Electric bicycle safety risk dynamic assessment method and system

Abstract

The invention belongs to the field of edge intelligence and safety supervision, and discloses an electric bicycle safety risk dynamic assessment method and system, and the method comprises the steps: obtaining a global risk feature set, constructing a risk assessment matrix, dynamically adjusting a first recognition threshold value and a second recognition threshold value of key risk features, and calculating an early risk probability to judge an early risk level; extracting global temperature time sequence data, generating a temperature change curve, combining an early risk level, constructing a risk evolution model, outputting a risk evolution probability sequence in future hours, calculating a risk evolution rate, dynamically adjusting a risk judgment threshold value, and generating a dynamic risk evolution curve; the risk state feature vector is obtained, the first matching degree is calculated in combination with the current risk level, the matching confidence coefficient is determined, matching judgment is conducted, if not, the risk prompt level is adjusted, the early risk level and the dynamic risk evolution curve are updated, the timeliness of electric bicycle safety risk prevention and control is improved, and the false alarm rate is reduced.

Description

Technical Field

[0001] This invention relates to a method and system for dynamic assessment of safety risks of electric bicycles, belonging to the field of edge intelligence and safety supervision technology. Background Technology

[0002] With the rapid advancement of technology, electric bicycles have become a mainstream mode of transportation, but significant safety hazards remain. During parking and charging, electric bicycles are prone to battery thermal runaway due to malfunctions or accidents such as damaged internal battery separator dendrites, failure of the thermal management system, and overcharging or over-discharging. This can lead to safety accidents. Therefore, the existing regulatory model, which relies on manual inspections and post-incident handling, lacks the ability to predict the temporal evolution of safety risks and cannot accurately anticipate risk trends to effectively prevent such risks. Thus, it is necessary to prevent safety incidents from occurring at the source.

[0003] A Chinese patent application with publication number CN114997714A discloses a method, system, and computer device for risk identification of electric bicycle riders. This method establishes a risk dataset based on preset traffic violation information. Then, based on this risk dataset, a fuzzy clustering algorithm is used to analyze the traffic safety risks of electric bicycle riders, classifying them according to preset risk levels. Finally, based on rider attribute information and electric bicycle vehicle characteristic information, electric bicycle rider groups at each risk level are further divided into different risk categories. This method can accurately locate electric bicycle groups with different safety risk levels, enabling the development of different management strategies for each risk level, allowing for differentiated management of electric bicycle rider groups, and providing a basis for formulating electric bicycle governance measures.

[0004] Although there is an existing method, system, and computer equipment for identifying the risks of electric bicycle riders, which collects vehicle characteristic information of electric bicycles, preset road traffic violation information and attribute information of riders, establishes a risk dataset based on the preset road traffic violation information, uses fuzzy clustering algorithm to classify preset risk levels, and further subdivides risk types by combining rider attributes and vehicle characteristics, thus achieving accurate positioning and differentiated management of electric bicycle rider groups and providing a reliable basis for the formulation of electric bicycle governance measures, it has problems such as limited detection area, accuracy being easily affected by environmental interference, and risk identification being in the middle and late stages, resulting in an increased false alarm rate. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for dynamic assessment of safety risks of electric bicycles. This method involves collecting multi-source heterogeneous data and aligning it spatiotemporally at the edge; using adaptive RLS filtering for noise reduction and an attention mechanism to enhance early risk features; generating a global risk feature set; constructing a risk assessment matrix to dynamically adjust the first and second identification thresholds; calculating the probability of early risks to determine the early risk level; and constructing a risk evolution model to output future... The risk evolution probability sequence within an hour is then used to adjust the risk warning level through closed-loop matching calibration, thereby improving the accuracy of assessment and the reliability of prevention and control, reducing the false alarm rate, and ultimately ensuring the safety of electric bicycles.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The dynamic assessment method for safety risks of electric bicycles includes:

[0008] Collect battery status data, risk precursor data, global temperature distribution data and environmental interference data, align them spatiotemporally at the edge, increase the weight of risk precursor features, filter environmental noise signals, and generate a global risk feature set.

[0009] Construct a risk assessment matrix, dynamically adjust the first and second identification thresholds for key risk characteristics, and calculate the probability of early risks to determine the level of early risks;

[0010] Extract global temperature time-series data, generate temperature change curves, combine with the early risk levels to construct a risk evolution model, and output future risk data. Calculate the risk evolution probability sequence within an hour, calculate the risk evolution rate, dynamically adjust the risk judgment threshold, and generate a dynamic risk evolution curve;

[0011] Obtain the risk status feature vector, calculate the first matching degree in combination with the current risk level, determine the matching confidence, and make a matching judgment. If there is no match, adjust the risk warning level and update the early risk level and the dynamic risk evolution curve.

[0012] Specifically, the steps for generating a global risk feature set include:

[0013] Collect battery status data, early warning data, global temperature distribution data, and environmental interference data of electric bicycles;

[0014] For multi-source heterogeneous data at the edge, the time delay of different sensors is corrected, and after spatiotemporal alignment, a multidimensional fusion dataset is formed.

[0015] Dynamically adjust the core parameters of the adaptive RLS filtering algorithm and calculate the signal-to-noise ratio.

[0016] If the signal-to-noise ratio is less than the preset signal-to-noise ratio threshold, continue filtering to select valid signals; otherwise, stop filtering to form a denoised multidimensional fusion dataset.

[0017] An attention mechanism network model is used to input a denoised multidimensional fusion dataset, extract risk precursor features, construct an initial weight library, calculate and dynamically adjust the basic weights of each risk precursor feature, and output a weighted fusion dataset.

