpaulin state image recognition freight security monitoring method and system, and relates to the technical field of freight security monitoring.The method comprises the following steps: continuously collecting multispectral image data of the surface of a tarpaulin; generating a topological guide map based on skeleton data and inputting the pre-constructed visual Transformer model to extract tarpaulin wrinkle feature data and rope knot structure feature data; calculating local wind pressure data of the tarpaulin wrinkle area according to the displacement mapping of the wrinkle features and the multispectral image; combining real-time tension data to perform online compensation on the local wind pressure data, obtaining compensated tension distribution data, and generating a tarpaulin tension thermal map; generating a tarpaulin critical wind speed value based on the above feature data, and outputting a warning when the value is lower than a preset threshold.The method fuses structural features and mechanical state data, realizes accurate identification and safety warning of the tarpaulin state, and improves the safety and reliability of the railway freight process.

Description

Technical Field

[0001] This invention relates to the field of freight safety monitoring technology, specifically to a freight safety monitoring method and system based on tarpaulin status image recognition. Background Technology

[0002] The role of tarpaulins in securing and protecting goods during railway freight transport is crucial for ensuring cargo safety, and their reliability directly affects the protective effect on goods during transit. However, tarpaulins are often affected by various factors during transportation, such as wind, rope tension, and weather, which can cause wrinkles, loosening, or even damage to the tarpaulins, leading to safety hazards.

[0003] Existing image-based tarpaulin condition detection technologies mostly focus on identifying obvious damaged areas or judging the integrity of tarpaulin coverage. They lack sufficiently precise dynamic analysis capabilities for features such as subtle changes in tarpaulin surface wrinkles and early signs of loosening rope knots. In actual transportation, changes in tarpaulin condition are not only affected by tension but also coupled with multiple factors such as wind speed, wind direction, rope knot tightness, and the distribution pattern of tarpaulin wrinkles. Especially under high-speed rail travel or strong winds, the local wind pressure distribution on the tarpaulin surface changes complexly. Traditional methods relying on single-point or limited tension monitoring cannot comprehensively reflect the overall stability of the tarpaulin, easily leading to prediction lag or misjudgment, thus reducing the practical value of early warning.

[0004] Therefore, there is an urgent need in this field for a freight safety monitoring technology that uses dynamic prediction based on the condition of tarpaulins. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a freight safety monitoring method and system based on tarpaulin status image recognition.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention discloses a freight safety monitoring method based on tarpaulin status image recognition, comprising the following steps:

[0008] Continuously acquire multispectral image data of the tarpaulin surface;

[0009] Based on the multispectral image data, the skeleton data of the tarpaulin outline is extracted, and the skeleton data is used to generate a topology guidance map;

[0010] The topology guidance graph is input into a pre-built visual Transformer model, and the output is tarpaulin fold feature data and rope knot structure feature data.

[0011] Based on the displacement mapping relationship between the tarpaulin fold feature data and the multispectral image data, calculate the local wind pressure data of the tarpaulin fold area;

[0012] Receive real-time tension data collected in real time, and perform online compensation on the local wind pressure data based on the real-time tension data to obtain compensated tension distribution data;

[0013] Based on the tarpaulin fold feature data, the rope knot structure feature data, and the compensation tension distribution data, a critical wind speed value for the predicted tarpaulin is generated.

[0014] When the critical wind speed value is lower than the preset wind speed threshold, warning data is generated and output.

[0015] Secondly, the present invention discloses a freight safety monitoring system based on tarpaulin status image recognition, comprising the following methods:

[0016] The multispectral imaging module is used to continuously acquire multispectral image data of the tarpaulin surface;

[0017] The skeleton extraction module is used to extract the skeleton data of the tarpaulin outline based on the multispectral image data and generate a topology guide map.

[0018] The visual recognition module is used to input the topology guidance map into a pre-built visual Transformer model and output tarpaulin fold feature data and rope knot structure feature data.

[0019] The wind pressure calculation module is used to calculate the local wind pressure data of the tarpaulin fold area based on the displacement mapping relationship between the tarpaulin fold feature data and the multispectral image data.

[0020] The tension compensation module is used to receive real-time tension data and perform online compensation on the local wind pressure data to obtain compensated tension distribution data;

[0021] The critical generation module is used to generate a critical wind speed value for the tarpaulin based on the tarpaulin fold feature data, the rope knot structure feature data, and the compensation tension distribution data.

[0022] The early warning module is used to generate and output early warning data when the critical wind speed value is lower than the preset wind speed threshold.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] 1. This invention effectively enhances the expressive power of tarpaulin surface structural features by continuously acquiring multispectral image data of the tarpaulin surface and combining it with a topological guidance map generated from skeleton data. A pre-constructed visual Transformer model is used to conduct in-depth analysis of the topological guidance map, achieving high-precision identification of tarpaulin folds and rope knots. This breakthrough overcomes the limitations of traditional two-dimensional image recognition, fully utilizing skeleton topological information to enhance spatial correlation and improve the ability to capture complex details of the tarpaulin surface, providing accurate and reliable basic data for subsequent wind pressure calculations and tension analysis.

[0025] 2. Based on the displacement mapping relationship between the tarpaulin fold features and multispectral images, this invention can accurately calculate local wind pressure and its impact on the tarpaulin tension distribution. Through multi-data source fusion and real-time compensation mechanism, it breaks through the single-view limitation of traditional tension monitoring, effectively reflects the spatiotemporal dynamic changes of the tarpaulin under stress in complex environments, enhances the perception of local stress concentration areas, and improves the precision and real-time performance of monitoring.

[0026] 3. This invention achieves more scientific and accurate safety early warning by comprehensively analyzing the characteristics of tarpaulin folds, the structural characteristics of rope knots, and the compensated tension distribution data, as well as predicting critical wind speeds. This prediction mechanism combines multi-dimensional information on structural characteristics and mechanical states, overcoming the limitations of existing technologies that rely on a single indicator in critical wind speed prediction. This enhances the reliability and foresight of tarpaulin condition monitoring, reduces the risk of tarpaulin failure caused by wind, and ensures freight safety. Attached Figure Description

[0027] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0028] Figure 1 This is a flowchart of the steps of the present invention;

[0029] Figure 2 This is a schematic diagram illustrating the working principle of the present invention;

[0030] Figure 3 This is a flowchart of the skeleton extraction and feature fusion process of the present invention;

[0031] Figure 4 This is the critical wind speed early warning logic diagram of the present invention;

[0032] Figure 5 This is a system module connection diagram of the present invention;

[0033] Figure 6 This is a diagram illustrating the module data interaction of the present invention. Detailed Implementation

[0034] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0035] In existing technologies, monitoring the condition of freight tarpaulins often relies on single sensor data or simple image processing methods, making it difficult to simultaneously achieve real-time performance and detection accuracy. Traditional methods cannot effectively reflect the complex deformation process of tarpaulins under wind load and tension, especially under the interactive influence of tarpaulin folds and rope knots, resulting in significant errors in real-time compensation of tension distribution and prediction of critical wind speed values. Existing systems lack coupled analysis of the multi-dimensional characteristics of tarpaulins, making it difficult to achieve accurate freight safety risk assessment and early warning, and failing to meet the high standards of safety monitoring required by modern intelligent logistics.

[0036] To address the aforementioned technical bottlenecks, this application proposes a freight safety monitoring method based on tarpaulin state image recognition, integrating multispectral imaging, visual Transformer recognition, and tension compensation. This method, based on tarpaulin fold features, rope knot structure features, and real-time tension data, establishes a multi-dimensional linkage analysis framework encompassing tarpaulin morphology, stress, and wind load through the coupling of a physical model and a deep learning model. By introducing multi-level algorithms for instantaneous velocity field correction, tension compensation, and critical wind speed prediction, accurate dynamic detection and risk assessment of tarpaulin state are achieved, enhancing the system's robustness and practicality.

