Vehicle gear device anti-collision strip water immersion alarm system based on graph neural network

Abstract

The invention discloses a vehicle gear device anti-collision strip water immersion alarm system based on a graph neural network, and the system comprises a sensing data collection module which collects the time series data of multiple sensing nodes; the graph structure generation module is used for constructing an adjacent matrix and generating a direction weight matrix according to the length direction of the anti-collision strip; the node feature and stability construction module is used for generating a node feature sequence and a stability factor; the frequency domain decomposition module is used for executing structure sensing atlas transformation on the node features in the improved FAGCN model to obtain a first frequency domain component and a second frequency domain component; the gating weight correction module is used for generating a corrected gating weight in combination with the stability factor and the frequency domain component; the direction modulation module is used for performing direction weighting on the frequency domain components; the node weighting calculation module is used for generating a convolution output sequence; and the consistency screening and alarm output module outputs an immersion alarm signal. The water immersion state of the anti-collision strip is reflected more reliably, and stable warning is provided.

Description

Technical Field

[0001] This invention relates to the field of graph neural networks and vehicle safety monitoring technology, and in particular to a vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural networks. Background Technology

[0002] With the increasing integration of vehicle electronic systems and the prevalence of complex driving environments, the risk of critical components being submerged in water during vehicle operation in flooded or waterlogged areas is becoming increasingly prominent. This is especially true for the gear shifter and its protective structure, located inside or in the lower part of the vehicle. Water ingress can easily lead to functional failure, short circuits, corrosion, and even safety hazards. Therefore, real-time monitoring and warning technologies for vehicle submersion have gradually become an important research direction in the field of vehicle safety monitoring. Existing technologies for vehicle submersion or flood warning schemes mostly use single-point or a small number of water level sensors, combined with threshold comparison and rate of change judgment to achieve alarm output. Some schemes also upload water level information to the vehicle system or remote terminal via a communication module to prompt the driver to take evasive action.

[0003] However, in real-world applications, vehicle gear shift bumper strips are typically elongated structures, and water immersion often spreads gradually along these structures, exhibiting significant spatial propagation characteristics. Existing monitoring methods based on single points or a limited number of sensors struggle to comprehensively reflect the state changes at different locations along the bumper strip, easily leading to situations where localized water immersion goes undetected or is misjudged as overall immersion. Furthermore, the sensor nodes in the bumper strip area are susceptible to environmental temperature changes, vehicle vibration, and signal jitter during operation, resulting in time-series observation data exhibiting significant non-stationarity and noise interference. Traditional judgment methods based on fixed thresholds or simple rules struggle to distinguish between genuine water immersion signals and transient interference signals, leading to a high false alarm rate.

[0004] Furthermore, as the number of sensing nodes increases, the spatial relationships between nodes and the direction of signal propagation gradually become key factors affecting the accuracy of judgments. However, existing technologies generally lack systematic modeling of the spatial structure of multiple nodes and fail to effectively utilize the arrangement order and adjacency relationship of nodes along the length of the crash barrier. Some methods that introduce machine learning or neural networks focus primarily on temporal feature modeling, ignoring the structural constraints and directional differences between nodes, and lack a quantification mechanism for node reliability within the model, making it easy for unstable nodes to have an excessive impact on the overall judgment result. At the same time, existing solutions typically make judgments based on only a single time point or abnormal results within a short period during the alarm output stage, lacking a joint screening mechanism for temporal and spatial continuity, making it difficult to effectively suppress isolated abnormal alarms caused by transient noise.

[0005] Therefore, how to provide a vehicle gear shifter anti-collision strip water immersion warning system based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural networks. The invention constructs a multi-sensor node collaborative modeling mechanism based on graph structure, incorporating the spatial arrangement of the vehicle gear shifter anti-collision strip and the direction of water immersion propagation into the data processing flow. By combining node stability factors and an improved FAGCN model, it achieves structural perception frequency domain decomposition of multi-node time-series signals. Through gating weight correction, direction modulation, and time and space consistency screening, it effectively suppresses instantaneous interference and false alarms, thereby improving the accuracy and stability of the anti-collision strip water immersion alarm.

[0007] A vehicle gear shifter anti-collision strip water immersion warning system based on a graph neural network according to an embodiment of the present invention includes: The sensor data acquisition module is used to collect time-series observation data from multiple sensor nodes arranged along the anti-collision strip of the vehicle gear shifter. The graph structure generation module is used to generate an adjacency matrix based on the physical installation location of the sensor nodes, and to generate a direction weight matrix based on the arrangement order of the sensor nodes along the length of the anti-collision strip. The node feature and stability construction module is used to construct node feature sequences and generate node stability factors; The frequency domain decomposition module is used to perform structure-aware map transformation on the node feature sequences in the improved FAGCN model to generate the first frequency domain component and the second frequency domain component. The gating weight correction module generates corrected gating weights based on the node stability factor and the first and second frequency domain components. The directional modulation module performs directional weighting and synthesis processing on the first frequency domain component and the second frequency domain component based on the directional weight matrix; The node weighted calculation module performs node-level weighted synthesis processing on the direction-modulated processing components based on the corrected gating weights to generate a convolutional output sequence. The consistency screening and alarm output module generates a water immersion alarm sequence based on the convolution output sequence and outputs a water immersion alarm signal.