[0018] According to the coverage integrity verification rules, identify and supplement the central monitoring blind spots or missing data areas in the weighted fusion dataset.

[0019] Battery state risk features, temperature risk features, environmental associated risk features, and risk precursor features are extracted from the verified weighted fusion dataset to generate a global risk feature set.

[0020] Specifically, the steps for determining the early risk level include:

[0021] Extract key risk characteristics, including those inhibited by the environment and those stimulated by the environment;

[0022] Extract the critical values ​​of each key risk characteristic in the early stage of risk, revise and determine the benchmark thresholds of each key risk characteristic, and construct a risk assessment matrix.

[0023] For the environmentally inhibited features, an environmental interference factor is obtained, an amplification coefficient is introduced, a reference threshold for the concentration of the feature gas is calibrated, and a calibrated first identification threshold is generated.

[0024] For the environmental excitation characteristics, the local maximum heating rate is obtained, a sensitivity coefficient is introduced, a reference threshold for the local temperature is calibrated, and a calibrated second recognition threshold is generated.

[0025] The risk assessment matrix is ​​dynamically updated based on the first identification threshold and the second identification threshold.

[0026] Specifically, the steps for determining the early risk level also include:

[0027] Based on the dynamically updated risk assessment matrix, a first confidence level is calculated for the environmentally inhibited features based on the first identification threshold and combined with the real-time values ​​of the environmentally inhibited features;

[0028] For the environmentally stimulated feature, a second confidence level is calculated based on the second identification threshold and the real-time value of the environmentally stimulated feature;

[0029] A first weight and a second weight are set, and the early risk probability is calculated by combining the first confidence level and the second confidence level. ;

[0030] Set risk assessment threshold , To determine the early risk level;

[0031] like If so, it is classified as a Level 1 risk; if If so, it is judged as a level two risk; if If so, it is classified as a level three risk.

[0032] Specifically, the steps for generating the temperature change curve include:

[0033] Extract global temperature distribution data, label spatial coordinates, and use the K-Means clustering algorithm to remove outlier global temperature distribution data, retaining valid temperature data;

[0034] The valid temperature data are sorted according to timestamps to form continuous global temperature time series data;

[0035] The sliding window size is dynamically selected based on the aforementioned early risk level;

[0036] Level 1 risk adopts Minute window, Level 2 risk adopted Minute window, Level 3 risk adopted Minute window, and ;

[0037] For the global temperature time series data within each window, a smooth temperature value is generated by fitting using the Savitzky-Golay filtering algorithm, and a temperature change curve is generated by sliding point by point.

[0038] Specifically, outputting the future The specific steps for the risk evolution probability sequence within an hour include:

[0039] Differential features are extracted from the temperature change curve and concatenated with the smoothed temperature value to form a feature-enhanced time series sequence;

[0040] Construct a risk evolution model, including an input layer, a risk LSTM layer, a feature fusion layer, and an output layer;

[0041] The input layer is used to receive the feature-enhanced temporal sequence and the early risk level;

[0042] The risk LSTM layer is used to learn the evolutionary relationship between temperature changes and the early risk levels, before descending sorting. Each attention weight time step generates a risk temporal feature vector;

[0043] The feature fusion layer maps the early risk level to a temporal feature space and fuses it with the risk temporal feature vector to generate a comprehensive risk feature vector;

[0044] The output layer is used to output the future risk based on the comprehensive risk feature vector and the softmax activation function. Hourly risk probability distribution;

[0045] Obtain a historical risk case dataset, train the risk evolution model using the Adam optimizer, and output the future... Risk probability sequence within hours.

[0046] Specifically, the steps for generating a dynamic risk evolution curve include:

[0047] Based on the risk probability sequence, calculate the risk probability difference between adjacent time steps;

[0048] Set time step weights and calculate the risk evolution rate by weighting each time step. , shaping the future Risk evolution rate sequence within hours;

[0049] Set rate threshold , Dynamically adjust the risk assessment thresholds for each early risk level;

[0050] like If so, the risk assessment threshold for the early risk level will be raised, and the early fire risk level will be upgraded by one level.

[0051] like If so, the risk assessment threshold of the early risk level will be maintained, and a security risk warning will be triggered;

[0052] like If so, the risk assessment threshold for the early risk level remains unchanged, and monitoring continues;

[0053] By integrating the temperature change curve, the risk probability sequence, the dynamically adjusted risk judgment threshold, and the risk level transition point, a dynamic fire risk evolution curve is generated.

[0054] Specifically, the steps for determining the matching confidence level include:

[0055] Real-time acquisition of trigger status data, generation of raw feature data stream, smoothing of risk warning intensity signal in the raw feature data stream, and generation of risk status feature vector;

[0056] An edge computing algorithm is used to calculate the ambient noise intensity in real time and set the normal noise range.

[0057] If the ambient noise intensity exceeds the normal noise range, the generated data quality label is low quality; otherwise, the generated data quality label is high quality.

[0058] Extract the current risk level from the dynamic risk evolution curve to form a quantitative value of the current risk level, quantify the risk state feature vector into a risk state feature value, and calculate the first matching degree;

[0059] If the data quality label is low quality, then determine the confidence coefficient and calculate the matching confidence score in combination with the first matching score;

[0060] If the data quality label is high quality, then the first match score is directly used as the match confidence score.

[0061] Specifically, the steps for determining matching include:

[0062] Set a matching threshold to determine the matching accuracy.

[0063] If the matching confidence score is greater than or equal to the matching threshold, then a match is determined; otherwise, a mismatch is determined.

