Hazardous chemical substance transportation early warning system based on fuzzy neural network

Abstract

The invention discloses a hazardous chemical substance transportation early warning system based on a fuzzy neural network, and relates to the technical field of traffic transportation and road safety monitoring. According to the hazardous chemical substance transportation early warning system based on the fuzzy neural network, road geometric influence in original data of a sensor is accurately stripped from dynamic response of a vehicle; according to the method, false risk signals caused by road design can be eliminated, the method focuses on instability risks really caused by dynamic factors such as vehicle control and load transfer, and the rollover risk can be judged more essentially and more accurately; converting the preprocessed characteristic quantity into a fuzzy linguistic variable, reasoning by simulating a fuzzy rule of human expert experience, and finally outputting continuous risk indexes and levels; the processing mode not only better conforms to the imprecise characteristic of risk assessment, but also enhances the adaptability and robustness of the system to different vehicle types, loads and road conditions.

Description

Technical Field

[0001] This invention relates to the field of transportation and road safety monitoring technology, and in particular to a hazardous chemical transportation early warning system based on fuzzy neural networks. Background Technology

[0002] In the field of hazardous chemical transportation safety monitoring, bridge sections are high-risk areas for accidents due to their special structure. Current early warning systems generally use computer vision technology to monitor transport vehicles in real time through video analysis combined with multi-sensor fusion. Such systems rely on deep learning algorithms to identify vehicle shapes and combine inertial measurement unit data to judge changes in vehicle attitude, thereby achieving early warning of dangerous conditions such as rollover.

[0003] Existing technologies primarily determine rollover risk by setting fixed thresholds; when the vehicle tilt angle exceeds a predetermined value, an alarm mechanism is triggered, and parameters such as vehicle speed and acceleration are used for auxiliary judgment; some systems also introduce multi-sensor data fusion technology to obtain richer vehicle status information through radar, laser ranging and other devices to improve recognition accuracy.

[0004] However, in actual bridge applications, the normal tilting of vehicles when driving on curves has similar characteristics to the signs of rollover, leading to false alarms from the system. Different bridges have significant differences in geometric features such as radius of curvature and slope, making it difficult for fixed thresholds to adapt to varying road conditions. In addition, changes in vehicle load distribution and uneven road surfaces can also affect the accuracy of attitude judgment, and the ability of existing solutions to distinguish these interference factors needs to be improved. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a hazardous chemical transportation early warning system based on fuzzy neural networks to solve the problem of high false alarm rate in existing hazardous chemical vehicle rollover early warning systems in bridge curve scenarios, due to the inability to distinguish between road geometric inclination and vehicle instability.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides a hazardous chemical transportation early warning system based on a fuzzy neural network, comprising:

[0009] The data acquisition module is configured to synchronously acquire multi-source data from the transport vehicle, including at least vehicle attitude data, road environment data, and vehicle operating status data.

[0010] The data preprocessing and fusion module is configured to perform time alignment, outlier removal, missing measurement imputation and feature normalization on the multi-source data, and to perform fusion estimation and geometric compensation on attitude and kinematic quantities.

[0011] The fuzzy neural network processing module, including a fuzzification layer, a rule reasoning layer, and a defuzzification layer, is configured to input preprocessed features into the fuzzy neural network to output a rollover risk index and risk level.

[0012] The early warning decision and human-computer interaction module is configured to generate graded early warnings based on the risk level and duration strategy, and to provide vehicle-side prompts and external communication.

[0013] The data storage and adaptive update module is configured to store historical data and early warning records, and to adaptively update the membership parameters and fuzzy rules based on the historical data.

[0014] As a preferred embodiment of the hazardous chemical transportation early warning system based on fuzzy neural network described in this invention, the vehicle attitude data includes roll angle, pitch angle, lateral acceleration and yaw rate.

[0015] The road environment data includes road curvature, longitudinal slope, and road surface adhesion.

[0016] The vehicle operating status data includes vehicle speed, steering angle, and load distribution information.

[0017] As a preferred embodiment of the hazardous chemical transportation early warning system based on fuzzy neural network described in this invention, the data preprocessing and fusion module is configured to perform superelevation and longitudinal slope compensation on the roll angle, and estimate the geometric components of lateral acceleration based on vehicle speed and road curvature.

[0018] The steps for estimating and separating the superelevation and longitudinal slope compensation of the roll angle and the geometric components of the lateral acceleration include:

[0019] Using the road's lateral superelevation angle and longitudinal slope angle as geometric references, the roll and pitch of the vehicle-mounted IMU are aligned with the road coordinate system to form reference attitude quantities for subsequent compensation and separation.