[0008] Optionally, modules can be integrated using the following methods: S1. Construct a graph structure of multiple sensor nodes arranged along the anti-collision strip of the vehicle gear shifter, generate an adjacency matrix, and generate an orientation weight matrix based on the spatial arrangement order of the nodes on the anti-collision strip. S2. Construct node feature sequences based on time-series observation data of sensor nodes, and calculate node stability factors based on short-term node changes, temperature compensation, and historical difference components. S3. Input the node feature sequence into the improved FAGCN model, perform spectral transformation processing, and generate the first frequency domain component and the second frequency domain component; S4. Based on the node stability factor, perform a correction process on the gating weights of the improved FAGCN model to generate corrected gating weights; S5. Perform directional modulation processing on the first frequency domain component and the second frequency domain component based on the directional weight matrix; S6. Based on the modified gating weights, perform weighted calculations on the first and second processed components after direction modulation in the improved FAGCN model to generate a convolutional output sequence; S7. Based on the time variation pattern of the convolutional output sequence, perform consistency filtering processing to generate a water immersion alarm sequence and output a water immersion alarm signal. Optionally, S1 specifically includes: S11. Obtain the physical installation location data of multiple sensor nodes arranged along the anti-collision strip of the vehicle gear shifter; S12. Perform node connection relationship determination processing based on the physical installation distance between two adjacent sensing nodes. When the physical installation distance between adjacent nodes is within the preset distance threshold range, the adjacent nodes are defined as having a graph connection relationship, and the connection relationship is recorded in the adjacency matrix. S13. Generate a node arrangement sequence based on the physical installation order of all sensor nodes, and generate a direction index sequence based on the arrangement order of adjacent nodes in the anti-collision strip length direction in the node arrangement sequence. S14. Perform direction weight calculation processing based on the direction index sequence. For each pair of adjacent sensing nodes in the arrangement sequence, assign corresponding direction weight values ​​according to the order of arrangement of the nodes on the anti-collision strip, and record the direction weight values ​​in the direction weight matrix. S15. Construct the graph structure data for the input of the improved FAGCN model based on the adjacency matrix and the orientation weight matrix. The graph structure data includes the connection relationship data between nodes and the orientation weight data between nodes.

[0009] Optionally, S2 specifically includes: S21. Collect raw observation data of each sensor node arranged along the anti-collision strip of the vehicle gear shifter in a continuous time period, and perform time alignment processing on the raw observation data according to a uniform sampling time interval. S22. Based on the time-aligned original observation data, construct a node feature sequence according to the sensor node numbering order. Each time point in the node feature sequence corresponds to a node observation vector. S23. Perform short-time change calculation processing on the observation data of each sensing node within a preset time window, calculate the difference between the observation values ​​of adjacent sampling points within the time window and accumulate it. S24. Acquire the ambient temperature data corresponding to each sensing node, and perform temperature compensation processing on the node observation data based on the ambient temperature data and preset temperature compensation parameters to generate the temperature compensation amount. S25. Perform historical difference processing on the node observation vector of each sensing node in adjacent time windows, calculate the difference between the node observation vector of the current time window and the node observation vector of the previous time window, and generate historical difference components. S26. Based on the short-term change, the temperature compensation, and the historical difference, perform stability calculation processing on each sensing node to generate a corresponding node stability factor, and establish a correspondence between the node stability factor and the node feature sequence.

[0010] Optionally, S3 specifically includes: S31. In the improved FAGCN model, based on the adjacency matrix and the orientation weight matrix, the node feature sequence is subjected to joint normalization processing to generate normalized graph structure data containing node connection relationships and node orientation relationships. S32. Based on the normalized graph structure data and node feature sequence, perform structure-aware graph spectral transformation processing, and perform spectral domain mapping on the node feature sequence while simultaneously constraining the node connection relationship and the node arrangement direction relationship. S33. Based on the arrangement order of nodes in the anti-collision strip length direction in the spectral domain mapping result, perform spectral component reordering processing to generate a spectral component sequence that is consistent with the spatial arrangement order of nodes; S34. Based on the change amplitude of adjacent spectral components in the spectral component sequence, perform adaptive frequency domain partitioning processing to divide the spectral components into a first spectral component set and a second spectral component set. S35. Generate a first frequency domain component based on the first spectral component set, and generate a second frequency domain component based on the second spectral component set, and establish a correspondence between the first frequency domain component and the second frequency domain component and the node feature sequence respectively.

[0011] Optionally, S4 specifically includes: S41. In the improved FAGCN model, a corresponding initial gating weight is generated for each sensing node. The initial gating weight is determined based on the node feature sequence during the model training phase. S42. Based on the node stability factor, perform a first-stage correction process on the initial gating weights, and perform segmented proportional adjustment on the initial gating weights according to the numerical range corresponding to the node stability factor. S43. Based on the first frequency domain component and the second frequency domain component, calculate the relative change relationship between the two types of frequency domain components at the node dimension, and perform a second-stage correction process on the gate weights after the first-stage correction based on the relative change relationship. S44. Perform cross-window consistency constraint processing on the gating weight sequence corresponding to the same sensing node within adjacent time windows, and perform threshold pruning processing on the gating weight change amplitude within adjacent time windows. S45. The gate weights after segmentation ratio adjustment, frequency domain relationship correction and cross-window consistency constraint processing are determined as the modified gate weights, and participate in the weighted calculation of the first frequency domain component and the second frequency domain component in the improved FAGCN model.

[0012] Optionally, S5 specifically includes: S51. Based on the direction weight matrix, according to the arrangement order of the sensor nodes along the length direction of the anti-collision strip, determine the set of forward adjacent nodes and the set of backward adjacent nodes along the length direction for each sensor node. S52. For the first frequency domain component, traverse each sensing node in the node dimension, multiply the first frequency domain component corresponding to the forward adjacent node set by the first direction weight value, and multiply the first frequency domain component corresponding to the backward adjacent node set by the second direction weight value to generate the first direction weighted result. S53. For the second frequency domain component, traverse each sensing node in the node dimension, multiply the corresponding second frequency domain component in the forward adjacent node set by the third direction weight value, and multiply the corresponding second frequency domain component in the backward adjacent node set by the fourth direction weight value to generate the second direction weighted result. S54. Perform weighted synthesis processing on the first frequency domain component based on the first directional weighting result, and perform weighted synthesis processing on the second frequency domain component based on the second directional weighting result to generate the first processed component after directional modulation and the second processed component after directional modulation. S55. Store the first processed component after direction modulation and the second processed component after direction modulation as intermediate data in the improved FAGCN model.