[0064] For mismatches, calculate the matching gap. Set the gap threshold , Adjust the priority determination;

[0065] like If it is, then it is determined to be the first priority; if If it is, then it is determined to be the second priority; if If so, it is determined to be the third priority;

[0066] Establish gap-level rules: for the first priority, adjust the risk warning intensity across levels; for the second priority, adjust the risk warning intensity level by level; for the third priority, adjust the trigger frequency while keeping the current risk warning mode unchanged.

[0067] Risk status data is fed back to the risk assessment matrix to update early risk levels and dynamic risk evolution curves;

[0068] Recalculate the matching confidence score; stop iterative calibration when the matching confidence score is greater than or equal to the matching threshold; when the matching confidence score is continuous... If the calibration fails to meet the requirements in the next iteration, a manual review prompt will be triggered.

[0069] The electric bicycle safety risk dynamic assessment system includes: risk perception module, risk assessment module, risk evolution module, and matching calibration module;

[0070] The risk perception module is used to collect battery status data, risk precursor data, global temperature distribution data and environmental interference data. After spatiotemporal alignment, the weight of risk precursor features is increased and environmental noise signals are filtered to generate a global risk feature set.

[0071] The risk assessment module is used to construct a risk assessment matrix, dynamically adjust the first and second identification thresholds of key risk characteristics, and calculate the early risk probability in real time to determine the early risk level.

[0072] The risk evolution module is used to extract global temperature time-series data, generate temperature change curves, combine them with the early risk levels, construct a risk evolution model, and output future risk data. Calculate the risk evolution probability sequence within an hour, calculate the risk evolution rate, dynamically adjust the risk judgment threshold, and generate a dynamic risk evolution curve;

[0073] The matching calibration module is used to acquire the risk state feature vector, calculate the first matching degree in combination with the current risk level, determine the matching confidence level, and make a matching judgment. If there is no match, the risk warning level is adjusted and the early risk level and the dynamic risk evolution curve are updated.

[0074] The beneficial effects of this invention are:

[0075] 1. This invention collects multi-source heterogeneous data on battery status, risk precursors, global temperature, and environmental interference, and performs spatiotemporal alignment at the edge. It employs an adaptive RLS filtering algorithm to filter noise and an attention mechanism to enhance the weight of risk precursor features, generating a global risk feature set. This solves the feature distortion problems caused by heterogeneous data and noise interference in traditional methods. A risk assessment matrix is ​​constructed, and for features inhibited and excited by the environment, the first and second identification thresholds are dynamically calibrated in conjunction with environmental interference factors and local heating rates. Then, the early risk probability is calculated by weighting to determine the early risk level. This avoids the poor adaptability of fixed thresholds, effectively reduces the false alarm rate, and achieves effective assessment and discrimination of early safety risks of electric bicycles in complex environments.

[0076] 2. This invention, based on global temperature time-series data, generates temperature change curves through dynamic sliding windows and Savitzky-Golay filtering, constructs a risk evolution model, and outputs future... The system calculates the risk evolution probability sequence within an hour, determines the risk evolution rate, dynamically adjusts the risk assessment thresholds for each early risk level, and generates a dynamic risk evolution curve. This addresses the lag issue of traditional static assessments and enables early prediction of safety risk trends. By calculating the first matching degree and matching confidence level, the system dynamically adjusts the risk warning level according to priority, updates the early risk level and the dynamic risk evolution curve, and avoids mismatches between risk warnings and actual risk levels. When the next iteration of calibration fails to meet the standard, a manual review prompt is triggered to ensure the timeliness of risk alerts and improve the timeliness and reliability of electric bicycle safety risk prevention and control. Attached Figure Description

[0077] Figure 1A schematic diagram of a dynamic risk assessment method for electric bicycles.

[0078] Figure 2 This is a flowchart for determining the early risk level in this invention;

[0079] Figure 3 This is a flowchart illustrating the generation of dynamic risk evolution curves in this invention;

[0080] Figure 4 This is a flowchart of the matching determination process in this invention;

[0081] Figure 5 This is a structural diagram of a dynamic safety risk assessment system for electric bicycles. Detailed Implementation

[0082] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0083] Example 1

[0084] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a dynamic assessment method for the safety risks of electric bicycles, including the following steps:

[0085] Battery status data and early warning data of electric bicycles are collected by vehicle-mounted multimodal sensors. At the same time, ground-mounted thermal infrared sensors are deployed to cover the bottom of the electric bicycle, the ground and the surrounding area to collect global temperature distribution data and environmental interference data. For multi-source heterogeneous data at the edge, a dynamic time warping algorithm is used to perform spatiotemporal alignment using timestamps. An attention mechanism is used to increase the weight of early warning features, such as increasing the weight of characteristic gas concentration. Meanwhile, an adaptive filtering algorithm is used to filter environmental noise signals to generate a global risk feature set that removes environmental noise and has no blind spots. Among them, battery status data includes voltage, current and internal resistance; early warning data includes electrolyte evaporation, characteristic gas concentration and battery pack sealing parameters; multi-source heterogeneous data includes battery status data, early warning data, global temperature distribution data and environmental interference data; early warning features include abnormal characteristic gas concentration; and environmental interference data includes high dust concentration.

[0086] Based on the global risk feature set, an adaptive environmental calibration risk assessment matrix is ​​constructed. According to the currently collected environmental interference data and global temperature distribution data, the identification threshold of key risk features is dynamically adjusted. In high dust environment, the first identification threshold of characteristic gas concentration is increased to suppress false alarms. When the ground-mounted thermal infrared sensor detects local temperature rise, the second identification threshold of local temperature is reduced to improve sensitivity. The risk assessment matrix is ​​used to calculate the early risk probability in real time to determine the early risk level.