[0037] In practice, the system simultaneously acquires multispectral image data and real-time tension sensor data of the tarpaulin. A skeleton extraction module obtains the tarpaulin's outline skeleton and topological guidance map, while a visual Transformer model further extracts the tarpaulin's fold and knot structural features. Based on this multimodal data, physical deformation constraints are applied to the displacement field using finite element or differential methods to accurately correct the instantaneous velocity field. Subsequently, a compensating tension distribution is generated based on the corrected velocity field and real-time tension data. This distribution data, along with the structural stability index, is input into an LSTM model trained on historical wind pressure and failure events to predict the critical wind speed value of the tarpaulin in real time, enabling dynamic determination of safety thresholds and early warning triggering.

[0038] Compared to existing single-data-source monitoring methods, this solution effectively overcomes tension distribution errors caused by tarpaulin wrinkles and knot structures through multimodal data fusion and deep physical model coupling, achieving high-precision real-time monitoring of railway freight tarpaulin conditions. A dynamic critical wind speed prediction model combined with safe tension threshold calculation enhances the system's sensitivity to wind load risks and its response speed. The overall solution balances real-time performance and accuracy, improving the reliability of railway freight safety monitoring and providing strong technical support for accurately preventing tarpaulin failure and ensuring safe train operation.

[0039] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0040] Example 1:

[0041] like Figure 1 As shown, a freight safety monitoring method based on tarpaulin status image recognition includes the following steps:

[0042] Continuously acquire multispectral image data of the tarpaulin surface;

[0043] Skeleton data of the tarpaulin outline is extracted from multispectral image data, and a topology guidance graph is generated using the skeleton data; the topology guidance graph is a directed graph that contains the topological relationships of the skeleton line nodes and connections of the tarpaulin edge.

[0044] Input the topology guidance graph into the pre-built visual Transformer model and output tarpaulin fold feature data and rope knot structure feature data;

[0045] Based on the displacement mapping relationship between the tarpaulin fold feature data and multispectral image data, the local wind pressure data of the tarpaulin fold area is calculated.

[0046] It receives real-time tension data and performs online compensation on local wind pressure data based on the real-time tension data to obtain compensated tension distribution data; the real-time tension data is collected by tension sensors installed on the tarpaulin ropes or fixed points;

[0047] The critical wind speed value for the tarpaulin is generated based on the tarpaulin fold feature data, rope knot structure feature data, and compensation tension distribution data.

[0048] When the critical wind speed value is lower than the preset wind speed threshold, early warning data is generated and output.

[0049] The preset wind speed threshold is pre-configured based on the tensile strength of the tarpaulin material and transportation safety regulations. When the critical safe wind speed (i.e., the critical wind speed value) predicted by the system is lower than the system's preset minimum safe wind speed requirement (i.e., the preset wind speed threshold), it indicates that the tarpaulin structure is becoming unsafe and its wind resistance is insufficient, thus triggering an early warning. This ensures early intervention when the tarpaulin's condition deteriorates (e.g., due to a decrease in structural strength caused by changes in wind force or tension) to prevent tarpaulin tearing or cargo falling.

[0050] like Figure 2 As shown, the working principle of this application is as follows: Multispectral imaging equipment installed at different locations on the vehicle continuously acquires multispectral image data of the surface covering the tarpaulin. This multispectral imaging equipment can simultaneously acquire image information in different bands such as visible light and near-infrared, thus ensuring stable and clear acquisition of tarpaulin surface feature data under different lighting and weather conditions. When the ambient light intensity is below 50 lux or the satellite signal is lost, the system switches to near-infrared active illumination mode and uses inertial navigation to locate the tarpaulin reference point.

[0051] Image processing was performed on the acquired multispectral image data to extract the skeleton data of the tarpaulin outline. The skeleton data, obtained through edge detection and morphological thinning algorithms, reflects the spatial geometric structure of the tarpaulin edges. A topology guidance map was generated based on the skeleton data to characterize the topological features of the tarpaulin surface boundaries, providing structured prior information for subsequent depth feature extraction. This topology guidance map was then input into a pre-built and trained visual Transformer model. This model uses the topology guidance map and the original multispectral image data for feature fusion, and extracts tarpaulin fold feature data and rope knot structure feature data through a multi-head self-attention mechanism. The fold feature data describes the spatial location and morphological parameters of the folds on the tarpaulin surface, while the rope knot structure feature data annotates the position, type, and morphological features of the knots.

[0052] Based on the displacement mapping relationship between tarpaulin fold feature data and multispectral image data, local wind pressure data in the tarpaulin fold area is calculated. This calculation obtains spatial distribution information reflecting the wind stress state of the tarpaulin by comparing the displacement changes of the fold area at adjacent times and deriving the local wind pressure value by combining an aerodynamic model.

[0053] The displacement mapping calibration experiment is as follows: 20 marker points were set on the surface of a 10m×8m tarpaulin. The actual displacement was measured using a total station and compared with the image displacement mapping results. The environmental conditions were wind speed of 8m / s and vehicle speed of 60km / h.

[0054]

[0055] Conclusion: The displacement mapping error is ≤0.04m, which meets the accuracy requirements for wind pressure calculation.

[0056] Based on this, the system receives real-time tension sensor data. Tension sensors are installed on the tarpaulin ropes or anchor points to monitor the stress on the tarpaulin in real time. The system performs online compensation between the real-time tension data and local wind pressure data, generating compensated tension distribution data through a data fusion algorithm to correct the tension distribution calculated solely from wind pressure, making it closer to the actual physical stress state.

[0057] Furthermore, based on tarpaulin fold feature data, rope knot structure feature data, and compensation tension distribution data, the system calculates and predicts the critical wind speed value of the tarpaulin. This critical wind speed value reflects the maximum wind speed that the tarpaulin can withstand under the current stress and structural conditions without structural failure. When the predicted critical wind speed value is lower than a preset wind speed threshold, the system generates warning data and sends it to the driver's terminal and the remote monitoring center. Simultaneously, it executes automated tarpaulin reinforcement or vehicle deceleration commands locally on the vehicle. The preset wind speed threshold is pre-configured based on the tensile strength of the tarpaulin material and transportation safety regulations to ensure that the monitoring strategy complies with industry standards and material safety limits.

[0058] This application enables real-time monitoring and safety assessment of the condition of tarpaulins on freight vehicles. By comprehensively considering the tarpaulin's structural form, wind pressure, and actual tension distribution, it accurately predicts the safety limits of the tarpaulin under different operating conditions, thereby providing early warnings and taking measures to reduce safety risks such as tarpaulin tearing and cargo falling, and improving the reliability and safety of freight transportation.

[0059] like Figure 3 As shown, this application further proposes the following specific steps for extracting skeleton data of the tarpaulin outline based on multispectral image data and generating a topology guidance map using the skeleton data:

[0060] Edge enhancement and binarization processing are performed on multispectral image data, binarized contour lines are extracted using morphological thinning algorithms, and tarpaulin edge skeleton line data are generated through topology preservation constraints.

[0061] The topology data of nodes and connections is established based on the edge skeleton line data of the tarpaulin, and a topology guidance diagram is constructed based on the topology data.

[0062] The topology data consists of a set of nodes. Sum of edges Stored in the form of This represents the i-th skeleton node. Indicates connection to skeleton nodes With skeleton nodes The edge.

[0063] In the implementation of this invention, the acquired multispectral image data is first preprocessed. Specifically, this includes performing edge enhancement operations on the multispectral image to improve the contrast between the tarpaulin boundary and the background, and binarizing the image using adaptive thresholding or global thresholding methods to obtain a binary image that highlights the tarpaulin outline. Subsequently, a morphological thinning algorithm is applied to the binary image to remove redundant pixels and retain the center path of the outline, resulting in tarpaulin edge skeleton line data that maintains the geometric shape.