[0013] Optionally, S6 specifically includes: S61. Obtain the first processed component and the second processed component after direction modulation, and construct the corresponding node processing sequence according to the sensor node numbering order. S62. Obtain the corrected gating weights and construct the corresponding gating weight sequence according to the sensor node number order; S63. Traverse each sensing node along the node dimension, and multiply the first processed component after direction modulation and the second processed component after direction modulation by the corresponding node's correction gate weight and complementary weight value, respectively, to generate a first weighted result and a second weighted result; the complementary weight value is determined according to the numerical range of the correction gate weight by a preset interval mapping rule, the interval mapping rule being: when the correction gate weight increases, the complementary weight value decreases, and when the correction gate weight decreases, the complementary weight value increases. S64. Perform a weighted summation on the first weighted result and the second weighted result at the node dimension to generate node convolution results that correspond one-to-one with each sensing node. S65. Combine the convolution results of each node according to the arrangement order of the sensing nodes along the length of the anti-collision strip to generate a convolution output sequence.

[0014] Optionally, S7 specifically includes: S71. Obtain the convolution output sequence and construct node time series data according to the sensor node number order and time order; S72. Perform sliding window segmentation on the node time series data within a preset time window to generate node subsequences corresponding to multiple consecutive time periods; S73. Calculate the change between convolution output values ​​at adjacent time points in each node subsequence, construct a node change direction sequence according to the sign of the change, and generate a node trend sequence based on the difference range between adjacent changes. S74. Based on the consistency of the change direction of the node trend sequence in multiple consecutive time periods, perform consistency marking processing on the corresponding node time periods to generate stable abnormal segments. S75. Based on the distribution of stable abnormal segments along the length of the anti-collision strip, perform spatial continuity filtering to generate a water immersion alarm sequence and output a water immersion alarm signal.

[0015] The beneficial effects of this invention are: This invention introduces a graph-based multi-sensor node collaborative modeling mechanism in the scenario of a vehicle gear shifter anti-collision strip. Addressing the limitations of traditional single-point or small-sensor immersion monitoring schemes, which struggle to characterize local water diffusion processes and are susceptible to transient interference, this invention first constructs an adjacency matrix and a direction weight matrix based on the physical arrangement of sensor nodes along the length of the anti-collision strip. This explicitly incorporates the spatial structure of the anti-collision strip and the direction of water propagation into the data processing flow. During the node feature construction stage, short-term changes, temperature compensation, and historical differences are integrated to generate a node stability factor, enabling the reliability differences of sensor nodes to be quantified within the model. Furthermore, an improved FAGCN model is used to perform structure-aware graph transformation and frequency domain decomposition, dividing the multi-node time-series signals into different categories reflecting continuous spatial changes and abrupt spatial changes. The frequency domain components are reordered based on the node arrangement order, thus avoiding the problem of insufficient characterization of water immersion propagation characteristics in traditional time-domain or single-frequency-domain analysis. Furthermore, this invention introduces a multi-stage gating weight correction mechanism within the model, combining node stability factors, relative changes between frequency domain components, and consistency constraints across time windows into the gating weight generation process. Combined with a dual-weight fusion strategy of directional modulation and complementary weight mapping, differentiated weighting of different frequency domain information in different spatial directions is achieved, effectively suppressing interference from unstable nodes and transient noise on the convolution output. In the alarm generation stage, joint screening of temporal trend consistency and spatial continuity is performed on the convolution output sequence to avoid false alarms triggered by isolated anomalies, thereby achieving stable identification and reliable alarm for the water immersion status of the anti-collision strip. Through the above technical solutions, this invention significantly improves the accuracy, stability, and practicality of vehicle gear shifter anti-collision strip water immersion alarms under complex environmental disturbance conditions. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is a schematic diagram of the structure of a vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural network proposed in this invention; Figure 2 This is a flowchart illustrating the process of a vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural networks proposed in this invention. Figure 3 This is a schematic diagram of the processing flow of perceptual spectrum transformation and frequency domain decomposition in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1-3 A vehicle gear shifter anti-collision strip water immersion warning system based on graph neural network, comprising: The sensor data acquisition module is used to collect time-series observation data from multiple sensor nodes arranged along the anti-collision strip of the vehicle gear shifter. The graph structure generation module is used to generate an adjacency matrix based on the physical installation location of the sensor nodes, and to generate a direction weight matrix based on the arrangement order of the sensor nodes along the length of the anti-collision strip. The node feature and stability construction module is used to construct node feature sequences and generate node stability factors; The frequency domain decomposition module is used to perform structure-aware map transformation on the node feature sequences in the improved FAGCN model to generate the first frequency domain component and the second frequency domain component. The gating weight correction module generates corrected gating weights based on the node stability factor and the first and second frequency domain components. The directional modulation module performs directional weighting and synthesis processing on the first frequency domain component and the second frequency domain component based on the directional weight matrix; The node weighted calculation module performs node-level weighted synthesis processing on the direction-modulated processing components based on the corrected gating weights to generate a convolutional output sequence. The consistency screening and alarm output module generates a water immersion alarm sequence based on the convolution output sequence and outputs a water immersion alarm signal.