[0020] By separating vehicle body mounting errors from terrain effects under road reference, the road alignment roll angle used for risk assessment is obtained:

[0021] ,

[0022] in, This indicates the road alignment with the roll angle, measured in rad. This indicates the roll angle measured by the IMU, in rad. This indicates the road's superelevation angle, measured in rad. This indicates the pitch angle measured by the IMU, in rad. Indicates the longitudinal slope angle of the road, in rad;

[0023] The projection of gravity onto the vehicle's lateral axis is removed from the measured lateral acceleration to obtain the gravity-free lateral acceleration:

[0024] ,

[0025] in, This represents the compensated lateral acceleration, in units of... , This indicates the lateral acceleration measured by the IMU, in units of... , Represents gravitational acceleration, with units of . ;

[0026] Geometric roll drive intensity is characterized by road curvature and vehicle speed, and is expressed as:

[0027] ,

[0028] in, Represents geometric lateral acceleration, in units of , This indicates vehicle speed, measured in m / s. Represents road curvature, in units of ;

[0029] The compensated lateral acceleration is separated from the geometric component to obtain the dynamically dominant lateral component:

[0030] ,

[0031] in, Represents the lateral acceleration component of the dynamics, in units of ;

[0032] An estimate of the equilibrium superelevation angle is provided at the road design reference speed for consistency verification and compensation reference:

[0033] ,

[0034] in, This indicates the equilibrium superelevation angle under the reference operating conditions, in rad. This indicates the road design reference speed, expressed in m / s.

[0035] Four types of quantities are provided to the blurring layer:

[0036] Attitude geometry alignment Path geometric strength Exercise intensity and the dominant dynamic tilt force The remaining intermediate quantities are used only for internal compensation and verification.

[0037] As a preferred embodiment of the hazardous chemical transportation early warning system based on fuzzy neural networks described in this invention, the fuzzification layer divides roll angle, road curvature, and vehicle speed into at least three membership degree sets.

[0038] The membership function is a triangular or trapezoidal function, and its center and width parameters can be adjusted online by the adaptive update module.

[0039] As a preferred embodiment of the hazardous chemical transportation early warning system based on fuzzy neural networks described in this invention, the rule reasoning layer includes multiple antecedent fuzzy rules in the form of if-then, where the antecedents include at least a set of roll angles, a set of road curvatures, and a set of vehicle speeds, and the consequent is a fuzzy set of rollover risk indicators.

[0040] The deblurring layer uses the centroid method to output continuous risk indicators and maps them to risk levels based on preset grading thresholds.

[0041] As a preferred embodiment of the hazardous chemical transportation early warning system based on fuzzy neural network described in this invention, the data preprocessing and fusion module includes a time synchronization unit, which is configured to align multi-sensor data based on timestamp interpolation and drift correction, and to denoise and estimate attitude and acceleration through a filter.

[0042] As a preferred embodiment of the hazardous chemical transportation early warning system based on fuzzy neural networks described in this invention, the adaptive update adopts a phased strategy:

[0043] Without affecting online inference latency, the membership degree and rule parameters are periodically updated using historical samples and early warning feedback;

[0044] Perform quick fine-tuning and set parameter boundaries when drift detection is triggered.

[0045] As a preferred embodiment of the hazardous chemical transportation early warning system based on fuzzy neural networks described in this invention, the early warning decision and human-computer interaction module is equipped with a tiered early warning and hysteresis mechanism.

[0046] High risk triggers an audible and visual alarm and maintains it for at least one preset time window; medium risk triggers a voice prompt; low risk does not trigger an alarm.

[0047] When risk indicators fluctuate around the grading threshold, hysteresis bands are used to suppress frequent switching.

[0048] As a preferred embodiment of the hazardous chemical transportation early warning system based on fuzzy neural network described in this invention, the external communication includes sending the early warning information to the monitoring center and emergency platform through the vehicle-mounted communication link, and displaying the risk level, key features and suggested operations on the vehicle-mounted display terminal.

[0049] As a preferred embodiment of the hazardous chemical transportation early warning system based on fuzzy neural network described in this invention, the data preprocessing and fusion module uses a sliding window for online processing, and the window length and update step size are dynamically adjusted according to vehicle speed and road curvature.

[0050] The window length and update step size are dynamically adjusted based on vehicle speed and road curvature as follows:

[0051] Establish a coupling condition between processing time and statistical window, using the upper limit of end-to-end latency as a constraint:

[0052] ,

[0053] in, This indicates that for a length of The computational delay for processing one window, measured in seconds. This indicates the update step size, in seconds. This represents the upper limit of the target delay, in seconds. Indicates the number of samples within the window; dimensionless.

[0054] Approximately:

[0055] ,

[0056] in, This represents fixed costs, expressed in seconds (s). This represents the processing cost per sample, expressed in seconds per sample.