[0064] After obtaining the skeleton data, the connectivity and branching structure of the skeleton lines are analyzed using a topology-preserving constraint algorithm to identify the relationships between nodes and connections in the skeleton, forming initial topology data. The topology data is stored in the form of node sets and edge sets. The node sets record the positions of key inflection points, endpoints, and forks in the skeleton, while the edge sets record the connection information and geometric lengths between adjacent nodes.

[0065] The generation of the topology-guided graph includes: using nodes from the skeleton data as graph nodes, edges as connections, and node attributes including local curvature values. and mean spectral reflectance Specifically, each node attribute includes a local curvature value. (Used to reflect the degree of boundary curvature at this point) and mean spectral reflectance (Calculated from the multispectral image within the neighborhood of this node).

[0066] The edge weights of the topology guided graph are calculated by weighted summation of the curvature differences and spectral differences between adjacent nodes:

[0067]

[0068] in, The weights of the edges connecting nodes i and j in the skeleton data are given by a and b, which are preset weight coefficients used to allocate the weight ratio between the curvature difference and the spectral reflectance difference, so that the guide graph can reflect both the differences in geometric structure and the differences in the spectral spectrum of the material surface. For example, after grid search optimization, a=0.65, b=0.35. The absolute difference between the local curvature values ​​at skeleton node i and skeleton node j is the value of the difference. The larger the difference, the more drastic the bending change of the tarpaulin surface between the two points. The absolute difference between the average spectral reflectance at skeleton node i and skeleton node j is the result of shadows, stains, wear, or different materials. This difference helps identify non-uniform regions. The resulting topology guidance graph is a weighted directed graph that can be used as additional positional encoding information in the subsequent visual Transformer model input. This helps the model retain the spatial structural association between tarpaulin folds and rope knots during the feature extraction stage.

[0069] This application characterizes the geometric connections of the tarpaulin edge skeleton and integrates local curvature and spectral features to form a topological graph structure data that combines structural and material property differences. In the subsequent visual Transformer feature extraction process, the topological guidance graph, as prior structural information, effectively improves the accuracy of identifying wrinkles and knots, making subsequent wind pressure calculations, tension compensation, and critical wind speed predictions more accurate and reliable, thereby enhancing the responsiveness and robustness of the entire freight safety monitoring system. This application is applicable to railway freight scenarios, covering train operation conditions at speeds of 60-160 km / h, and is compatible with tarpaulin safety monitoring in complex environments such as tunnels and bridges.

[0070] like Figure 3 As shown, this application further proposes that the specific steps for inputting the topology guidance graph into a pre-built visual Transformer model and outputting tarpaulin fold feature data and rope knot structure feature data include:

[0071] The topology guidance graph is converted into topology location encoding vectors by mapping the skeleton node coordinates through a sine function, and the multispectral image data is converted into corresponding image patch feature vectors;

[0072] The topological location encoding vector and the image patch feature vector are concatenated at the input layer of the visual Transformer model;

[0073] In the encoder of the visual Transformer model, the spatial correlation between node features and the spectral correlation between image patch features are calculated and fused through a multi-head self-attention mechanism.

[0074] The decoder of the visual Transformer model outputs tarpaulin fold feature data and rope knot structure feature data.

[0075] Among them, the tarpaulin fold feature data is a set of fold positions and shape parameters of the tarpaulin marked in a spatial coordinate system, and the rope knot structure feature data is a set of knot positions and shape parameters of the tarpaulin marked in the same spatial coordinate system; the spatial coordinate system is a three-dimensional grid coordinate system covering the surface of the tarpaulin.

[0076] After obtaining the weighted topological guidance graph, the present invention further inputs it into a pre-built visual Transformer model to achieve accurate extraction of tarpaulin fold feature data and rope knot structure feature data.

[0077] The pre-built visual Transformer model is as follows:

[0078] (1) Input image size: 512×512 pixels (multispectral 3 channels);

[0079] (2) Image patch size: 16×16 (feature vector dimension = 768);

[0080] (3) Number of encoder layers = 12, number of multi-head attention heads = 16;

[0081] (4) Pre-training dataset: contains 100,000 tarpaulin fold / knot labeled images.

[0082] Specifically, firstly, by using a sine and cosine function mapping method, the two-dimensional or three-dimensional coordinates of the skeleton nodes in the topology guidance graph are converted into topological position encoding vectors. This encoding not only preserves the geometric position information of the nodes but also periodically maps the relative positional relationships of adjacent nodes to enhance the model's ability to capture spatial structural patterns. Simultaneously, multispectral image data is divided into several fixed-size image blocks according to an image segmentation strategy, and these blocks are converted into corresponding image block feature vectors through a convolutional feature extraction network, thereby representing the detailed information of the tarpaulin surface in both spectral and texture dimensions.

[0083] In the input layer of the visual Transformer model, the topological location encoding vector and the corresponding spatial location image patch feature vector are concatenated along the channel dimension, allowing the model to simultaneously receive structural priors and image spectral information before entering the encoding stage. In the encoder part, a multi-head self-attention mechanism is used to calculate the spatial correlation between topological nodes and the spectral correlation between image patches, and then weighted and merged in the attention fusion layer to ensure that the spatial dependencies of local structures such as folds and knots are not lost during global feature extraction.

[0084] In the decoder section, the model generates two types of outputs based on the fused feature representations: one is the tarpaulin fold feature data, which includes the precise position, length, curvature, and morphological parameters of the folds in the three-dimensional spatial coordinate system; the other is the rope knot structure feature data, which includes parameters such as knot position, morphological features, and tightness. The three-dimensional spatial coordinate system adopts a mesh-based modeling method covering the tarpaulin surface to ensure the correspondence between the feature data and the actual physical space, so as to directly call upon it in subsequent wind pressure calculations, tension compensation, and critical wind speed predictions.

[0085] This application utilizes a visual Transformer model that leverages not only the rich texture and spectral information of multispectral images but also introduces structured prior information provided by a topological guidance graph. This allows the feature extraction process to simultaneously focus on global contours and local details. In wrinkle and knot detection, this fused input reduces the impact of illumination variations, image noise, or local occlusion on recognition accuracy, thereby providing higher-confidence input data for subsequent mechanical analysis and early warning decisions, and enhancing the intelligence and stability of the entire freight safety monitoring system.

[0086] This application further proposes that, based on the displacement mapping relationship between tarpaulin fold feature data and multispectral image data, the specific steps for calculating local wind pressure data in the tarpaulin fold area include:

[0087] Extract the pixel coordinate set of the tarpaulin fold area corresponding to adjacent time points, and establish a displacement mapping relationship based on the pixel coordinate set;

[0088] The displacement vector field of the tarpaulin fold area at adjacent time points is calculated using the optical flow method to obtain the velocity components of each pixel in the pixel coordinate set on the two-dimensional plane.

[0089] The acquisition time interval between adjacent image frames based on time-aligned velocity components and multispectral image data Calculate the instantaneous velocity field in the folded region of the tarpaulin ;

[0090] Instantaneous velocity field Substitute into the local wind pressure calculation formula:

[0091]

[0092] in, The value represents the local wind pressure, k is a preset proportionality coefficient based on wind tunnel test calibration, and ρ is the air density, for example, 1.225 kg / m³ (standard atmospheric conditions). The wind pressure correction factor is 0.1-0.3. The standard deviation of the velocity component ( This represents the velocity measurements at all N sampling points within a specific time window. Relative to its average value A measure of the degree of dispersion of wind speed (a measure used to reflect the contribution of wind speed instability to wind pressure).