[0020] In this embodiment, the modules are interconnected using the following method: S1. Construct a graph structure of multiple sensor nodes arranged along the anti-collision strip of the vehicle gear shifter, generate an adjacency matrix, and generate an orientation weight matrix based on the spatial arrangement order of the nodes on the anti-collision strip. S2. Construct node feature sequences based on time-series observation data of sensor nodes, and calculate node stability factors based on short-term node changes, temperature compensation, and historical difference components. S3. Input the node feature sequence into the improved FAGCN model, perform spectral transformation processing, and generate the first frequency domain component and the second frequency domain component; S4. Based on the node stability factor, perform a correction process on the gating weights of the improved FAGCN model to generate corrected gating weights; S5. Perform directional modulation processing on the first frequency domain component and the second frequency domain component based on the directional weight matrix; S6. Based on the modified gating weights, perform weighted calculations on the first and second processed components after direction modulation in the improved FAGCN model to generate a convolutional output sequence; S7. Based on the time variation law of the convolution output sequence, perform consistency screening processing to generate a water immersion alarm sequence and output a water immersion alarm signal.

[0021] For applications where vehicle gear shift bumper strips exhibit a linear or near-linear arrangement along their length and the water immersion state gradually spreads along the structural direction, this invention structurally improves the FAGCN model. It introduces a directional weight matrix to characterize the forward and backward relationships between nodes, enabling the graph convolution process to distinguish between different propagation directions. Simultaneously, a gated weight correction mechanism based on node stability factors is introduced during frequency domain fusion, allowing the weights of different nodes participating in frequency domain component fusion to dynamically change with node observation stability within different time windows. Furthermore, this invention reorders and adaptively partitions spectral components in the frequency domain processing, combining the physical arrangement order of nodes to maintain a consistent correspondence between frequency domain components and the spatial structure of the bumper strip. Through these improvements, the improved FAGCN model can simultaneously characterize the structural directionality between nodes, differences in node reliability, and the spatial propagation characteristics of the water immersion state while retaining the original frequency domain adaptive graph convolution framework. This makes it more suitable for multi-sensor node collaborative modeling in vehicle gear shift bumper bumper water immersion alarm scenarios.

[0022] In this embodiment, S1 specifically includes: S11. Obtain the physical installation location data of multiple sensor nodes arranged along the anti-collision strip of the vehicle gear shifter; S12. Perform node connection relationship determination processing based on the physical installation distance between two adjacent sensing nodes. When the physical installation distance between adjacent nodes is within the preset distance threshold range, the adjacent nodes are defined as having a graph connection relationship, and the connection relationship is recorded in the adjacency matrix. S13. Generate a node arrangement sequence based on the physical installation order of all sensor nodes, and generate a direction index sequence based on the arrangement order of adjacent nodes in the anti-collision strip length direction in the node arrangement sequence. S14. Perform direction weight calculation processing based on the direction index sequence. For each pair of adjacent sensing nodes in the arrangement sequence, assign corresponding direction weight values ​​according to the order of arrangement of the nodes on the anti-collision strip, and record the direction weight values ​​in the direction weight matrix. S15. Construct the graph structure data for the input of the improved FAGCN model based on the adjacency matrix and the orientation weight matrix. The graph structure data includes the connection relationship data between nodes and the orientation weight data between nodes.

[0023] In this embodiment, S2 specifically includes: S21. Collect raw observation data of each sensor node arranged along the anti-collision strip of the vehicle gear shifter in a continuous time period, and perform time alignment processing on the raw observation data according to a uniform sampling time interval. S22. Based on the time-aligned original observation data, construct a node feature sequence according to the sensor node numbering order. Each time point in the node feature sequence corresponds to a node observation vector. S23. Perform short-time change calculation processing on the observation data of each sensing node within a preset time window, calculate the difference between the observation values ​​of adjacent sampling points within the time window and accumulate it. S24. Acquire the ambient temperature data corresponding to each sensing node, and perform temperature compensation processing on the node observation data based on the ambient temperature data and preset temperature compensation parameters to generate the temperature compensation amount. S25. Perform historical difference processing on the node observation vector of each sensing node in adjacent time windows, calculate the difference between the node observation vector of the current time window and the node observation vector of the previous time window, and generate historical difference components. S26. Based on the short-term change, the temperature compensation, and the historical difference, perform stability calculation processing on each sensing node to generate a corresponding node stability factor, and establish a correspondence between the node stability factor and the node feature sequence.

[0024] This invention constructs a stability factor reflecting the stability of node observations by jointly calculating short-term variations, temperature compensation, and historical differences in node observation data, thus ensuring comparability between different nodes within the same time window. This stability factor is not a single statistic but comprehensively characterizes the short-term fluctuation amplitude of nodes, environmental temperature disturbances, and trends across time windows. This effectively avoids the amplified impact of a single outlier sampling point on subsequent graph convolution operations, thereby providing a reliable data foundation for the gating weight correction in the improved FAGCN model.

[0025] In this embodiment, S3 specifically includes: S31. In the improved FAGCN model, based on the adjacency matrix and the orientation weight matrix, the node feature sequence is subjected to joint normalization processing to generate normalized graph structure data containing node connection relationships and node orientation relationships. S32. Based on the normalized graph structure data and node feature sequence, perform structure-aware graph spectral transformation processing, and perform spectral domain mapping on the node feature sequence while simultaneously constraining the node connection relationship and the node arrangement direction relationship. S33. Based on the arrangement order of nodes in the anti-collision strip length direction in the spectral domain mapping result, perform spectral component reordering processing to generate a spectral component sequence that is consistent with the spatial arrangement order of nodes; S34. Based on the change amplitude of adjacent spectral components in the spectral component sequence, perform adaptive frequency domain partitioning processing to divide the spectral components into a first spectral component set and a second spectral component set. S35. Generate a first frequency domain component based on the first spectral component set, and generate a second frequency domain component based on the second spectral component set, and establish a correspondence between the first frequency domain component and the second frequency domain component and the node feature sequence respectively.