[0057] Define lateral demand metrics for adjusting the window:

[0058] ,

[0059] in, This represents a lateral demand indicator, in units of... , This indicates vehicle speed, expressed in m / s. Represents road curvature, in units of ;

[0060] The time windows are given based on the principle that high cornering speeds shorten the window and low straight-line speeds lengthen the window:

[0061] ,

[0062] in, This represents the unsmoothed time window, in seconds. This indicates an operator that truncates based on upper and lower limits. and These represent the lower and upper limits of the time window, respectively, in seconds. and This represents the adjustment coefficient, with units of s and... , This indicates a small quantity to prevent the denominator from being zero, and the unit is 1. ;

[0063] The update step size is set using a fixed overlap rate, and the result is verified using a delay constraint.

[0064] ,

[0065] in, This indicates the update step size, in seconds. This indicates the effective time window, in seconds. This represents the scaling factor between the step size and the window size; it is dimensionless. , These represent its lower and upper bounds, respectively, and are dimensionless.

[0066] right Perform first-order smoothing and add relative threshold hysteresis:

[0067] ,

[0068] like Then keep ,

[0069] in, Indicates the first The time window after the update takes effect is in seconds. Represents the smoothing coefficient, dimensionless. This represents the hysteresis threshold, which is dimensionless.

[0070] Map the time quantity to the sample quantity using the sampling rate, and set a lower stability limit:

[0071] ,

[0072] ,

[0073] in, This represents the number of samples corresponding to the step size, and is dimensionless. This represents the sampling rate, measured in Hz. The minimum sample size required to represent statistical stability; dimensionless.

[0074] If found Then adjust them according to priority:

[0075] Prioritize increasing But not exceeding If the condition is still not met, then the value should be reduced accordingly. or and not less than and constraint.

[0076] The beneficial effects of this invention are as follows: This invention effectively solves the fundamental problem of high false alarm rates in traditional methods through data preprocessing and fusion mechanisms; it does not simply rely on absolute values ​​of vehicle attitude, but precisely separates the road geometric influences (such as superelevation, longitudinal slope, and curve curvature) from the vehicle's own dynamic response in the raw sensor data. This eliminates false risk signals caused by road design itself, focusing on the instability risks truly caused by dynamic factors such as vehicle handling and load transfer, thus achieving a more fundamental and accurate judgment of rollover hazards.

[0077] The application of fuzzy neural networks in this invention endows the system with a powerful ability to handle uncertainties and nonlinear problems. It transforms preprocessed features into fuzzy linguistic variables, and uses fuzzy rules simulating human expert experience for reasoning, ultimately outputting continuous risk indicators and levels. This approach not only better reflects the inherent imprecision of risk assessment but also enhances the system's adaptability and robustness to different vehicle types, loads, and road conditions.

[0078] The system provided by this invention possesses excellent engineering practicality and adaptability. The dynamic window adjustment mechanism ensures an optimal balance between real-time processing and data statistical stability, enabling rapid response to sudden risks while maintaining the confidence level of judgments. Simultaneously, the self-learning update function of parameters and rules allows the system to continuously optimize performance and adapt to changes in vehicle status and the external environment over long-term use. The tiered early warning and hysteresis strategy provides a clear and user-friendly human-machine interaction experience, promptly alerting to high risks while avoiding frequent switching near thresholds that could interfere with the driver. This invention constructs a complete early warning closed loop, from data acquisition, intelligent processing, risk decision-making to information output and remote communication, providing comprehensive and intelligent technical support for the safety of hazardous chemical transportation and significantly improving the industry's safety management level. Attached Figure Description

[0079] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0080] Figure 1 This is a schematic diagram of the framework of the hazardous chemical transportation early warning system based on fuzzy neural networks in the embodiment. Detailed Implementation

[0081] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0082] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0083] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0084] This application proposes a hazardous chemical transportation early warning system based on fuzzy neural networks, combined with... Figure 1 As shown, the system includes:

[0085] The data acquisition module is configured to synchronously collect multi-source data from transport vehicles, including at least vehicle attitude data, road environment data, and vehicle operating status data.

[0086] In this embodiment, the acquisition criteria for vehicle attitude data, road environment data, and vehicle operating status data are as follows: attitude data includes roll angle, pitch angle, lateral acceleration, and yaw rate output by the onboard inertial measurement unit; road curvature, longitudinal slope, and superelevation angle in the road environment data are obtained from a consistent source, either from a map or online estimation, ensuring coordinate system consistency; vehicle speed, steering angle, and load distribution information in the vehicle operating status data are provided by the vehicle bus or equivalent sensor input. Specifically, the default sampling rate... The frequency is set to 100Hz, and can be tuned within the range of 50-200Hz, determined by a trade-off between the vibration environment at the installation point and the capabilities of the computing platform. The timestamp resolution is no less than 1ms to ensure subsequent time synchronization accuracy. Optionally, road curvature and longitudinal slope are estimated online by vehicle kinematics when no map is available, keeping the input dimensions of the road environment data items unchanged. When load distribution measurements are missing, the most recent valid value is kept within 10 seconds; if the interval exceeds the limit, the field is marked as unknown and continues to be output downstream, while other fields are processed normally.