[0087] Continuous global temperature time-series data is extracted from the global risk feature set. A sliding window smoothing algorithm is used to generate temperature change curves. Combined with early risk levels, a risk evolution model is constructed. Finally, an LSTM time-series prediction algorithm is used to output future... Calculate the risk evolution rate from the hourly risk evolution probability sequence. The risk assessment threshold for early risk levels is dynamically adjusted based on the rate of risk evolution, generating a real-time updated dynamic risk evolution curve.

[0088] The system collects trigger status data from the audio-visual warning device, such as the risk warning intensity. It uses a sliding window mid-range filter to generate a risk status feature vector. Through an edge computing algorithm, it calculates the first degree of matching between the risk status feature vector and the current risk level indicated by the dynamic risk evolution curve in real time. It then determines the matching between the risk warning intensity of the audio-visual warning device and the early risk level. If there is a mismatch, it immediately adjusts the risk warning level and feeds the risk status data back to the risk assessment matrix to re-determine the early risk level and update the dynamic risk evolution curve until the first degree of matching reaches the preset standard.

[0089] Specifically, the steps for generating a global risk feature set include:

[0090] The vehicle-mounted multimodal sensor collects battery status data and early warning data of electric bicycles. At the same time, the ground-mounted thermal infrared sensor covers the bottom of the electric bicycle, the ground and the surrounding area to collect temperature distribution data and environmental interference data.

[0091] For multi-source heterogeneous data at the edge, a dynamic time warping algorithm is used to correct the time delay of different sensors, and timestamps are used for spatiotemporal alignment to form a multidimensional fusion dataset.

[0092] An interference feature library was constructed based on historical environmental interference data and safety risk simulation experimental data. Environmental interference data was extracted from the multidimensional fusion dataset and template matched with the interference feature library to mark noise signal segments corresponding to high dust concentration and ground reflective heat, thereby determining the interference range of noise on risk precursor identification and temperature monitoring. Among these measures, a dust threshold was set based on the statistical correlation analysis results between dust concentration and sensor false alarm rate in historical environmental interference data. When the dust concentration in the environmental interference data exceeds the dust threshold, it is considered a high dust concentration.

[0093] To address the noise characteristics of different data types, the core parameters of the adaptive RLS filtering algorithm are dynamically adjusted. For example, for voltage, current, and internal resistance data, the forgetting factor is increased to filter out contact noise caused by high dust concentrations; for global temperature distribution data, a sliding window is used to smooth out instantaneous temperature peaks caused by ground reflection heat; for electrolyte evaporation and characteristic gas concentration data, the characteristic gas concentration in vehicle-free areas is used as a baseline, and differential correction is performed to eliminate baseline noise from ambient background gases. During the filtering process, the signal-to-noise ratio (SNR) is calculated in real time, and a SNR threshold is set according to sensor performance. If the SNR is less than the threshold, filtering continues, and valid signals are selected; otherwise, filtering stops to ensure that valid signals are not distorted. The resulting denoised battery status data, global temperature distribution data, and environmental interference data form a denoised multidimensional fusion dataset. The core parameters include, but are not limited to, the forgetting factor.

[0094] An attention-based network model is employed, using a denoised multidimensional fusion dataset as input. Risk precursor features are extracted from the denoised multidimensional fusion dataset. An initial weight library is constructed based on the statistical analysis results of historical risk cases and the experience of industry experts. The basic weights of each risk precursor feature are calculated, and the basic weights are dynamically adjusted based on real-time multi-source heterogeneous data. If the detected electrolyte evaporation exceeds the preset evaporation threshold, the basic weight of electrolyte evaporation is increased. If the detected abnormal concentration of characteristic gas is detected, the basic weight of characteristic gas concentration is increased. The result is a weighted fusion dataset with enhanced risk precursor features. The preset evaporation threshold is determined through statistical analysis of battery thermal runaway experimental data and the standard electrolyte evaporation rate specification.

[0095] According to the coverage integrity verification rules, the spatial dimension label of each data in the weighted fusion dataset is checked to identify monitoring blind spots or areas with missing data that are not covered by sensors. Gaussian interpolation is used to supplement the data to ensure that there are no blind spots in the monitoring of the electric bicycle parking area and its surrounding area. In this embodiment, the coverage integrity verification rules are determined by industry safety monitoring specifications, the results of sensor deployment optimization in experimental scenarios, and the priority judgment criteria of experts for key monitoring areas.

[0096] In the battery state risk dimension, battery state risk features, such as voltage anomaly, are extracted from the verified weighted fusion dataset. In the temperature distribution risk dimension, temperature risk features, such as local maximum temperature, are extracted from the verified weighted fusion dataset. In the environmental association risk dimension, environmental association risk features, such as strong interference type, are extracted from the verified weighted fusion dataset. In the risk precursor dimension, risk precursor features, such as characteristic gas concentration anomaly level, are extracted from the verified weighted fusion dataset. The battery state risk features, temperature risk features, environmental association risk features, and risk precursor features are integrated and labeled with weight values, device ID, timestamp, and spatial location to generate a standardized global risk feature set.

[0097] Specifically, the steps for determining the early risk level include:

[0098] Key risk features are extracted from the global risk feature set, including environmentally inhibited features and environmentally stimulated features. Among them, environmentally inhibited features are risk features whose risk indication effectiveness decreases and whose correlation with the target risk weakens after being affected by environmental factors. Environmentally inhibited features include, but are not limited to, feature gas concentrations. Environmentally stimulated features are risk features whose feature intensity increases and whose correlation with the target risk is enhanced after being affected by environmental factors. Environmentally stimulated features include, but are not limited to, local temperature.