[0026] In the anti-collision strip structure, the sensing nodes are continuously arranged along the length, and the node observation data exhibits a significant order correlation in the spatial dimension. This invention introduces node arrangement direction constraints during the spectral transformation process, ensuring that the spectral domain mapping results maintain a spatial structural expression consistent with the physical installation order. By reordering the spectral components and adaptively partitioning them based on the variation amplitude of adjacent spectral components, it can distinguish between spatially continuous change patterns and spatially abrupt change patterns, avoiding over-smoothing or anomaly masking problems caused by using fixed frequency thresholds. This structure-aware frequency domain decomposition method allows the improved FAGCN model to better reflect the actual characteristics of water diffusion along the structure of the anti-collision strip while maintaining the original spectral processing framework, enhancing its ability to distinguish between local anomalies and gradual changes.

[0027] In this embodiment, S4 specifically includes: S41. In the improved FAGCN model, a corresponding initial gating weight is generated for each sensing node. The initial gating weight is determined based on the node feature sequence during the model training phase. S42. Based on the node stability factor, perform a first-stage correction process on the initial gating weights, and perform segmented proportional adjustment on the initial gating weights according to the numerical range corresponding to the node stability factor. S43. Based on the first frequency domain component and the second frequency domain component, calculate the relative change relationship between the two types of frequency domain components at the node dimension, and perform a second-stage correction process on the gate weights after the first-stage correction based on the relative change relationship. S44. Perform cross-window consistency constraint processing on the gating weight sequence corresponding to the same sensing node within adjacent time windows, and perform threshold pruning processing on the gating weight change amplitude within adjacent time windows. S45. The gate weights after segmentation ratio adjustment, frequency domain relationship correction and cross-window consistency constraint processing are determined as the modified gate weights, and participate in the weighted calculation of the first frequency domain component and the second frequency domain component in the improved FAGCN model.

[0028] In the scenario of detecting water immersion of crash barriers, the observation data of sensor nodes are simultaneously affected by the node's own stability, spatial diffusion state, and temporal evolution. This invention introduces a multi-stage correction mechanism during the generation of gating weights. First, based on node stability, the gating weights are adjusted in segments to keep highly volatile nodes under control in the weight distribution. Then, considering the relative changes between the first and second frequency domain components, the gating weights are corrected using frequency domain correlation, causing the gating weights to change with the degree of spatial abrupt change. Finally, cross-window consistency constraints are applied to the gating weight sequence in the temporal dimension to prevent abnormal weight jumps caused by instantaneous fluctuations.

[0029] In this embodiment, S5 specifically includes: S51. Based on the direction weight matrix, according to the arrangement order of the sensor nodes along the length direction of the anti-collision strip, determine the set of forward adjacent nodes and the set of backward adjacent nodes along the length direction for each sensor node. S52. For the first frequency domain component, traverse each sensing node in the node dimension, multiply the first frequency domain component corresponding to the forward adjacent node set by the first direction weight value, and multiply the first frequency domain component corresponding to the backward adjacent node set by the second direction weight value to generate the first direction weighted result. S53. For the second frequency domain component, traverse each sensing node in the node dimension, multiply the corresponding second frequency domain component in the forward adjacent node set by the third direction weight value, and multiply the corresponding second frequency domain component in the backward adjacent node set by the fourth direction weight value to generate the second direction weighted result. S54. Perform weighted synthesis processing on the first frequency domain component based on the first directional weighting result, and perform weighted synthesis processing on the second frequency domain component based on the second directional weighting result to generate the first processed component after directional modulation and the second processed component after directional modulation. S55. Store the first processed component after direction modulation and the second processed component after direction modulation as intermediate data in the improved FAGCN model.

[0030] In the anti-collision strip structure of a vehicle gear shifter, sensing nodes are arranged linearly or nearly linearly along the length direction, and the water propagation along the anti-collision strip exhibits a clear propagation sequence characteristic during immersion. This invention introduces a directional weight matrix in the frequency domain processing stage, mapping the forward and backward adjacency relationships of nodes in the physical structure to numerical weights in different directions, which are then applied to different frequency domain components, maintaining an asymmetric expression of frequency domain features in spatial direction. By performing direction-dependent weighted synthesis processing on the frequency domain components of adjacent nodes, the inverse noise is avoided from being treated as equivalent to the forward change, thus introducing a direction-sensitive feature modulation mechanism without altering the graph structure. This processing method differs from traditional undirected graph convolution or symmetric weighting methods, making the model more closely reflect the actual characteristics of water diffusion along the structure of the anti-collision strip.

[0031] In this embodiment, S6 specifically includes: S61. Obtain the first processed component and the second processed component after direction modulation, and construct the corresponding node processing sequence according to the sensor node numbering order. S62. Obtain the corrected gating weights and construct the corresponding gating weight sequence according to the sensor node number order; S63. Traverse each sensing node along the node dimension, and multiply the first processed component after direction modulation and the second processed component after direction modulation by the corresponding node's correction gate weight and complementary weight value, respectively, to generate a first weighted result and a second weighted result; the complementary weight value is determined according to the numerical range of the correction gate weight by a preset interval mapping rule, the interval mapping rule being: when the correction gate weight increases, the complementary weight value decreases, and when the correction gate weight decreases, the complementary weight value increases. S64. Perform a weighted summation on the first weighted result and the second weighted result at the node dimension to generate node convolution results that correspond one-to-one with each sensing node. S65. Combine the convolution results of each node according to the arrangement order of the sensing nodes along the length of the anti-collision strip to generate a convolution output sequence.