[0087] The data preprocessing and fusion module is configured to perform time alignment, outlier removal, missing data imputation and feature normalization on multi-source data, and to perform fusion estimation and geometric compensation of attitude and kinematic quantities.

[0088] For example, time alignment uses a unified time axis and interpolation alignment, prioritizing the vehicle time as the master clock. If stable drift exists, linear drift correction is performed. Outlier removal uses a combination of two-sided limiting and median consistency criteria. Missing data imputation uses a time limit and marks data as unreliable if it exceeds an acceptable duration. The default interpolation step size is the same as the sampling period. The outlier limiting threshold is set to ±5 times the median absolute deviation of the historical moving median of the corresponding signal. The maximum imputation duration is 0.5s, which can be tuned within 0.2-1.0s. Optionally, feature normalization uses two commonly used methods: zero mean unit variance or interval scaling, to maintain the stability of the input scale of subsequent fuzzification layers. When any key field has no valid value and continues to exceed the set duration, the data quality status is simultaneously output downstream as degraded, so that the decision-making end can apply a more conservative threshold strategy.

[0089] The fuzzy neural network processing module, including a fuzzification layer, a rule reasoning layer, and a defuzzification layer, is configured to input preprocessed features into the fuzzy neural network to output a rollover risk index and risk level.

[0090] The early warning decision and human-computer interaction module is configured to generate graded early warnings based on risk level and duration strategies, and to provide vehicle-side prompts and external communication.

[0091] The data storage and adaptive update module is configured to store historical data and early warning records, and adaptively update the membership parameters and fuzzy rules based on the historical data.

[0092] In one embodiment, vehicle attitude data includes roll angle, pitch angle, lateral acceleration, and yaw rate.

[0093] Road environmental data includes road curvature, longitudinal slope, and pavement adhesion.

[0094] Vehicle operating status data includes vehicle speed, steering angle, and load distribution information;

[0095] In one embodiment, the data preprocessing and fusion module is configured to perform superelevation and longitudinal slope compensation on the roll angle and estimate the geometric components of lateral acceleration based on vehicle speed and road curvature, thereby distinguishing between geometric roll and dynamic roll.

[0096] The steps for estimating and separating the geometric components of roll angle superelevation, longitudinal slope compensation, and lateral acceleration include:

[0097] Using the road's lateral superelevation angle and longitudinal slope angle as geometric references, the roll and pitch of the vehicle-mounted IMU are aligned with the road coordinate system to form reference attitude quantities for subsequent compensation and separation.

[0098] By separating vehicle body mounting errors from terrain effects under road reference, the road alignment roll angle used for risk assessment is obtained:

[0099] ,

[0100] in, This indicates the road alignment with the roll angle, measured in rad. This indicates the roll angle measured by the IMU, in rad. This indicates the road's superelevation angle, measured in rad. This indicates the pitch angle measured by the IMU, in rad. Indicates the longitudinal slope angle of the road, in rad;

[0101] The projection of gravity onto the vehicle's lateral axis is removed from the measured lateral acceleration to obtain the gravity-free lateral acceleration:

[0102] ,

[0103] in, This represents the compensated lateral acceleration, in units of... , This indicates the lateral acceleration measured by the IMU, in units of... , Represents gravitational acceleration, with units of . ;

[0104] Geometric roll drive intensity is characterized by road curvature and vehicle speed, and is expressed as:

[0105] ,

[0106] in, Represents geometric lateral acceleration, in units of , This indicates vehicle speed, measured in m / s. Represents road curvature, in units of ;

[0107] The compensated lateral acceleration is separated from the geometric component to obtain the dynamically dominant lateral component:

[0108] ,

[0109] in, Represents the lateral acceleration component of the dynamics, in units of ;

[0110] An estimate of the equilibrium superelevation angle is provided at the road design reference speed for consistency verification and compensation reference:

[0111] ,

[0112] in, This indicates the equilibrium superelevation angle under the reference operating conditions, in rad. This indicates the road design reference speed, expressed in m / s.

[0113] Four types of quantities are provided to the blurring layer:

[0114] Attitude geometry alignment Path geometric strength Exercise intensity and the dominant dynamic tilt force The remaining intermediate quantities are used only for internal compensation and verification.

[0115] Furthermore, , , and Before entering the fuzzification layer, the output is uniformly based on the dimensions and symbols of the road coordinate system to ensure comparability between different road segments. By default, detrending and amplitude limiting are performed on the above four types of quantities. The amplitude limiting range is determined by expanding the historical 95% confidence range by 10%. The update cycle is consistent with the data frame. , and The source remains consistent with the aforementioned, and its nominal range is based on common highway design values, with specific values ​​provided by the data acquisition module. Optionally, when When short-term fluctuations are large, robust statistical values ​​from the most recent samples are used to replace instantaneous values ​​to suppress jitter. If any quantity is determined to be unreliable, the previous reliable output is retained in the current frame along with an input degradation flag.