[0099] The critical values ​​of key risk characteristics in the early stages of risk are extracted from historical risk cases. Combined with industry safety standards, the critical values ​​of each key risk characteristic are modified. For example, the critical value of local temperature is reduced according to industry safety standards. Finally, the benchmark threshold of each key risk characteristic is determined.

[0100] Construct a risk assessment matrix with key risk characteristics as rows and benchmark thresholds for each key risk characteristic as columns;

[0101] For features suppressed by the environment, the characteristic signal of the characteristic gas concentration is distorted due to the smoke scattering noise caused by high dust concentration. Therefore, the environmental interference factor is obtained from the global risk feature set, and the amplification coefficient is introduced through the noise attenuation formula to calibrate the benchmark threshold of the characteristic gas concentration and generate the calibrated first identification threshold.

[0102] For environmentally excited features, the local maximum heating rate is obtained from the global risk feature set. A sensitivity coefficient is introduced through a sensitivity function to calibrate the local temperature reference threshold and generate a calibrated second identification threshold.

[0103] The risk assessment matrix is ​​dynamically updated based on the calibrated first identification threshold and the calibrated second identification threshold.

[0104] Based on the dynamically updated risk assessment matrix, for environmentally inhibited features, a first confidence level is calculated using a first identification threshold as a benchmark and combined with the real-time value of the environmentally inhibited features; for environmentally stimulated features, a second confidence level is calculated based on a second identification threshold and the real-time value of the environmentally stimulated features.

[0105] Based on the risk contribution of key risk characteristics in historical risk cases and expert experience, a first and second weight are established. Combining the first and second confidence levels, a weighted summation algorithm is used to calculate the early risk probability. ;

[0106] Risk assessment thresholds are set based on the risk probability distribution of historical risk cases and industry safety early warning standards. , ,and To determine the early risk level; if If so, it is classified as a Level 1 risk, and normal monitoring is maintained; if If it is determined to be a level 2 risk, a security warning will be triggered; if If so, it is determined to be a level three risk, triggering emergency response.

[0107] Specifically, the steps for generating a dynamic risk evolution curve include:

[0108] Global temperature distribution data is extracted from the global risk feature set, and the spatial coordinates corresponding to each global temperature distribution data are labeled. The K-Means clustering algorithm is used to remove abnormal outliers of global temperature distribution data, retain the valid temperature data, and sort the valid temperature data according to the timestamp to form continuous global temperature time series data.

[0109] The sliding window size is dynamically selected based on the early risk level; for level one risk, the sliding window size is adopted. Minute window, Level 2 risk adopted Minute window, Level 3 risk adopted Minute window, and Using the Savitzky-Golay filtering algorithm, a smooth temperature value at the center point of the window is generated by fitting the global temperature time series data within each window, and the temperature change curve is generated by sliding point by point.

[0110] Differential features are extracted from temperature change curves based on early risk levels. For Level 1 risk, the average temperature and temperature standard deviation are extracted. For Level 2 risk, the rate of temperature change and the maximum local temperature difference are extracted. For Level 3 risk, the temperature acceleration is extracted. The differential features are then concatenated with smoothed temperature values ​​to form a feature-enhanced time series.

[0111] Construct a risk evolution model, including an input layer, a risk LSTM layer, a feature fusion layer, and an output layer;

[0112] The input layer is used to receive feature-enhanced time-series sequences and early risk levels;

[0113] The risk LSTM layer is used to introduce a temporal attention mechanism to learn the evolutionary relationship between temperature changes and early risk levels, before descending sorting. Each attention weight time step generates a risk temporal feature vector;

[0114] The feature fusion layer maps early risk levels to the temporal feature space through dimension adaptation, and uses nonlinear interactive operation and gating modulation to fuse risk temporal feature vectors and early risk levels to generate a comprehensive risk feature vector.

[0115] The output layer is used to output the future risk based on the comprehensive risk feature vector and the softmax activation function. Hourly risk probability distribution;

[0116] Historical risk case data is acquired and preprocessed to generate a historical risk case dataset, which is then divided into a training set and a validation set according to a preset ratio. The training set is used to train a risk evolution model using the Adam optimizer, while the validation set is used to evaluate the performance of the risk evolution model. Feature-enhanced time series sequences are input into the trained risk evolution model, and the model iteratively outputs future... Risk probability sequence within an hour; wherein, in this embodiment, the preset ratio is set to 7:3;

[0117] Based on the risk probability sequence, the risk probability difference between adjacent time steps is calculated. Time step weights are set according to industry expert experience, and the risk evolution rate is calculated by weighting each time step. , shaping the future Risk evolution rate sequence within hours;

[0118] Set the rate threshold based on industry expert experience. , Dynamically adjust the risk assessment thresholds for each early risk level. If so, the risk assessment threshold for the early risk level is raised, and the early risk level is upgraded by one level; if If so, the risk assessment threshold of the early risk level will be maintained, and a security risk warning will be triggered; if If so, the risk assessment threshold for the early risk level remains unchanged, and monitoring continues;

[0119] Integrating temperature change curves, future The system generates a real-time updated dynamic risk evolution curve by annotating the risk probability sequence, dynamically adjusted risk judgment threshold, and risk level transition points within each hour, along with safety risk warning information at each time step. The risk level transition point is the time step and corresponding risk state node where the risk probability value first exceeds the adjusted judgment threshold after the risk evolution rate reaches a preset upward adjustment threshold, triggering an upward adjustment of the early risk level judgment threshold. The preset upward adjustment threshold is determined through a combination of industry expert experience, historical risk evolution rate statistics, and the characteristic patterns of the target risk type.