[0032] During the water immersion detection of bumper strips, low-frequency changes and high-frequency abrupt changes often coexist at the same sensing node, and their evolution with spatial direction and time exhibits different patterns. This invention introduces a dual-channel weighting structure based on modified gating weights and complementary weights in the convolution output generation stage, enabling the first and second processing components to form a controlled combination relationship at the node level. By performing dual-weighting on each node and synthesizing the results at the node dimension, the invention avoids a single frequency domain component dominating the convolution result under local anomalies or background fluctuations. This node-level weighted synthesis method differs from the direct superposition or uniform weighting methods in traditional graph convolution, allowing the convolution output to simultaneously reflect the spatial direction modulation result and the gating weight correction result, thereby forming a node convolution representation that matches the water diffusion characteristics of the bumper strip.

[0033] In this embodiment, S7 specifically includes: S71. Obtain the convolution output sequence and construct node time series data according to the sensor node number order and time order; S72. Perform sliding window segmentation on the node time series data within a preset time window to generate node subsequences corresponding to multiple consecutive time periods; S73. Calculate the change between convolution output values ​​at adjacent time points in each node subsequence, construct a node change direction sequence according to the sign of the change, and generate a node trend sequence based on the difference range between adjacent changes. S74. Based on the consistency of the change direction of the node trend sequence in multiple consecutive time periods, perform consistency marking processing on the corresponding node time periods to generate stable abnormal segments. S75. Based on the distribution of stable abnormal segments along the length of the anti-collision strip, perform spatial continuity filtering to generate a water immersion alarm sequence and output a water immersion alarm signal.

[0034] During the water immersion detection of vehicle bumper strips, abnormal outputs at a single time point are easily affected by transient noise or sensor jitter. This invention introduces a dual consistency screening mechanism based on temporal trend and spatial distribution at the alarm generation stage. By constructing a sequence of changing directions in the temporal dimension through node convolution output and performing continuity judgment, isolated anomalies caused by short-term fluctuations are eliminated. Simultaneously, combining the node arrangement relationship along the length of the bumper strip, spatial continuity screening is performed on anomaly segments with consistent time, ensuring that the alarm results simultaneously meet the requirements of temporal continuity and spatial adjacency. This screening structure differs from conventional methods based on single thresholds or single node judgments, structuring the alarm judgment process into a time- and space-coupled processing flow, thereby improving the stability and reliability of water immersion alarms.

[0035] Example 1: To verify the feasibility and technical effectiveness of this invention in a real vehicle environment, the vehicle gear shift anti-collision strip water immersion warning system based on the improved FAGCN model described in this invention was applied to a real-vehicle test scenario of a passenger car model. The vehicle model is a front-engine, automatic transmission structure, with the gear shift area located in the center of the front passenger compartment. An anti-collision strip extending along the length of the gear shift is installed on the outer side of the gear shift, serving as a buffer and protection against minor collisions or vibrations. Because this anti-collision strip is structurally long and narrow, and easily becomes the primary path for water immersion inside the vehicle during heavy rain or flooded roads, this location was chosen as the typical and representative target for water immersion monitoring.

[0036] In this implementation scenario, 12 sensor nodes were evenly arranged along the length of the crash barrier. Each sensor node could collect electrical response signals related to the water immersion state and continuously output time-series observation data with a uniform sampling period of 100 milliseconds. During the test, the vehicle experienced various states, including driving in a dry environment, slow seepage of slight water accumulation, local water diffusion through the crash barrier, and complete immersion, accompanied by changes in ambient temperature and vehicle idling vibration, to simulate complex working conditions under real road conditions. The collected multi-node time-series data was first time-aligned in the onboard computing unit to ensure that the data from each node participated in subsequent calculations under a unified time reference.

[0037] In practical applications, the system of this invention first constructs the node arrangement relationship based on the physical installation position of each sensing node on the anti-collision strip, and generates an adjacency matrix and a direction weight matrix, so that the connection relationship between adjacent nodes and the forward and backward relationships along the length of the anti-collision strip can be clearly expressed. Subsequently, the system processes the observation data of each node within a continuous time window, calculating short-term changes, generating a temperature compensation amount based on environmental temperature data collected by an external temperature sensor, and obtaining historical difference components through the observation vector difference between adjacent time windows. These three components are used together to generate a node stability factor. This stability factor is used in subsequent model processing to reflect the differences in data reliability of different nodes under the current environment, thereby avoiding adverse effects on the overall judgment due to jitter or noise amplification of individual nodes.

[0038] After constructing node features and stability, the system inputs the node feature sequences into the improved FAGCN model. During processing, the model simultaneously constrains node connectivity and orientation relationships, performs structure-aware spectral transformation on the node feature sequences, and reorders and adaptively partitions the results in the spectral domain, thereby generating a first frequency domain component reflecting continuous spatial variation and a second frequency domain component reflecting abrupt spatial changes. Combined with a node stability factor, the system performs multi-stage corrections to the model's internal gating weights, ensuring that stable nodes receive higher weights during frequency domain fusion, while effectively suppressing the influence of unstable nodes. Simultaneously, the system applies directional modulation to both types of frequency domain components based on the directional weight matrix, ensuring that changes along the water propagation direction of the crash barrier are strongly represented in the model output.

[0039] In the convolutional output generation stage, the system further introduces a node-level weighting mechanism with modified gating weights and complementary weights. This mechanism weights and synthesizes the two types of processed components after direction modulation to generate a convolutional output sequence that corresponds one-to-one with each node. Finally, the system performs sliding window segmentation and trend analysis on the convolutional output sequence in the time dimension, and performs consistency screening based on the spatial distribution of nodes along the length of the crash barrier. Only when the abnormal changes are continuous in time and exhibit adjacent diffusion characteristics in space will a water immersion alarm sequence be generated and an alarm signal be output.