[0116] Specifically, a geometric benchmark is established using superelevation and longitudinal slope. Roll angle is then compensated under this benchmark to obtain an alignment roll quantity that can be compared across road sections and installation conditions. Attitude-gravity decoupling of lateral acceleration eliminates spurious signals caused by gravity projection, making the lateral quantity closer to the tire's lateral action. Geometric lateral acceleration, given by vehicle speed and curvature, reflects the centripetal demand caused by the curve. This is then subtracted from the compensated measurement to obtain the dominant dynamic component. This component is more closely related to load transfer, center of gravity height, and tire saturation, making it suitable for inclusion in fuzzy rules. Combining the equilibrium superelevation angle given by the road design benchmark provides a consistent reference when data reliability and map errors exist, improving compensation robustness. The four inputs entering the fuzzification layer each perform their respective functions: attitude alignment describes geometric roll, curvature and vehicle speed describe path and motion intensity, and the dynamic component focuses on controllable risk sources, laying the foundation for continuous risk indicators output by the rule layer and defuzzification layer.

[0117] In one embodiment, the fuzzification layer divides roll angle, road curvature, and vehicle speed into at least three membership degree sets.

[0118] The membership function is a triangular or trapezoidal function, and its center and width parameters can be adjusted online by the adaptive update module;

[0119] In this embodiment, at least three levels are configured by default as three-level divisions, which can be expanded to five levels to improve resolution when needed. The initial center and width are determined by the quantiles and dispersion of the most recent historical data. The initial center is set to the 25%, 50%, and 75% quantiles, and the width is set to 0.8 times the corresponding quantile interval, which can be tuned within the range of 0.5-1.5 times. The minimum step size and update cycle are adjusted online in accordance with the adaptive update strategy. Optionally, the triangular and trapezoidal functions can be interchanged under the principle of equivalent coverage without changing the output dimension. When online updates are unavailable or limited, the most recent stable parameter set is maintained, and the parameter freeze state is recorded.

[0120] In one embodiment, the rule reasoning layer includes multiple antecedent fuzzy rules of the if-then form, where the antecedents include at least a set of roll angles, a set of road curvatures, and a set of vehicle speeds, and the consequent is a fuzzy set of rollover risk indicators.

[0121] The deblurring layer uses the centroid method to output continuous risk indicators and maps them to risk levels based on preset grading thresholds;

[0122] Similarly, continuous risk indicators undergo monotonicity and boundary checks before being mapped to levels to ensure the predictability and consistency of level switching. The default grading thresholds are set to initial values ​​with equal intervals for the three levels, and are fine-tuned during the adaptive update phase based on a trade-off between historical false positives and false negatives; the time window and hysteresis width are kept consistent with the early warning decision module. An optional implementation is to use an equal-frequency or equal-risk-density threshold arrangement instead of the equal-interval method without changing the number of output levels. When continuous indicators are unavailable, the most recent level output is retained, and a conservative mode flag is triggered to indicate upper-level delay tolerance.

[0123] In one embodiment, the data preprocessing and fusion module includes a time synchronization unit configured to align multi-sensor data based on timestamp interpolation and drift correction, and to denoise and estimate attitude and acceleration using a filter.

[0124] Optionally, the time synchronization interpolation strategy prioritizes zero-order hold, switching to linear interpolation to reduce phase distortion when the interval between adjacent samples exceeds twice the sampling period. The default maximum allowed time drift is ±5ms, with a drift correction period of 1s. The default cutoff frequency for attitude and acceleration filtering is set based on the natural frequencies of the vehicle suspension and chassis, with a 20% safety margin. When timestamp jitter exceeds the allowable range or there are jumps, the affected frames are marked as not participating in fusion, and normal processing is resumed in the subsequent frame.

[0125] In one embodiment, adaptive updates employ a phased strategy:

[0126] Without affecting online inference latency, the membership degree and rule parameters are periodically updated using historical samples and early warning feedback;

[0127] When drift detection is triggered, a quick fine-tuning is performed, and parameter boundaries are set to ensure system stability;

[0128] Specifically, periodic updates are based on sample statistics within a rolling time window and feedback labels indicating warnings, maintaining an update step size smaller than the stability boundary. Drift detection is triggered by statistical indicators of input distribution or output deviation. The default periodic update cycle is 60 seconds, which can be tuned within 30-300 seconds; rapid fine-tuning lasts no more than 5 seconds and adheres to parameter boundaries to avoid over-adjustment. When feedback labels are unavailable, small adjustments to the center and width are performed only based on unsupervised input distribution drift. When multiple consecutive updates fail to converge or the trigger frequency is too high, parameter freezing is automatically initiated, and the system reverts to the most recently stable parameter set.