[0120] Specifically, the steps for determining the match between the intensity of risk warnings and the early risk level include:

[0121] By utilizing edge computing nodes deployed on the bottom and surrounding areas of electric bicycles, trigger status data of audio-visual prompt devices, such as the intensity of risk prompts, are collected in real time, and raw feature data streams are generated through preprocessing.

[0122] A sliding window mid-value filter is used on the edge side to smooth the risk warning intensity signal in the original feature data stream to remove instantaneous peak noise, and dimensionality normalization is performed to generate a standardized risk state feature vector.

[0123] An edge computing algorithm is used to calculate the ambient noise intensity in real time. The normal noise range is set according to the working environment of the audio-visual prompt device. If the ambient noise intensity exceeds the normal noise range, the data quality label is generated as low quality; otherwise, the data quality label is generated as high quality.

[0124] The current risk level is extracted from the dynamic risk evolution curve. The quantitative value of the current risk level is obtained through one-hot encoding and numerical mapping. The risk state feature vector is quantized into risk state feature value. The first matching degree between the risk state feature value and the quantitative value of the current risk level is calculated.

[0125] If the data quality label is low quality, the confidence coefficient is determined based on the degree of noise exceeding the standard and empirical weights. The product of the confidence coefficient and the first matching degree is calculated to obtain the matching confidence degree. If the data quality label is high quality, the first matching degree is directly defined as the matching confidence degree.

[0126] A matching threshold is set based on the risk evolution rate and industry expert experience. A matching assessment is conducted. If the matching confidence is greater than or equal to the matching threshold, the risk status feature vector is determined to match the current risk level, and no risk warning status adjustment is made; otherwise, the risk status feature vector is determined to not match the current risk level.

[0127] For mismatches, the absolute difference between the matching confidence score and the matching threshold is calculated based on the matching threshold to obtain the matching gap. To quantify the current degree of mismatch, gap thresholds are set based on the experience of industry experts. , Adjust the priority determination, if If it is, then it is determined to be the first priority; if If it is, then it is determined to be the second priority; if If so, it is determined to be the third priority;

[0128] Establish gap grading rules and adjust risk warning levels; for the first priority, adjust the risk warning intensity across levels, such as raising the risk warning intensity level and switching the risk warning mode to continuous mode when the risk level is level three; for the second priority, adjust the risk warning intensity level step by step; for the third priority, adjust the trigger frequency while keeping the current risk warning mode unchanged.

[0129] The adjusted risk warning status, current risk level, and matching confidence level are integrated into a risk status data package and fed back to the risk assessment matrix. Incremental learning is used to update the risk assessment matrix, and the updated early risk level is output.

[0130] The updated early risk level is input into the risk evolution model, and the future risk level is recalculated using the LSTM algorithm. The risk probability sequence within an hour is dynamically adjusted, the risk judgment threshold is updated, and the dynamic risk evolution curve is updated. Based on the updated dynamic risk evolution curve, the matching confidence level is recalculated, and a matching determination is made. When the matching confidence level is greater than or equal to the matching threshold, the iterative calibration stops. If the calibration fails in the next iteration, a manual review prompt will be triggered; among other things... The specific size can be set by those skilled in the art according to actual needs, and this embodiment does not limit it.

[0131] Example 2

[0132] Please see Figure 5 Another embodiment of the present invention provides an electric bicycle safety risk dynamic assessment system, comprising: a risk perception module, a risk assessment module, a risk evolution module, and a matching calibration module;

[0133] The risk perception module is used to collect battery status data, risk precursor data, global temperature distribution data and environmental interference data of electric bicycles, perform spatiotemporal alignment at the edge, increase the weight of risk precursor features through an attention mechanism, filter environmental noise signals using an adaptive filtering algorithm, and generate a global risk feature set.

[0134] The risk assessment module is used to construct a risk assessment matrix based on the global risk feature set, dynamically adjust the identification threshold of key risk features, increase the first identification threshold of characteristic gas concentration in high dust environment, and decrease the local temperature when local heating is detected. The risk assessment matrix is ​​used to calculate the early risk probability in real time to determine the early risk level.

[0135] The risk evolution module extracts continuous global temperature time-series data, generates temperature change curves, combines early risk levels to construct a risk evolution model, and then uses the LSTM time-series prediction algorithm to output future risk data. Calculate the risk evolution probability sequence within an hour, calculate the risk evolution rate, dynamically adjust the risk assessment threshold for early risk levels, and generate a dynamic risk evolution curve.

[0136] The matching calibration module is used to collect trigger status data, generate risk status feature vectors, calculate the first matching degree between the risk status feature vectors and the current risk level in real time based on edge computing, determine the matching confidence level, and make a matching judgment. If there is no match, the risk warning level is adjusted, and the risk status data is fed back to the risk assessment matrix to update the early risk level and dynamic risk evolution curve until the first matching degree reaches the preset standard.

[0137] Working principle and effects:

[0138] By collecting battery status data, risk precursor data, global temperature distribution data, and environmental interference data of electric bicycles, multi-source heterogeneous data are integrated and spatiotemporally aligned at the edge. The weight of risk precursor features is increased through an attention mechanism, and environmental noise signals are filtered out using an adaptive filtering algorithm to generate a global risk feature set. This solves the problems of limited detection area and distortion of risk features caused by noise interference in traditional data collection.