[0040] To verify the effectiveness of the method of the present invention, the system of the present invention was compared and analyzed with the traditional immersion alarm scheme based on single threshold determination and the graph neural network scheme without introducing node stability and orientation modulation. The results are shown in Table 1: Table 1. Performance Comparison of Different Water Immersion Alarm Solutions in Real-Vehicle Testing

[0041] As can be seen from the comparison results in Table 1 above, the performance differences of different water immersion alarm schemes in complex vehicle environments are quite significant. Traditional single-threshold judgment schemes, relying solely on instantaneous observations, are prone to false alarms under conditions of vehicle vibration and ambient temperature fluctuations, averaging 6.8 false alarms per hour, and the alarm duration is short, failing to reflect the continuity of the water immersion process. Conventional graph neural network schemes, after introducing inter-node correlation modeling, improve both the number of false alarms and recognition accuracy, but in scenarios where water spreads along the direction of the bumper strip, the characterization of the propagation trend is still insufficient. In contrast, the scheme of this invention, through node stability constraints, direction modulation, and temporal and spatial consistency screening, effectively reduces the frequency of false alarms, controlling the average number of false alarms to 0.9 per hour, while significantly improving the lead time for water immersion recognition and the duration of alarm stability, achieving a comprehensive recognition accuracy of 96.7%. This indicates that the present invention can more reliably reflect the water immersion status of the bumper strip and provide stable alarms in practical applications.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A vehicle gear shifter anti-collision strip water immersion warning system based on graph neural network, characterized in that, include: The sensor data acquisition module is used to collect time-series observation data from multiple sensor nodes arranged along the anti-collision strip of the vehicle gear shifter. The graph structure generation module is used to generate an adjacency matrix based on the physical installation location of the sensor nodes, and to generate a direction weight matrix based on the arrangement order of the sensor nodes along the length of the anti-collision strip. The node feature and stability construction module is used to construct node feature sequences and generate node stability factors; The frequency domain decomposition module is used to perform structure-aware map transformation on the node feature sequences in the improved FAGCN model to generate the first frequency domain component and the second frequency domain component. The gating weight correction module generates corrected gating weights based on the node stability factor and the first and second frequency domain components. The directional modulation module performs directional weighting and synthesis processing on the first frequency domain component and the second frequency domain component based on the directional weight matrix; The node weighted calculation module performs node-level weighted synthesis processing on the direction-modulated processing components based on the corrected gating weights to generate a convolutional output sequence. The consistency screening and alarm output module generates a water immersion alarm sequence based on the convolution output sequence and outputs a water immersion alarm signal.

2. The vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural network according to claim 1, characterized in that, The modules are connected in the following way: S1. Construct a graph structure of multiple sensor nodes arranged along the anti-collision strip of the vehicle gear shifter, generate an adjacency matrix, and generate an orientation weight matrix based on the spatial arrangement order of the nodes on the anti-collision strip. S2. Construct node feature sequences based on time-series observation data of sensor nodes, and calculate node stability factors based on short-term node changes, temperature compensation, and historical difference components. S3. Input the node feature sequence into the improved FAGCN model, perform spectral transformation processing, and generate the first frequency domain component and the second frequency domain component; S4. Based on the node stability factor, perform a correction process on the gating weights of the improved FAGCN model to generate corrected gating weights; S5. Perform directional modulation processing on the first frequency domain component and the second frequency domain component based on the directional weight matrix; S6. Based on the modified gating weights, perform weighted calculations on the first and second processed components after direction modulation in the improved FAGCN model to generate a convolutional output sequence; S7. Based on the time variation law of the convolution output sequence, perform consistency screening processing to generate a water immersion alarm sequence and output a water immersion alarm signal.

3. A vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural network according to claim 2, characterized in that, S1 specifically includes: S11. Obtain the physical installation location data of multiple sensor nodes arranged along the anti-collision strip of the vehicle gear shifter; S12. Perform node connection relationship determination processing based on the physical installation distance between two adjacent sensing nodes. When the physical installation distance between adjacent nodes is within the preset distance threshold range, the adjacent nodes are defined as having a graph connection relationship, and the connection relationship is recorded in the adjacency matrix. S13. Generate a node arrangement sequence based on the physical installation order of all sensor nodes, and generate a direction index sequence based on the arrangement order of adjacent nodes in the anti-collision strip length direction in the node arrangement sequence. S14. Perform direction weight calculation processing based on the direction index sequence. For each pair of adjacent sensing nodes in the arrangement sequence, assign corresponding direction weight values ​​according to the order of arrangement of the nodes on the anti-collision strip, and record the direction weight values ​​in the direction weight matrix. S15. Construct the graph structure data for the input of the improved FAGCN model based on the adjacency matrix and the orientation weight matrix. The graph structure data includes the connection relationship data between nodes and the orientation weight data between nodes.

4. A vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural network according to claim 2, characterized in that, S2 specifically includes: S21. Collect raw observation data of each sensor node arranged along the anti-collision strip of the vehicle gear shifter in a continuous time period, and perform time alignment processing on the raw observation data according to a uniform sampling time interval. S22. Based on the time-aligned original observation data, construct a node feature sequence according to the sensor node numbering order. Each time point in the node feature sequence corresponds to a node observation vector. S23. Perform short-time change calculation processing on the observation data of each sensing node within a preset time window, calculate the difference between the observation values ​​of adjacent sampling points within the time window and accumulate it. S24. Acquire the ambient temperature data corresponding to each sensing node, and perform temperature compensation processing on the node observation data based on the ambient temperature data and preset temperature compensation parameters to generate the temperature compensation amount. S25. Perform historical difference processing on the node observation vector of each sensing node in adjacent time windows, calculate the difference between the node observation vector of the current time window and the node observation vector of the previous time window, and generate historical difference components. S26. Based on the short-term change, the temperature compensation, and the historical difference, perform stability calculation processing on each sensing node to generate a corresponding node stability factor, and establish a correspondence between the node stability factor and the node feature sequence.