[0093] Local wind pressure values ​​for all tarpaulin folds A weighted average is performed to generate local wind pressure data that corresponds to the spatial position of the tarpaulin surface and matches the real-time tension data in both time and space.

[0094] In this application, when calculating the local wind pressure data of the tarpaulin fold region based on the displacement mapping relationship between the tarpaulin fold feature data and multispectral image data, the corresponding fold region is first located in consecutive multispectral image frames. The pixel coordinate set of this region in adjacent frames is extracted, and a displacement mapping relationship matrix is ​​constructed to record the two-dimensional coordinate changes of each pixel. Subsequently, the velocity vector field is calculated using optical flow methods (such as dense optical flow or sparse optical flow). ,in, Represents the horizontal component (X direction) of the velocity vector. The vertical component (Y direction) of the velocity vector is represented, and the instantaneous velocity field is obtained through timestamp alignment. .

[0095] To improve the stability of wind pressure calculations, local wind pressure correction is used to obtain local wind pressure. Ultimately, this will reduce the local wind pressure in all folded areas. According to its area weight A weighted average is performed to obtain global and local wind pressure data. The formula is as follows:

[0096]

[0097] This application improves the accuracy and stability of wind pressure estimation under dynamic wind conditions by introducing a velocity fluctuation correction term and a weighted average mechanism, providing more reliable basic data for subsequent tension compensation.

[0098] This application further proposes to calculate the instantaneous velocity field in the folded region of the tarpaulin. The specific steps include:

[0099] The acquisition timestamps of adjacent image frames are matched with the acquisition timestamps of real-time tension data, and the real-time tension change is calculated by time difference calculation of the real-time tension data. ;

[0100] Real-time tension change Injecting a velocity field correction model, the output is an instantaneous velocity field corrected by curvature constraints and tension compensation. :

[0101]

[0102] in, The original velocity component (obtained by optical flow method) is represented by α, the material creep coefficient is represented by E, and the elastic modulus is represented by E. α and E are obtained from a tarpaulin material parameter database. Where A is the characteristic length of the folded region (calculated from the folded feature data), and A is the unit area of ​​the tarpaulin. This represents the time interval between the acquisition of adjacent image frames.

[0103] The tarpaulin material parameter database is a local pre-configured database that stores the creep coefficient α and elastic modulus E of different tarpaulin materials. The database includes standard mechanical parameters for polyester fiber and PVC coating, and supports user updates through the system interface.

[0104] Material parameter database (partial examples):

[0105]

[0106] Among them, the acquisition timestamps of adjacent image frames With the acquisition timestamp of real-time tension data The two types of data are matched using time interpolation and linear regression. Time difference calculation is then performed on the aligned real-time tension data T(t) to obtain the real-time tension change at each feature point.

[0107]

[0108] in, This represents the tension of the tarpaulin at the i-th feature point at time t. This represents the tension of the tarpaulin at the same characteristic point after a time interval Δt following time t. It represents the real-time tension change. The injection velocity field correction model is used to obtain the instantaneous velocity field after curvature constraint and tension compensation. .

[0109] It also includes obtaining the fold curvature radius of the tarpaulin based on the fold feature data of the tarpaulin. And using the rate of change of the normal vector of the folded surface For displacement vector field Apply physical constraints:

[0110]

[0111] in, This is the normal vector of the folded surface.

[0112] In the finite element method implementation, the tarpaulin surface is meshed into triangular or quadrilateral elements with a mesh resolution of ≥20×20 pixels / element. Constraint equations for displacement and normal vectors are established at the element nodes, and the displacement field is solved by assembling the overall stiffness matrix and load vectors. In the finite difference method implementation, the curvature constraint equations are discretized into a system of difference equations in a two-dimensional mesh coordinate system. The pixel displacement vectors are iteratively updated until convergence to a stable solution satisfying the curvature condition. This approach not only ensures the geometric consistency of the displacement field but also enables efficient computation in large-area tarpaulin scenarios.

[0113] By introducing tension compensation, curvature constraints, and finite element / differential solution mechanisms, this method deeply integrates optical flow velocity field calculation with the mechanical properties and geometric deformation characteristics of tarpaulin materials, effectively eliminating wind pressure estimation errors caused by material creep, tension relaxation, and non-uniform deformation. Compared with simple image displacement calculation, this method can maintain high-precision prediction even under complex stress environments and provides a more reliable physical basis for determining critical wind speeds for tarpaulins and for freight safety early warning.

[0114] This application utilizes finite element discretization to solve the problem, ensuring that the velocity field not only conforms to the dynamic compensation of tension changes but also matches the physical laws of fold geometric deformation. By combining real-time tension changes with fold curvature constraints in the instantaneous velocity field calculation method, this application can not only perform high-precision modeling of the partial motion of the tarpaulin caused by wind but also compensate for calculation errors caused by material creep, tension relaxation, and non-uniform deformation. When combined with local wind pressure calculation formulas, this method can improve the spatial resolution and temporal response accuracy of wind pressure estimation, ultimately enhancing the accuracy of critical wind speed prediction and enabling early warning of the tarpaulin's safety status during freight transport.

[0115] This application further proposes that the specific steps for obtaining compensated tension distribution data by online compensation of local wind pressure data based on real-time tension data include:

[0116] Based on real-time tension data, the installation position of the tension sensor is mapped to the spatial coordinate system corresponding to the surface of the tarpaulin;

[0117] Under the same spatial coordinates in the spatial coordinate system, the corresponding aligned local wind pressure values ​​are directly extracted, and a matching data pair of tension and wind pressure is established;

[0118] The matched data pairs are fused using a time-series fusion algorithm based on adaptive Kalman filtering to calculate the fused tension distribution, which is then used as the compensation tension distribution data. The adaptive Kalman filter achieves dynamic fusion of tension and wind pressure data by adjusting the noise covariance matrix in real time.

[0119] Among them, the matching data pairs maintain temporal and spatial consistency during the fusion process.

[0120] When performing online compensation for local wind pressure data based on real-time tension data, this application first maps the spatial coordinates of the tension sensor's measuring points to the three-dimensional spatial coordinate system of the tarpaulin, and performs spatial registration with the wind pressure measuring points. To ensure time consistency, a two-way time synchronization algorithm is used to match the tension and wind pressure data, forming a matched data pair. .

[0121] An adaptive Kalman filter (AKF) is used for time-series fusion during the data fusion process. The filter update formula is as follows:

[0122]

[0123]

[0124] in, This is the updated estimate of the tension distribution; The observation vector (including wind pressure, tension, etc.); Based on the state at time k−1, we predict the tension distribution at time k; H is the observation matrix, which establishes the relationship between the state variables (estimated tension) and the observed quantities (measured tension). For Kalman gain, Let represent the prior estimate of the covariance matrix, and represent the tension distribution estimate. The degree of uncertainty. The observation noise covariance matrix represents the measured values. The unreliability (noise level) of the wind pressure measurement is adaptively adjusted to cope with fluctuations in wind pressure measurement.

[0125] The Kalman filter parameters are the initial state covariance matrix. =diag([0.1, 0.1, 0.1]), sets the uncertainty of the initial state (initial tension value) when the filter starts running. The value on the diagonal (0.1) indicates a large uncertainty in the initial guess; the observation noise covariance R=0.05 (adaptive adjustment range 0.02-0.1); the process noise covariance Q=0.01 (based on wind pressure measurement error calibration).

[0126] This application achieves real-time compensation for tension changes caused by local wind pressure by introducing a time-series fusion method based on adaptive Kalman filtering, thereby improving the real-time performance and accuracy of tension distribution estimation. The compensated tension distribution data output after fusion is perfectly aligned with the stress state of the tarpaulin in both time and space.