[0139] A risk assessment matrix is ​​constructed based on a global risk feature set. The key risk feature identification thresholds are dynamically adjusted according to environmental interference data. The risk assessment matrix is ​​used to calculate the early risk probability and determine the early risk level in real time, ensuring that the risk assessment is adapted to environmental differences. This solves the problem of inaccurate risk assessment caused by high false alarm rate in traditional methods.

[0140] Based on global temperature time-series data, the risk evolution module generates temperature change curves, constructs a risk evolution model by combining early risk levels, and outputs future risk predictions using the LSTM time-series prediction algorithm. The risk evolution probability sequence within an hour is used to calculate the risk evolution rate and dynamically adjust the risk judgment threshold to generate a dynamic risk evolution curve, thereby realizing early prediction of risk trends and solving the problem of lag in traditional static prediction of risk trends.

[0141] The system collects trigger status data from audio-visual alert devices, generates risk status feature vectors, calculates the first degree of matching between the risk status feature vectors and the current risk level in real time based on edge computing, and determines the matching confidence level. If there is no match, the risk alert level is adjusted, and the risk status data is fed back to the risk assessment matrix to re-determine the early risk level and update the dynamic risk evolution curve, thus solving the problem of risk alert failure.

[0142] Overall, through a four-layer architecture of risk perception, risk assessment, risk evolution, and matching calibration, and by integrating equipment status, environmental characteristics, and risk trends, the system achieves perception, dynamic assessment, and intelligent matching calibration of electric bicycle safety risks. This effectively solves the problems of limited detection area, susceptibility to environmental interference, and late-stage risk identification in electric bicycle safety risk prevention and control, which lead to increased false alarm rates. It provides an intelligent prevention and control solution for electric bicycle safety risks.

[0143] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic risk assessment method for electric bicycles, characterized in that, include: Collect battery status data, risk precursor data, global temperature distribution data and environmental interference data, align them spatiotemporally at the edge, increase the weight of risk precursor features, filter environmental noise signals, and generate a global risk feature set. Construct a risk assessment matrix, dynamically adjust the first and second identification thresholds for key risk characteristics, and calculate the probability of early risks to determine the level of early risks; Extract global temperature time-series data, generate temperature change curves, combine with the early risk levels to construct a risk evolution model, and output future risk data. Calculate the risk evolution probability sequence within an hour, calculate the risk evolution rate, dynamically adjust the risk judgment threshold, and generate a dynamic risk evolution curve; Obtain the risk status feature vector, calculate the first matching degree in combination with the current risk level, determine the matching confidence, and make a matching judgment. If there is no match, adjust the risk warning level and update the early risk level and the dynamic risk evolution curve.

2. The method for dynamic assessment of safety risks of electric bicycles according to claim 1, characterized in that, The specific steps for generating a global risk feature set include: Collect battery status data, early warning data, global temperature distribution data, and environmental interference data of electric bicycles; For multi-source heterogeneous data at the edge, the time delay of different sensors is corrected, and after spatiotemporal alignment, a multidimensional fusion dataset is formed. Dynamically adjust the core parameters of the adaptive RLS filtering algorithm and calculate the signal-to-noise ratio. If the signal-to-noise ratio is less than the preset signal-to-noise ratio threshold, continue filtering to select valid signals; otherwise, stop filtering to form a denoised multidimensional fusion dataset. An attention mechanism network model is used to input a denoised multidimensional fusion dataset, extract risk precursor features, construct an initial weight library, calculate and dynamically adjust the basic weights of each risk precursor feature, and output a weighted fusion dataset. According to the coverage integrity verification rules, identify and supplement the central monitoring blind spots or missing data areas in the weighted fusion dataset. Battery state risk features, temperature risk features, environmental associated risk features, and risk precursor features are extracted from the verified weighted fusion dataset to generate a global risk feature set.

3. The method for dynamic assessment of safety risks of electric bicycles according to claim 2, characterized in that, The specific steps for determining the early risk level include: Extract key risk characteristics, including those inhibited by the environment and those stimulated by the environment; Extract the critical values ​​of each key risk characteristic in the early stage of risk, revise and determine the benchmark thresholds of each key risk characteristic, and construct a risk assessment matrix. For the environmentally inhibited features, an environmental interference factor is obtained, an amplification coefficient is introduced, a reference threshold for the concentration of the feature gas is calibrated, and a calibrated first identification threshold is generated. For the environmental excitation characteristics, the local maximum heating rate is obtained, a sensitivity coefficient is introduced, a reference threshold for the local temperature is calibrated, and a calibrated second recognition threshold is generated. The risk assessment matrix is ​​dynamically updated based on the first identification threshold and the second identification threshold.

4. The method for dynamic assessment of safety risks of electric bicycles according to claim 3, characterized in that, The specific steps for determining the early risk level also include: Based on the dynamically updated risk assessment matrix, a first confidence level is calculated for the environmentally inhibited features based on the first identification threshold and combined with the real-time values ​​of the environmentally inhibited features; For the environmentally stimulated feature, a second confidence level is calculated based on the second identification threshold and the real-time value of the environmentally stimulated feature; A first weight and a second weight are set, and the early risk probability is calculated by combining the first confidence level and the second confidence level. ; Set risk assessment threshold , To determine the early risk level; like If so, it is classified as a Level 1 risk; if If so, it is judged as a level two risk; if If so, it is classified as a level three risk.