[0129] In one embodiment, the early warning decision-making and human-computer interaction module is equipped with a tiered early warning and hysteresis mechanism:

[0130] High risk triggers an audible and visual alarm and maintains it for at least one preset time window; medium risk triggers a voice prompt; low risk does not trigger an alarm.

[0131] When risk indicators fluctuate around the grading threshold, hysteresis band is used to suppress frequent switching.

[0132] Furthermore, the tiered early warning system applies different uplink and downlink trigger boundaries for the entry and exit of continuous risk indicators to ensure the necessary stability of the alarm. The default preset time window is 1.0s, which can be adjusted within 0.5-2.0s; the default relative threshold for the hysteresis band is 10%, which can be adjusted within 5%-20%. Optionally, the shortest silence interval for suppressing repeated alerts during periods of sustained high risk remains consistent with the preset time window to ensure that alerts are not too frequent. When the data quality status is degraded, a more conservative decision-making approach is adopted, appropriately expanding the hysteresis band and extending the retention time.

[0133] In one embodiment, external communication includes sending early warning information to the monitoring center and emergency platform via the vehicle communication link, and displaying the risk level, key features and suggested actions on the vehicle display terminal, while also supporting event playback;

[0134] In this embodiment, the fields carried by external communication include risk level, continuous risk indicators, key input quantities, and data quality status, enabling the receiving end to reproduce the decision context. The default upload period is 1 second in non-high-risk states and 0.2 seconds in high-risk states; the default window for event playback is approximately 30 seconds, which can be adjusted within 10-120 seconds. When communication is unstable, local buffering is switched and uploads are delayed, while maintaining the field structure. When communication interruption exceeds a preset duration, complete prompts and records are maintained only on the vehicle end, and playback segments are re-uploaded in batches after uplink recovery.

[0135] In one embodiment, the data preprocessing and fusion module uses a sliding window for online processing. The window length and update step size are dynamically adjusted according to vehicle speed and road curvature to maintain sufficient statistical stability while ensuring real-time performance.

[0136] For example, the linkage between the sliding window and the update step size follows a fixed overlap rate principle, ensuring that the algorithm's end-to-end latency is controlled and adapts to the vehicle's dynamic intensity. Default end-to-end latency upper limit. The default timeout is 200ms, which can be tuned within 100-300ms; the default overlap rate is... Take 0.5, corresponding to and The values ​​are set to 0.33 and 0.67 respectively. Optionally, under steady-state conditions with low vehicle speed and small curvature, the overlap ratio can be moderately reduced to improve throughput. In abnormal situations, such as a short-term increase in computational load, the overlap ratio should be increased first. Up to no more than The level is set to meet the delay constraint.

[0137] The window length and update step size are dynamically adjusted based on vehicle speed and road curvature as follows:

[0138] Establish a coupling condition between processing time and statistical window, using the upper limit of end-to-end latency as a constraint:

[0139] ,

[0140] in, This indicates that for a length of The computational delay for processing one window, measured in seconds. This indicates the update step size, in seconds. This represents the upper limit of the target delay, in seconds. Indicates the number of samples within the window; dimensionless.

[0141] To be implemented online, Approximately:

[0142] ,

[0143] in, This represents fixed costs, expressed in seconds (s). This represents the processing cost per sample, expressed in seconds per sample.

[0144] Define lateral demand metrics for adjusting the window:

[0145] ,

[0146] in, This represents a lateral demand indicator, in units of... , This indicates vehicle speed, expressed in m / s. Represents road curvature, in units of ;

[0147] The time windows are given based on the principle that high cornering speeds shorten the window and low straight-line speeds lengthen the window:

[0148] ,

[0149] in, This represents the unsmoothed time window, in seconds. This indicates an operator that truncates based on upper and lower limits. and These represent the lower and upper limits of the time window, respectively, in seconds. and This represents the adjustment coefficient, with units of s and... , This indicates a small quantity to prevent the denominator from being zero, and the unit is 1. ;

[0150] The update step size is set using a fixed overlap rate, and the result is verified using a delay constraint.

[0151] ,

[0152] in, This indicates the update step size, in seconds. This indicates the effective time window, in seconds. This represents the scaling factor between the step size and the window size; it is dimensionless. , These represent its lower and upper bounds, respectively, and are dimensionless.

[0153] right Perform first-order smoothing and add relative threshold hysteresis:

[0154] ,

[0155] like Then keep ,

[0156] in, Indicates the first The time window after the update takes effect is in seconds. Represents the smoothing coefficient, dimensionless. This represents the hysteresis threshold, which is dimensionless.

[0157] Map the time quantity to the sample quantity using the sampling rate, and set a lower stability limit:

[0158] ,

[0159] ,

[0160] in, This represents the number of samples corresponding to the step size, and is dimensionless. This represents the sampling rate, measured in Hz. The minimum sample size required to represent statistical stability; dimensionless.