[0127] This application further proposes that the freight safety monitoring method also includes generating tarpaulin tension distribution heat map data based on compensation tension distribution data, the steps of which are as follows:

[0128] The compensation tension distribution data and multispectral image data are registered in a spatial coordinate system to obtain the pixel position of the compensation tension distribution data in the image;

[0129] The surface of the tarpaulin is divided into regions based on the pixel position and the grid cells of a predetermined number of bits. The tension value of the compensation tension distribution data corresponding to each grid cell is mapped to the corresponding color-coded value. The range of the color-coded value is mapped to the range of the tension value in a linear proportional relationship.

[0130] The mapped grid cells are interpolated and smoothed to generate a continuous color matrix of tension distribution as heat map data of tarpaulin tension distribution.

[0131] Specifically, the compensated tension distribution data obtained through tension compensation and curvature constraint correction is registered with the corresponding multispectral image data in a three-dimensional spatial coordinate system, so that each tension measurement point in the compensated tension distribution data has pixel position coordinates (u,v) in the image. The spatial registration is completed based on the three-dimensional grid coordinate system of the fold and knot features output by the aforementioned visual Transformer, thereby ensuring a one-to-one correspondence between the tension data and the image position.

[0132] Next, the surface of the tarpaulin is divided into multiple grid cells according to a preset grid resolution (e.g., 20×20, 40×40, etc.), with each grid cell containing a number of pixels. For each grid cell, the average tension value of the compensation tension distribution data within its coverage area is calculated. And map it to color-coded values :

[0133]

[0134] in, and These are the minimum and maximum tension values ​​of all grid cells at the current moment, respectively. and This represents the lower and upper limits of the color mapping range (e.g., from blue to red).

[0135] To avoid color-blocked boundaries in the tension heatmap, bilinear interpolation or Gaussian kernel-based convolution smoothing is applied to the color matrix of the mapped grid cells, resulting in more continuous color transitions and softer edges. The resulting smooth color matrix is ​​the tarpaulin tension distribution heatmap data. This heatmap can be directly overlaid on a multispectral image or bound to the tarpaulin's geometric model in a 3D visualization interface to achieve real-time spatial force visualization.

[0136] This application utilizes a tarpaulin tension distribution heatmap to transform high-dimensional physical tension data into a simple, intuitive, and readable color image. This allows operators to quickly identify areas of abnormal stress and potential failure risks without relying on complex numerical analysis. Combined with wind pressure calculation and tension compensation methods, the visualization accuracy and timeliness of the tarpaulin tension distribution heatmap are improved, facilitating early warnings and guiding on-site handling under severe weather or high-speed transportation conditions, thereby enhancing freight safety.

[0137] This application further proposes that the specific steps for generating the critical wind speed value of the tarpaulin based on tarpaulin fold feature data, rope knot structure feature data, and compensation tension distribution data include:

[0138] The overall stability index of the tarpaulin is calculated based on the tarpaulin fold feature data and the rope knot structure feature data.

[0139] The stability index and compensation tension distribution data are input into the critical wind speed prediction model built on LSTM neural network, and the critical wind speed value is output.

[0140] Among them, the critical wind speed prediction model is trained based on the correlation data between historical wind pressure data and tarpaulin failure events, and establishes a mapping relationship between stability index, compensation tension distribution data and critical wind speed value.

[0141] Based on the tarpaulin fold feature data and rope knot structure feature data output by the visual Transformer model, parameters such as fold amplitude, radius of curvature distribution, number of knots, and knot position are extracted. These parameters are then combined with the aforementioned compensation tension distribution data to calculate the overall stability index S of the tarpaulin. The stability index is used to comprehensively measure the structural integrity and stress balance of the tarpaulin, and its calculation formula is as follows:

[0142]

[0143] in, It is a set of fold features (including mean curvature, curvature variance, fold density, etc.). This is a set of knot features (including the number of knots, knot density, and distribution uniformity). To obtain compensation tension distribution data The extracted tension feature set (including tension mean, maximum, coefficient of variation, etc.); These are the normalization functions for the corresponding feature sets, which map features of different dimensions to the interval [0,1]. The weighting coefficients are determined by fitting historical data.

[0144] The weights for fold features (reflecting the impact of folds on stability, such as fold density and curvature variance); The feature weight of the rope knot (reflecting the contribution of the number and tightness of knots to structural integrity); The tension distribution weight (reflects the influence of the mean and coefficient of variation of the compensating tension on the force balance).

[0145] In the specific implementation process, The weight range was determined through a three-factor, three-level orthogonal experiment, and a grid search fitting was performed using historical failure data. The specific steps are as follows:

[0146] Three types of typical tarpaulins (polyester fiber, PVC coating, and canvas) were selected, and wind speed environments of level 5 (10.8-13.8 m / s), level 6 (13.9-17.1 m / s), and level 7 (17.2-20.7 m / s) were simulated in a wind tunnel laboratory. 100 sets of valid failure samples (including wrinkle, knot, and tension data) were collected.

[0147] The following indicators were determined: wrinkle feature set Mean curvature (0-1), curvature variance (0-0.5), fold density (0-0.3); knot feature set : Number of knots (0-10), uniformity of distribution (0-1), tightness (0-1); tension feature set : Mean tension (0-500N), coefficient of variation (0-0.4).

[0148] A grid search method is used to traverse the interval [0.1, 0.5]. The optimal weights are determined by combining the stability index S with the goodness of fit R² to the actual failure probability as the objective function. =0.35, =0.25, =0.4 (Goodness of fit R²=0.92, Mean squared error MSE=0.03).

[0149] Next, the stability index S and the compensation tension distribution data are fed as joint inputs into a critical wind speed prediction model constructed based on a Long Short-Term Memory (LSTM) network. The input sequence of this model consists of data from several past time points. Corresponding value, The stability index represents time t. This represents the compensation tension distribution data at time t, and the output is the predicted critical wind speed value. Historical wind pressure data was used during model training. Data associated with tarpaulin failure events (including rupture and detachment) are used as a tag set, and a mapping relationship between stability index, tension distribution, and critical wind speed value is established by sliding a time window.

[0150] To improve prediction accuracy, the loss function of the critical wind speed prediction model may include a weighted combination of the mean squared error (MSE) term and the classification cross-entropy term:

[0151]

[0152] Where N refers to the number of samples in a training batch. and These are labels representing the actual and predicted failure status of the tarpaulin, respectively. This represents the cross-entropy loss function, which measures the difference between the distribution of the true labels and the distribution of the predicted probabilities. and These are weighting coefficients. and These are the actual and predicted critical wind speed values ​​for the i-th sample, respectively.

[0153] In the specific implementation process, initialization is based on the sample imbalance coefficient. =0.6, =0.4 (30% of the samples were failed, and 70% were non-failed).

[0154] Dynamic adjustment and :

[0155] Early training phase (first 100 rounds): =0.7, =0.3 (prioritize optimizing wind speed regression accuracy);

[0156] Mid-training period (100-500 rounds): =0.5, =0.5 (Balanced Regression and Classification);

[0157] Later in training (after 500 rounds): Adaptive adjustment based on the validation set F1 score; when the classification F1 score < 0.85, Increase by 0.05, or vice versa. Increase by 0.05.

[0158] After 5-fold cross-validation (validation set accuracy = 0.91, F1 score = 0.89), the optimal weights are: =0.55, =0.45.

[0159] Finally, the output critical wind speed value can be compared with the real-time wind speed data of the current meteorological monitoring system. When the real-time wind speed approaches or exceeds the preset proportional threshold of the critical wind speed value (e.g., 0.85 or 0.9), the system will trigger a safety warning.