5. The method for dynamic assessment of safety risks of electric bicycles according to claim 4, characterized in that, The specific steps for generating a temperature change curve include: Extract global temperature distribution data, label spatial coordinates, and use the K-Means clustering algorithm to remove outlier global temperature distribution data, retaining valid temperature data; The valid temperature data are sorted according to timestamps to form continuous global temperature time series data; The sliding window size is dynamically selected based on the aforementioned early risk level; Level 1 risk adopts Minute window, Level 2 risk adopted Minute window, Level 3 risk adopted Minute window, and ; For the global temperature time series data within each window, a smooth temperature value is generated by fitting using the Savitzky-Golay filtering algorithm, and a temperature change curve is generated by sliding point by point.

6. The method for dynamic assessment of safety risks of electric bicycles according to claim 5, characterized in that, Output the future The specific steps for the risk evolution probability sequence within an hour include: Differential features are extracted from the temperature change curve and concatenated with the smoothed temperature value to form a feature-enhanced time series sequence; Construct a risk evolution model, including an input layer, a risk LSTM layer, a feature fusion layer, and an output layer; The input layer is used to receive the feature-enhanced temporal sequence and the early risk level; The risk LSTM layer is used to learn the evolutionary relationship between temperature changes and the early risk levels, before descending sorting. Each attention weight time step generates a risk temporal feature vector; The feature fusion layer maps the early risk level to a temporal feature space and fuses it with the risk temporal feature vector to generate a comprehensive risk feature vector; The output layer is used to output the future risk based on the comprehensive risk feature vector and the softmax activation function. Hourly risk probability distribution; Obtain a historical risk case dataset, train the risk evolution model using the Adam optimizer, and output the future... Risk probability sequence within hours.

7. The method for dynamic assessment of safety risks of electric bicycles according to claim 6, characterized in that, The specific steps for generating a dynamic risk evolution curve include: Based on the risk probability sequence, calculate the risk probability difference between adjacent time steps; Set time step weights and calculate the risk evolution rate by weighting each time step. , shaping the future Risk evolution rate sequence within hours; Set rate threshold , Dynamically adjust the risk assessment thresholds for each early risk level; like If so, the risk assessment threshold for the early risk level will be raised, and the early risk level will be upgraded by one level. like If so, the risk assessment threshold of the early risk level will be maintained, and a security risk warning will be triggered; like If so, the risk assessment threshold for the early risk level remains unchanged, and monitoring continues; By integrating the temperature change curve, the risk probability sequence, the dynamically adjusted risk judgment threshold, and the risk level transition point, a dynamic risk evolution curve is generated.

8. The method for dynamic assessment of safety risks of electric bicycles according to claim 7, characterized in that, The specific steps for determining the match confidence level include: Real-time acquisition of trigger status data, generation of raw feature data stream, smoothing of risk warning intensity signal in the raw feature data stream, and generation of risk status feature vector; An edge computing algorithm is used to calculate the ambient noise intensity in real time and set the normal noise range. If the ambient noise intensity exceeds the normal noise range, the generated data quality label is low quality; otherwise, the generated data quality label is high quality. Extract the current risk level from the dynamic risk evolution curve to form a quantitative value of the current risk level, quantify the risk state feature vector into a risk state feature value, and calculate the first matching degree; If the data quality label is low quality, then determine the confidence coefficient and calculate the matching confidence score in combination with the first matching score; If the data quality label is high quality, then the first match score is directly used as the match confidence score.

9. The method for dynamic assessment of safety risks of electric bicycles according to claim 8, characterized in that, The specific steps for matching include: Set a matching threshold to determine the matching accuracy. If the matching confidence score is greater than or equal to the matching threshold, then a match is determined; otherwise, a mismatch is determined. For mismatches, calculate the matching gap. Set the gap threshold , Adjust the priority determination; like If it is, then it is determined to be the first priority; if If it is, then it is determined to be the second priority; if If so, it is determined to be the third priority; Establish gap-level rules: for the first priority, adjust the risk warning intensity across levels; for the second priority, adjust the risk warning intensity level by level; for the third priority, adjust the trigger frequency while keeping the current risk warning mode unchanged. Risk status data is fed back to the risk assessment matrix to update early risk levels and dynamic risk evolution curves; Recalculate the matching confidence score; stop iterative calibration when the matching confidence score is greater than or equal to the matching threshold; when the matching confidence score is continuous... If the calibration fails to meet the requirements in the next iteration, a manual review prompt will be triggered.

10. A dynamic safety risk assessment system for electric bicycles, used to implement the dynamic safety risk assessment method for electric bicycles as described in any one of claims 1-9, characterized in that, include: Risk perception module, risk assessment module, risk evolution module, and matching calibration module; The risk perception module is used to collect battery status data, risk precursor data, global temperature distribution data and environmental interference data. After spatiotemporal alignment, the weight of risk precursor features is increased and environmental noise signals are filtered to generate a global risk feature set. The risk assessment module is used to construct a risk assessment matrix, dynamically adjust the first and second identification thresholds of key risk characteristics, and calculate the early risk probability in real time to determine the early risk level. The risk evolution module is used to extract global temperature time-series data, generate temperature change curves, combine them with the early risk levels, construct a risk evolution model, and output future risk data. Calculate the risk evolution probability sequence within an hour, calculate the risk evolution rate, dynamically adjust the risk judgment threshold, and generate a dynamic risk evolution curve; The matching calibration module is used to acquire the risk state feature vector, calculate the first matching degree in combination with the current risk level, determine the matching confidence level, and make a matching judgment. If there is no match, the risk warning level is adjusted and the early risk level and the dynamic risk evolution curve are updated.

Images