5. A vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural network according to claim 2, characterized in that, S3 specifically includes: S31. In the improved FAGCN model, based on the adjacency matrix and the orientation weight matrix, the node feature sequence is subjected to joint normalization processing to generate normalized graph structure data containing node connection relationships and node orientation relationships. S32. Based on the normalized graph structure data and node feature sequence, perform structure-aware graph spectral transformation processing, and perform spectral domain mapping on the node feature sequence while simultaneously constraining the node connection relationship and the node arrangement direction relationship. S33. Based on the arrangement order of nodes in the anti-collision strip length direction in the spectral domain mapping result, perform spectral component reordering processing to generate a spectral component sequence that is consistent with the spatial arrangement order of nodes; S34. Based on the change amplitude of adjacent spectral components in the spectral component sequence, perform adaptive frequency domain partitioning processing to divide the spectral components into a first spectral component set and a second spectral component set. S35. Generate a first frequency domain component based on the first spectral component set, and generate a second frequency domain component based on the second spectral component set, and establish a correspondence between the first frequency domain component and the second frequency domain component and the node feature sequence respectively.

6. A vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural network according to claim 2, characterized in that, S4 specifically includes: S41. In the improved FAGCN model, a corresponding initial gating weight is generated for each sensing node. The initial gating weight is determined based on the node feature sequence during the model training phase. S42. Based on the node stability factor, perform a first-stage correction process on the initial gating weights, and perform segmented proportional adjustment on the initial gating weights according to the numerical range corresponding to the node stability factor. S43. Based on the first frequency domain component and the second frequency domain component, calculate the relative change relationship between the two types of frequency domain components at the node dimension, and perform a second-stage correction process on the gate weights after the first-stage correction based on the relative change relationship. S44. Perform cross-window consistency constraint processing on the gating weight sequence corresponding to the same sensing node within adjacent time windows, and perform threshold pruning processing on the gating weight change amplitude within adjacent time windows. S45. The gate weights after segmentation ratio adjustment, frequency domain relationship correction and cross-window consistency constraint processing are determined as the modified gate weights, and participate in the weighted calculation of the first frequency domain component and the second frequency domain component in the improved FAGCN model.

7. A vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural network according to claim 2, characterized in that, S5 specifically includes: S51. Based on the direction weight matrix, according to the arrangement order of the sensor nodes along the length direction of the anti-collision strip, determine the set of forward adjacent nodes and the set of backward adjacent nodes along the length direction for each sensor node. S52. For the first frequency domain component, traverse each sensing node in the node dimension, multiply the first frequency domain component corresponding to the forward adjacent node set by the first direction weight value, and multiply the first frequency domain component corresponding to the backward adjacent node set by the second direction weight value to generate the first direction weighted result. S53. For the second frequency domain component, traverse each sensing node in the node dimension, multiply the corresponding second frequency domain component in the forward adjacent node set by the third direction weight value, and multiply the corresponding second frequency domain component in the backward adjacent node set by the fourth direction weight value to generate the second direction weighted result. S54. Perform weighted synthesis processing on the first frequency domain component based on the first directional weighting result, and perform weighted synthesis processing on the second frequency domain component based on the second directional weighting result to generate the first processed component after directional modulation and the second processed component after directional modulation. S55. Store the first processed component after direction modulation and the second processed component after direction modulation as intermediate data in the improved FAGCN model.

8. A vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural network according to claim 2, characterized in that, S6 specifically includes: S61. Obtain the first processed component and the second processed component after direction modulation, and construct the corresponding node processing sequence according to the sensor node numbering order. S62. Obtain the corrected gating weights and construct the corresponding gating weight sequence according to the sensor node number order; S63. Traverse each sensing node along the node dimension, and multiply the first processed component after direction modulation and the second processed component after direction modulation by the corresponding node's correction gate weight and complementary weight value, respectively, to generate a first weighted result and a second weighted result; the complementary weight value is determined according to the numerical range of the correction gate weight by a preset interval mapping rule, the interval mapping rule being: when the correction gate weight increases, the complementary weight value decreases, and when the correction gate weight decreases, the complementary weight value increases. S64. Perform a weighted summation on the first weighted result and the second weighted result at the node dimension to generate node convolution results that correspond one-to-one with each sensing node. S65. Combine the convolution results of each node according to the arrangement order of the sensing nodes along the length of the anti-collision strip to generate a convolution output sequence.

9. A vehicle gear shifter anti-collision strip water immersion alarm system based on graph neural network according to claim 2, characterized in that, Specifically, S7 includes: S71. Obtain the convolution output sequence and construct node time series data according to the sensor node number order and time order; S72. Perform sliding window segmentation on the node time series data within a preset time window to generate node subsequences corresponding to multiple consecutive time periods; S73. Calculate the change between convolution output values ​​at adjacent time points in each node subsequence, construct a node change direction sequence according to the sign of the change, and generate a node trend sequence based on the difference range between adjacent changes. S74. Based on the consistency of the change direction of the node trend sequence in multiple consecutive time periods, perform consistency marking processing on the corresponding node time periods to generate stable outlier segments. S75. Based on the distribution of stable abnormal segments along the length of the anti-collision strip, perform spatial continuity filtering to generate a water immersion alarm sequence and output a water immersion alarm signal.

Images