[0160] This application introduces a stability index as a comprehensive characterization of structural stress and combines it with an LSTM neural network model trained using historical wind pressure and failure data. This embodiment can dynamically predict the wind resistance limit of the tarpaulin under its current condition. Compared to methods based solely on tension or wind pressure, this prediction method improves the accuracy and timeliness of early warnings, enabling operators to take measures such as reinforcement, speed reduction, or tarpaulin replacement before strong winds arrive, thereby effectively reducing cargo loss and the risk of safety accidents.

[0161] like Figure 4 As shown, this application further proposes that the process of inputting the stability index and compensation tension distribution data into the critical wind speed prediction model and outputting the critical wind speed value specifically includes:

[0162] The stability index and compensation tension distribution data are normalized to the same numerical range and then processed into a time series according to a predetermined number of bits to form a time series feature vector group.

[0163] The time-series feature vector groups are weighted and combined to generate a comprehensive stress index;

[0164] Real-time wind speed values ​​are collected, and a safe tension threshold is dynamically calculated. The comprehensive stress index is compared with the safe tension threshold. When the comprehensive stress index exceeds the safe tension threshold, the real-time wind speed value collected at this time is used.

[0165] The safety tension threshold is dynamically calculated using the material's tensile strength and stability index. The material's tensile strength is the tensile performance index value obtained from the tensile test results of the tarpaulin material sample or from the tarpaulin material parameter database.

[0166] The stability index S and the compensation tension distribution data at time t are used. The data are normalized to the same numerical range (e.g., [0,1]) and rearranged into a time series of length N according to a predetermined number of bits N, forming a time series feature vector group. .

[0167] Time series feature vector group The comprehensive stress index E is obtained by weighting and combining the results according to the weights.

[0168]

[0169] in, , To integrate weights, satisfy , For time-averaged tension:

[0170]

[0171] Where N is a time series of length N. The stability index represents time t. This represents the compensation tension distribution data at time t.

[0172] Safety tension threshold Dynamic calculations are performed using the material's tensile strength and stability index S, as shown in the following formula:

[0173]

[0174] in, β represents the tensile strength of the tarpaulin material, derived from the tensile test results of tarpaulin material samples or a tarpaulin material parameter database; β is the safety factor, typically taken as 0.7–0.9 to introduce a safety margin.

[0175] Compare the comprehensive stress index E with the safety tension threshold Comparison, when E≥ At that time, the corresponding wind speed value is calculated as the critical wind speed value. In calculations based on fluid dynamics principles, the critical wind speed can be deduced inversely from the wind pressure formula:

[0176]

[0177] in, ρ is the wind pressure coefficient (calibrated by wind tunnel experiments or numerical simulations), and ρ is the air density.

[0178] Collect real-time wind speed values ,when When η is the warning trigger ratio coefficient, for example, 0.85–0.9, the system triggers a critical wind speed warning and can automatically link with safety strategies such as reducing the speed of freight vehicles, reinforcing tarpaulins, or rerouting.

[0179] This application introduces a comprehensive stress index based on stability index and compensation tension distribution data, and dynamically calculates the safety tension threshold using the material's tensile strength. This embodiment can achieve dynamic and real-time prediction of critical wind speed values ​​by combining structural characteristics, stress state, and material ultimate performance. Compared with methods based solely on historical wind speed data or a single tension index, this scheme improves both the accuracy and adaptability of wind speed prediction, thus making the early warning of tarpaulin failure risk more accurate and reliable.

[0180] The following is a specific implementation example of freight safety monitoring using tarpaulin status image recognition:

[0181] To prevent the risk of tarpaulin tearing due to strong winds during steel transport, a safety monitoring system based on this invention was deployed on a railway train. The system consists of a multispectral imaging module (including a visible / near-infrared dual-band camera with a resolution of 1920×1080 and a sampling rate of 10Hz) installed on the roof of the train carriages, tension sensors (range 0-5000N, accuracy ±1%FS) positioned at the four corners of the tarpaulin and at rope anchor points, and real-time data processing via an onboard industrial computer.

[0182] During transport, images of the tarpaulin surface were continuously captured. When encountering crosswinds (real-time weather data showed a wind speed of 12 m / s), the system activated a high-frequency acquisition mode. Edge enhancement (Canny operator, thresholds T1=60, T2=180) and binarization were performed on the images. Morphological thinning algorithms were used to extract the tarpaulin edge skeleton lines, generating a topological guide graph. This graph uses a node set N={ , ,..., (15 feature nodes) and edge set (12 lines) storage, node attributes include local curvature (mean) and mean spectral reflectance (0.35±0.04).

[0183] A visual Transformer model (pre-trained on 100,000 labeled tarpaulin images) receives a topological guidance map and outputs fold feature data (fold position coordinates (350mm, 420mm), radius of curvature R=0.8m) and rope knot structure feature data (knot position (120mm, 280mm), tightness 0.72). Based on the pixel coordinate changes in the fold region of adjacent frames, the displacement vector field is calculated using the optical flow method, yielding the original velocity component vᵢ=1.2m / s. Combined with real-time tension data (sensor T3 acquisition value 2800N, timestamp matching error <0.1s), the instantaneous velocity field is calculated using a velocity field correction model.

[0184]

[0185] Substitute the parameters: α = 0.02 / h (creep coefficient of polyester fiber), ΔTᵢ = 150N (tension change), Δt = 0.1s (image frame interval). =1.2m (characteristic length of fold), E=3.5GPa (elastic modulus), A=0.01m² (unit area), calculated as follows: .

[0186] Will Substituting into the local wind pressure formula:

[0187]

[0188] Where k = 0.6 (wind tunnel test calibration coefficient), ρ = 1.225 kg / m³ (air density), and γ = 0.2 (wind pressure correction factor). =0.15m / s (standard deviation of velocity), calculated as follows =0.6×1.225×(1.32)² + 0.2×0.15≈1.31Pa.

[0189] By fusing wind pressure and tension data using adaptive Kalman filtering, compensated tension distribution data (mean 3200N, coefficient of variation 0.18) is obtained. The system generates a tarpaulin tension heatmap, showing a tension value of 3800N in the left rear corner area (color-coded as a red warning color). Based on the fold feature set ( ), Knot Feature Set ( ) and tension feature set ( Calculate the stability index:

[0190]

[0191] Substitution =0.35、 =0.25、 =0.4 (grid search optimization weight). =0.65 (normalized value of wrinkle feature). =0.82 (normalized value of knot feature). =0.70 (normalized value of tension characteristic), so S=0.35×0.65+ 0.25×0.82 + 0.4×0.70≈0.71.

[0192] Inputting the S and compensation tension data into the LSTM critical wind speed prediction model (input time series length 20, hidden layer dimension 128), the output critical wind speed value is 17.8 m / s. At this time, the real-time wind speed has reached 16.5 m / s (close to 92% of the critical value), and the system triggers a level 2 warning: the vehicle terminal displays a red alarm, the buzzer sounds, and at the same time, it links with the vehicle ECU via the CAN bus to limit the maximum vehicle speed to 60 km / h (originally 80 km / h).

[0193] After receiving the warning, the driver reported it. Railway staff stopped at the next station to inspect and found that the knot of the rope at the left rear corner was slightly loose (corresponding to the red area on the heat map). After tightening it again, the tension value dropped to 3000N, and the system predicted that the critical wind speed would rise to 20.5m / s. During the subsequent transportation process, despite encountering a momentary gust of 18m / s, the tarpaulin remained stable and ultimately arrived at its destination safely.