[0161] If found Then adjust them according to priority:

[0162] Prioritize increasing But not exceeding If the condition is still not met, then the value should be reduced accordingly. or and not less than and Constraints; if the above are still not satisfied, then set the delayed constraint not met state bit for processing by the upper-level strategy;

[0163] Optionally, the state bit indicating that the delay constraint was not met is maintained for at least one update cycle after being cleared, to avoid frequent entry into and exit from this state. (Default) Take 0.2s, Set the time to 2.0s; and The initial values ​​are set based on the historical straight-gradient-sharp-curve distribution, so that... Under typical straight road and sharp curve conditions, it falls near the upper and lower quartiles, respectively. and The values ​​are provided by offline benchmark tests on the deployment platform and are only updated after the software version or hardware platform changes, once the values ​​have stabilized. Approaching If the delay constraint still cannot be met, the statistical stability of the explicit reporting is limited and the minimum executable caliber is maintained until the constraint is lifted.

[0164] Specifically, the window and step size settings are transformed into optimization criteria under the constraint of delay upper limit, so as to establish a direct link between the time model and the overlap rate, which is convenient for reuse on different hardware platforms; the lateral demand index reflects the combined effect of curvature and speed, automatically tightening the time window when the curve is sharp or the speed is high, thereby improving the response to rapid changes in operating conditions, and lengthening the time window when the curve is straight or the speed is low to improve statistical stability; upper and lower limit truncation avoids fluctuations caused by extreme values, and the step size is obtained with a fixed ratio to keep the overlap rate of the sliding window within a controllable range; further, first-order smoothing and relative hysteresis are introduced to suppress parameter jitter near the threshold and reduce computational jitter and warning jumps caused by frequent reconfiguration; sample domain mapping and lower limit of sample number ensure minimum statistics and avoid insufficient confidence.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0166] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A hazardous chemical transportation early warning system based on fuzzy neural networks, characterized in that, include: The data acquisition module is configured to synchronously acquire multi-source data from the transport vehicle, including at least vehicle attitude data, road environment data, and vehicle operating status data. The data preprocessing and fusion module is configured to perform time alignment, outlier removal, missing measurement imputation and feature normalization on the multi-source data, and to perform fusion estimation and geometric compensation on attitude and kinematic quantities. The fuzzy neural network processing module, including a fuzzification layer, a rule reasoning layer, and a defuzzification layer, is configured to input preprocessed features into the fuzzy neural network to output a rollover risk index and risk level. The early warning decision and human-computer interaction module is configured to generate graded early warnings based on the risk level and duration strategy, and to provide vehicle-side prompts and external communication. The data storage and adaptive update module is configured to store historical data and early warning records, and to adaptively update the membership parameters and fuzzy rules based on the historical data.

2. The hazardous chemical transportation early warning system based on fuzzy neural networks as described in claim 1, characterized in that, The vehicle attitude data includes roll angle, pitch angle, lateral acceleration, and yaw rate; The road environment data includes road curvature, longitudinal slope, and road surface adhesion. The vehicle operating status data includes vehicle speed, steering angle, and load distribution information.

3. The hazardous chemical transportation early warning system based on fuzzy neural networks as described in claim 1, characterized in that, The data preprocessing and fusion module is configured to perform superelevation and longitudinal slope compensation on the roll angle, and estimate the geometric components of lateral acceleration based on vehicle speed and road curvature. The steps for estimating and separating the superelevation and longitudinal slope compensation of the roll angle and the geometric components of the lateral acceleration include: Using the road's lateral superelevation angle and longitudinal slope angle as geometric references, the roll and pitch of the vehicle-mounted IMU are aligned with the road coordinate system to form reference attitude quantities for subsequent compensation and separation. By separating vehicle body mounting errors from terrain effects under road reference, the road alignment roll angle used for risk assessment is obtained: ， in, This indicates the road alignment with the roll angle, measured in rad. This indicates the roll angle measured by the IMU, in rad. This indicates the road's superelevation angle, measured in rad. This indicates the pitch angle measured by the IMU, in rad. Indicates the longitudinal slope angle of the road, in rad; The projection of gravity onto the vehicle's lateral axis is removed from the measured lateral acceleration to obtain the gravity-free lateral acceleration: ， in, This represents the compensated lateral acceleration, in units of... , This indicates the lateral acceleration measured by the IMU, in units of... , Represents gravitational acceleration, with units of . ; Geometric roll drive intensity is characterized by road curvature and vehicle speed, and is expressed as: ， in, Represents geometric lateral acceleration, in units of , This indicates vehicle speed, measured in m / s. Represents road curvature, in units of ; The compensated lateral acceleration is separated from the geometric component to obtain the dynamically dominant lateral component: ， in, Represents the lateral acceleration component of the dynamics, in units of ; An estimate of the equilibrium superelevation angle is provided at the road design reference speed for consistency verification and compensation reference: ， in, This indicates the equilibrium superelevation angle under the reference operating conditions, in rad. This indicates the road design reference speed, expressed in m / s. Four types of quantities are provided to the blurring layer: Attitude geometry alignment Path geometric strength Exercise intensity and the dominant dynamic tilt force The remaining intermediate quantities are used only for internal compensation and verification.