[0194] This embodiment achieves precise monitoring of tarpaulin folds, knot structures, and stress states through multimodal data fusion of multispectral imaging and tension sensing. The topology guidance map enhances the visual Transformer's ability to recognize local features, the tension compensation algorithm corrects wind pressure errors from pure image calculations, and the LSTM model combined with the stability index dynamically predicts critical wind speeds, solving the problems of "delayed warnings" and "high false alarm rates" in traditional monitoring. The system triggers a warning 15 minutes in advance in strong winds, giving drivers time to react and avoiding the risk of cargo falling due to tarpaulin tearing, demonstrating the advantages of end-to-end intelligent monitoring.

[0195] Example 2:

[0196] like Figure 5 and Figure 6 As shown, a freight safety monitoring system based on tarpaulin status image recognition, using the aforementioned freight safety monitoring method based on tarpaulin status image recognition, includes:

[0197] The multispectral imaging module is used to continuously acquire multispectral image data of the tarpaulin surface;

[0198] The skeleton extraction module is used to extract the skeleton data of the tarpaulin outline based on multispectral image data and generate a topology guide map;

[0199] The visual recognition module is used to input the topology guidance diagram into a pre-built visual Transformer model and output tarpaulin fold feature data and rope knot structure feature data.

[0200] The wind pressure calculation module is used to calculate the local wind pressure data of the tarpaulin fold area based on the displacement mapping relationship between the tarpaulin fold feature data and the multispectral image data.

[0201] The tension compensation module is used to receive real-time tension data and perform online compensation on local wind pressure data to obtain compensated tension distribution data.

[0202] The critical generation module is used to generate the critical wind speed value of the tarpaulin based on the tarpaulin fold feature data, rope knot structure feature data, and compensation tension distribution data.

[0203] The early warning module is used to generate and output early warning data when the critical wind speed value is lower than the preset wind speed threshold.

[0204] The safety linkage interface module is configured to convert warning data into TCN train communication network protocol commands to trigger emergency braking or speed limit control.

[0205] The multispectral imaging module includes a multispectral camera array, a wide-angle lens assembly, and a mounting bracket. It is installed above or to the side of the cargo compartment of a freight vehicle to continuously acquire multispectral image data of the tarpaulin surface. The multispectral camera array has visible light, near-infrared, and short-wave infrared channels, supports high frame rate acquisition, and is synchronized with the system's main control unit via a time synchronization controller.

[0206] The skeleton extraction module includes an embedded GPU computing board and an image processing unit, which is used to perform edge detection, morphological processing and skeleton extraction on multispectral image data to generate skeleton data of the tarpaulin outline, and construct a topology guidance graph based on the connection relationship between skeleton nodes to provide structural constraints for subsequent visual feature extraction.

[0207] The visual recognition module includes a deep learning inference chip, a storage unit, and a pre-built visual Transformer model. This module inputs the topology guidance graph into the visual Transformer model to extract tarpaulin fold feature data and rope knot structure feature data. The visual Transformer model is pre-trained on a labeled dataset containing multiple types of tarpaulins and binding methods, and stored in the system's storage unit, supporting online updates of weight parameters.

[0208] The wind pressure calculation module includes a processor and a local cache. It is used to calculate the local wind pressure data of the tarpaulin fold area based on the displacement mapping relationship between the tarpaulin fold feature data and multispectral image data. It is then corrected by combining real-time data from vehicle speed and wind direction sensors, and outputs a local wind pressure distribution matrix.

[0209] The tension compensation module includes a tension sensor array, a signal acquisition unit, and a data processing unit. Tension sensors are positioned at the edges of the tarpaulin and along the main load-bearing ropes to collect real-time tension data and fuse it online with local wind pressure data to generate compensated tension distribution data. This module supports tension compensation curve adjustment based on a material parameter database.

[0210] The critical generation module includes a model inference unit and memory, and incorporates a critical wind speed prediction model built on an LSTM neural network. This module receives tarpaulin fold feature data, rope knot structure feature data, and compensation tension distribution data, calculates the tarpaulin stability index, and generates the predicted critical wind speed value.

[0211] The early warning module includes an in-vehicle display terminal, a buzzer, and a communication unit. When the critical wind speed value is lower than the preset wind speed threshold, it generates and outputs early warning data. This can be achieved through audible and visual alarms or by uploading the alarm information to the in-vehicle control system and remote dispatch center, enabling automatic intervention or manual handling. The early warning module is linked to the vehicle's ECU to limit the vehicle speed to less than or equal to 80% of the maximum speed.

[0212] Both the multispectral imaging module and the tension sensor passed the vibration test of GB / T 21563-2018 for rail transit equipment, with a positioning error of ≤±2mm under random vibration of 1-200Hz. When applied to high-speed railways (speed ≥120km / h), the system activates high-speed mode: the multispectral sampling rate is increased to 20Hz (originally 10Hz), the displacement mapping adopts a motion fuzzy compensation algorithm, and the wind pressure calculation module loads a high-speed railway-specific aerodynamic coefficient library.

[0213] During a specific railway freight transport operation, the system's multispectral imaging module continuously acquired images of the tarpaulin surface, the skeleton extraction module generated a topology guidance map in real time, and the visual recognition module identified tarpaulin folds and rope knots. The wind pressure calculation module calculated the local wind pressure distribution, and the tension compensation module performed compensation and correction based on tension sensor data. The critical generation module calculated the critical wind speed to be 18.5 m / s. When the early warning module determined that this value was below the set threshold (20 m / s), the system immediately triggered an audible and visual warning and notified railway staff through the communication unit, while simultaneously uploading the data to the monitoring platform for record-keeping.

[0214] The system in this embodiment achieves dynamic and accurate prediction of the stress state and critical wind speed of tarpaulins through the synergistic effect of multispectral imaging, skeleton extraction, visual Transformer recognition, physical wind pressure calculation, and real-time tension compensation. This system retains the interpretability of the physical model while utilizing deep learning to improve the accuracy of complex feature extraction and prediction, thereby enhancing the ability to detect and intervene in tarpaulin failure risks during freight transportation and reducing the incidence of cargo damage and traffic accidents.

[0215] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A freight security monitoring method of tarp status image recognition, characterized in that: The method comprises the following steps: continuously collecting multispectral image data of the tarpaulin surface; extracting skeleton data of the tarpaulin profile based on the multispectral image data, and generating a topological guide map using the skeleton data; inputting the topological guide map into a pre-constructed visual Transformer model to output tarpaulin wrinkle feature data and rope knot structure feature data; calculating local wind pressure data of the tarpaulin wrinkle area according to the displacement mapping relationship between the tarpaulin wrinkle feature data and the multispectral image data; receiving real-time tension data collected in real time, and compensating the local wind pressure data based on the real-time tension data to obtain compensated tension distribution data; generating a critical wind speed value of the predicted tarpaulin based on the tarpaulin wrinkle feature data, the rope knot structure feature data, and the compensated tension distribution data; generating and outputting warning data when the critical wind speed value is lower than a preset wind speed threshold; wherein the specific steps of calculating the local wind pressure data of the tarpaulin wrinkle area according to the displacement mapping relationship between the tarpaulin wrinkle feature data and the multispectral image data include: extracting a set of pixel point coordinates of the tarpaulin wrinkle area corresponding to adjacent time points, and establishing a displacement mapping relationship based on the set of pixel point coordinates; calculating the displacement vector field of the tarpaulin wrinkle area at adjacent time points using the optical flow method to obtain the velocity components of each pixel point in the set of pixel point coordinates in a two-dimensional plane; combining the time-aligned velocity components with the acquisition time intervals of adjacent image frames of the multispectral image data calculating a transient velocity field of the tarp wrinkle region ; The instantaneous velocity field Substitute into the local wind pressure calculation formula: ， wherein, is a local wind pressure value, k is a preset proportionality coefficient, ρ is air density, is a wind pressure correction factor, is a variance term of the velocity component, and σ is a velocity standard deviation. Local wind pressure values for all tarp wrinkle areas The weighted average is performed to generate the local wind pressure data corresponding to the spatial position of the tarp surface and matching the real-time tension data in both time and space.