4. The hazardous chemical transportation early warning system based on fuzzy neural networks as described in claim 1, characterized in that, The fuzzification layer divides roll angle, road curvature, and vehicle speed into at least three membership degree sets. The membership function is a triangular or trapezoidal function, and its center and width parameters can be adjusted online by the adaptive update module.

5. The hazardous chemical transportation early warning system based on a fuzzy neural network as described in claim 1, characterized in that, The rule reasoning layer contains multiple antecedent fuzzy rules in the form of if-then rules. The antecedents include at least a set of roll angles, a set of road curvatures, and a set of vehicle speeds, while the consequent is a fuzzy set of rollover risk indicators. The deblurring layer uses the centroid method to output continuous risk indicators and maps them to risk levels based on preset grading thresholds.

6. The hazardous chemical transportation early warning system based on fuzzy neural network as described in claim 1, characterized in that, The data preprocessing and fusion module includes a time synchronization unit, which is configured to align multi-sensor data based on timestamp interpolation and drift correction, and to perform denoising and estimation of attitude and acceleration through filters.

7. The hazardous chemical transportation early warning system based on fuzzy neural network as described in claim 1, characterized in that, The adaptive update employs a phased strategy: Without affecting online inference latency, the membership degree and rule parameters are periodically updated using historical samples and early warning feedback; Perform quick fine-tuning and set parameter boundaries when drift detection is triggered.

8. The hazardous chemical transportation early warning system based on fuzzy neural network as described in claim 1, characterized in that, The early warning decision-making and human-computer interaction module is equipped with a tiered early warning and hysteresis mechanism: High risk triggers an audible and visual alarm and maintains it for at least one preset time window; medium risk triggers a voice prompt; low risk does not trigger an alarm. When risk indicators fluctuate around the grading threshold, hysteresis bands are used to suppress frequent switching.

9. A hazardous chemical transportation early warning system based on a fuzzy neural network as described in claim 1, characterized in that, The external communication includes sending early warning information to the monitoring center and emergency platform through the vehicle-mounted communication link, and displaying the risk level, key characteristics and suggested actions on the vehicle-mounted display terminal.

10. A hazardous chemical transportation early warning system based on a fuzzy neural network as described in claim 3, characterized in that, The data preprocessing and fusion module uses a sliding window for online processing, and the window length and update step size are dynamically adjusted according to vehicle speed and road curvature. The window length and update step size are dynamically adjusted based on vehicle speed and road curvature as follows: Establish a coupling condition between processing time and statistical window, using the upper limit of end-to-end latency as a constraint: ， in, This indicates that for a length of The computational delay for processing one window, measured in seconds. This indicates the update step size, in seconds. This represents the upper limit of the target delay, in seconds. Indicates the number of samples within the window; dimensionless. Approximately: ， in, This represents fixed costs, expressed in seconds (s). This represents the processing cost per sample, expressed in seconds per sample. Define lateral demand metrics for adjusting the window: ， in, This represents a lateral demand indicator, in units of... , This indicates vehicle speed, expressed in m / s. Represents road curvature, in units of ; The time windows are given based on the principle that high cornering speeds shorten the window and low straight-line speeds lengthen the window: ， in, This represents the unsmoothed time window, in seconds. This indicates an operator that truncates based on upper and lower limits. and These represent the lower and upper limits of the time window, respectively, in seconds. and This represents the adjustment coefficient, with units of s and... , This indicates a small quantity to prevent the denominator from being zero, and the unit is 1. ; The update step size is set using a fixed overlap rate, and the result is verified using a delay constraint. ， ， in, This indicates the update step size, in seconds. This indicates the effective time window, in seconds. This represents the scaling factor between the step size and the window size; it is dimensionless. , These represent its lower and upper bounds, respectively, and are dimensionless. right Perform first-order smoothing and add relative threshold hysteresis: ， like Then keep , in, Indicates the first The time window after the update takes effect is in seconds. Represents the smoothing coefficient, dimensionless. This represents the hysteresis threshold, which is dimensionless. Map the time quantity to the sample quantity using the sampling rate, and set a lower stability limit: ， ， in, This represents the number of samples corresponding to the step size, and is dimensionless. This represents the sampling rate, measured in Hz. The minimum sample size required to represent statistical stability; dimensionless. If found Then adjust them according to priority: Prioritize increasing But not exceeding If the condition is still not met, then the value should be reduced accordingly. or and not less than and constraint.

Images