2. The cargo security monitoring method of tarp status image recognition according to claim 1, characterized in that: The specific steps of extracting skeleton data of the tarpaulin profile based on the multispectral image data and generating a topological guide map using the skeleton data include: performing edge enhancement and binarization processing on the multispectral image data, extracting a binary profile line using a morphological thinning algorithm, and generating tarpaulin edge skeleton line data through topological constraint; establishing node and connection topological structure data based on the tarpaulin edge skeleton line data, and constructing the topological guide map according to the topological structure data; The topology data is stored in the form of a node set and an edge set , represents the i-th skeleton node, represents an edge connecting the skeleton node and the skeleton node .

3. The cargo security monitoring method of tarp status image recognition according to claim 1, characterized in that: The specific steps of inputting the topological guide map into a pre-constructed visual Transformer model to output tarpaulin wrinkle feature data and rope knot structure feature data include: mapping the topological guide map through a sine function to convert the skeleton node coordinates into a topological position encoding vector, and converting the multispectral image data into a corresponding image block feature vector; performing feature splicing on the topological position encoding vector and the image block feature vector in the input layer of the visual Transformer model; in the encoder of the visual Transformer model, calculating the spatial correlation between node features and the spectral correlation between image block features through a multi-head self-attention mechanism, and performing fusion; outputting the tarpaulin wrinkle feature data and the rope knot structure feature data in the decoder of the visual Transformer model; wherein the tarpaulin wrinkle feature data is a set of wrinkle position and shape parameters of the tarpaulin labeled in a spatial coordinate system, and the rope knot structure feature data is a set of knot position and shape parameters of the tarpaulin labeled in the same spatial coordinate system; the spatial coordinate system is a three-dimensional grid coordinate system covering the tarpaulin surface.

4. The cargo security monitoring method of tarp status image recognition according to claim 1, characterized in that: computing a transient velocity field of the tarp wrinkle region The specific steps include: matching the acquisition time stamp of the adjacent image frame with the acquisition time stamp of the real-time tension data, calculating the real-time tension change amount by time difference calculation on real-time tension data ; The real-time tension change amount The injection velocity field correction model outputs a corrected instantaneous velocity field that is curvature-constrained and tension-compensated : ， wherein, is the original velocity component, a is the material creep coefficient, E is the elastic modulus, a and E are obtained through the tarpaulin material parameter database, is the characteristic length of the wrinkle region, A is the unit area of the tarpaulin, is the acquisition time interval of the adjacent image frames.

5. The cargo security monitoring method of tarp status image recognition according to claim 3, characterized in that: The specific steps of compensating the local wind pressure data based on the real-time tension data to obtain compensated tension distribution data include: Mapping the installation position of the tension sensor to the corresponding spatial coordinate system of the tarp surface based on the real-time tension data; Directly extracting the corresponding aligned local wind pressure value under the same spatial coordinate of the spatial coordinate system and establishing a matching data pair of tension and wind pressure; Calculating the fused tension distribution result through an adaptive Kalman filter-based time series fusion algorithm for the matching data pair, and taking the tension distribution result as the compensated tension distribution data; Wherein, the matching data pair maintains the consistency of time and space during the fusion process.

6. The cargo security monitoring method of tarp status image recognition according to claim 5, characterized in that: The freight security monitoring method further includes generating a tarp tension distribution heat map data based on the compensated tension distribution data, and the steps are as follows: Registering the compensated tension distribution data and the multispectral image data in the spatial coordinate system to obtain the pixel position of the compensated tension distribution data in the image; Based on the pixel position, the tarp surface is divided into regions according to a predetermined number of grid units, and the tension value of each grid unit corresponding to the compensated tension distribution data is mapped to a corresponding color coding value, wherein the range of the color coding value is linearly proportional to the range of the tension value. Interpolation smoothing processing is performed on the mapped grid units to generate a continuous tension distribution color matrix as the tarp tension distribution heat map data.

7. The tarp status image recognition freight security monitoring method according to claim 1, characterized in that: The specific steps of generating a predicted critical wind speed value of the tarp based on the tarp wrinkle feature data, the rope knot structure feature data, and the compensated tension distribution data include: Calculating the stability index of the tarp as a whole based on the tarp wrinkle feature data and the rope knot structure feature data; Inputting the stability index and the compensated tension distribution data into a critical wind speed prediction model based on an LSTM neural network to output the critical wind speed value; Wherein, the critical wind speed prediction model is trained based on historical wind pressure data and associated data of tarp failure events, and a mapping relationship between the stability index, the compensated tension distribution data, and the critical wind speed value is established.

8. The tarp status image recognition freight security monitoring method according to claim 7, characterized in that: The process of inputting the stability index and the compensated tension distribution data into the critical wind speed prediction model to output the critical wind speed value specifically includes: Normalizing the stability index and the compensated tension distribution data to the same numerical range, respectively, and performing time series processing according to a predetermined number to form a time series feature vector group; Combining the time series feature vector group to generate a comprehensive stress indicator; Collecting real-time wind speed values, dynamically calculating a safety tension threshold, comparing the comprehensive stress indicator with the safety tension threshold; when the comprehensive stress indicator exceeds the safety tension threshold, the real-time wind speed value collected at that time is taken as the critical wind speed value; Wherein, the safety tension threshold is dynamically calculated by the material tensile strength and the stability index, and the material tensile strength is an anti-tensile performance indicator value obtained from the tensile test results of tarp material samples or a tarp material parameter database.

9. A freight security monitoring system for tarp status image recognition, characterized by: The cargo security monitoring method using the tarpaulin state image recognition method according to any one of claims 1-8 comprises: a multispectral imaging module for continuously collecting multispectral image data of the tarpaulin surface; a skeleton extraction module for extracting skeleton data of the tarpaulin profile based on the multispectral image data and generating a topological guide map; a visual recognition module for inputting the topological guide map into a pre-constructed visual Transformer model and outputting tarpaulin wrinkle feature data and rope knot structure feature data; a wind pressure calculation module for calculating local wind pressure data of the tarpaulin wrinkle region according to a displacement mapping relationship between the tarpaulin wrinkle feature data and the multispectral image data; a tension compensation module for receiving real-time tension data and online compensating the local wind pressure data to obtain compensated tension distribution data; a criticality generation module for generating a critical wind speed value of the predicted tarpaulin based on the tarpaulin wrinkle feature data, the rope knot structure feature data, and the compensated tension distribution data; a warning module for generating and outputting warning data when the critical wind speed value is lower than a preset wind speed threshold; wherein the specific steps of calculating the local wind pressure data of the tarpaulin wrinkle region according to the displacement mapping relationship between the tarpaulin wrinkle feature data and the multispectral image data include: extracting a set of pixel point coordinates of the tarpaulin wrinkle region corresponding to adjacent time points, and establishing a displacement mapping relationship based on the set of pixel point coordinates; using an optical flow method to calculate the displacement vector field of the tarpaulin wrinkle region at adjacent time points to obtain the velocity components of each pixel point in the set of pixel point coordinates on a two-dimensional plane; combining the time-aligned velocity components with the acquisition time intervals of adjacent image frames of the multispectral image data calculating a transient velocity field of the tarp wrinkle region ; The instantaneous velocity field Substitute into the local wind pressure calculation formula: ， wherein, is the local wind pressure value, k is a preset proportionality coefficient, ρ is the air density, is the wind pressure correction factor, is the variance term of the velocity component, and σ is the velocity standard deviation. Local wind pressure values for all tarp wrinkle areas The weighted average is performed to generate local wind pressure data corresponding to the spatial position of the tarp surface and matching both the real-time tension data in time and space.

Images