Electric tricycle safe collision early warning system and method

Abstract

The invention relates to the technical field of electric tricycle safety, and discloses an electric tricycle safety collision early warning system and method, and the system comprises a multi-source sensor module, a feature extraction and fusion module, a collision risk assessment module, a path optimization module and a grading early warning control module. The method comprises the steps that the multi-source sensor module collects real-time environment and vehicle state data; the convolutional neural network extracts fusion environment data features to generate vectors, and the vectors are input into a Bayesian network model to output a collision risk probability value; constructing a multi-target path optimization model according to the probability value, and searching an optimal obstacle avoidance path by using an improved particle swarm optimization algorithm; and establishing a grading early warning control model according to the path data, and outputting brake pressure and sound-light alarm signals through a fuzzy logic controller. The system can accurately sense the environment, accurately evaluate the risk, optimize the obstacle avoidance path and perform graded early warning control, improves the driving safety of the electro-tricycle, also has sensor self-inspection and hot backup functions, and enhances the system reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of electric tricycle safety, and in particular to an electric tricycle safety collision warning system and method. Background Art

[0002] With the widespread use of electric tricycles in logistics, transportation, short-distance travel and other fields, their safety issues have attracted more and more attention. Electric tricycles usually travel in complex urban streets, rural roads and other environments, facing many potential collision risks.

[0003] In actual use scenarios, the safety awareness of electric tricycle drivers varies. Some drivers lack standardized driving skills and safety knowledge, and their ability to predict potential dangers is insufficient. At the same time, the traffic conditions on the road are complex and changeable, with pedestrians walking around at will and other vehicles driving illegally, all of which increase the possibility of collision accidents involving electric tricycles. According to relevant statistics, the number of traffic accidents involving electric tricycles has been on the rise in recent years, and a large part of these accidents are caused by drivers failing to detect danger in time and take effective measures.

[0004] The existing safety protection measures for electric tricycles are relatively simple. Most vehicles are only equipped with basic braking devices and lack active safety warning systems. Traditional braking systems are often unable to quickly stop the vehicle in an emergency, making collision accidents difficult to avoid. Moreover, even if some electric tricycles are equipped with some auxiliary safety equipment, such as reversing radar, their functions are relatively simple and can only work in specific scenarios. They cannot fully perceive the environment around the vehicle and lack effective assessment and warning capabilities for potential collision risks.

[0005] In terms of data collection and processing, existing technical means are difficult to meet the safety requirements of electric tricycles in complex environments. The application of a single sensor has limitations. For example, if only cameras are used for environmental monitoring, the image collection and recognition effects will be seriously affected in bad weather conditions (such as heavy rain and fog); and the use of radar alone, although the measurement of distance and speed is more accurate, cannot obtain the detailed shape and category information of obstacles. In addition, for the collected data, the existing processing algorithms are inefficient and cannot quickly and accurately extract key information, resulting in a significant reduction in the timeliness and accuracy of collision warnings.

[0006] In terms of path planning and obstacle avoidance, existing electric tricycles lack intelligent path planning capabilities. When encountering obstacles, drivers often rely on experience to avoid them, which is neither scientific nor safe. In narrow roads or busy traffic areas, unreasonable avoidance operations may lead to secondary accidents. Moreover, most existing obstacle avoidance algorithms do not fully consider the dynamic characteristics of electric tricycles, such as the turning radius of the vehicle, braking response time, etc., resulting in a low feasibility of the planned obstacle avoidance path in practical applications. Summary of the Invention

[0007] The purpose of the present invention is to provide an electric tricycle safety collision warning system and method to solve the problems raised in the above background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solution: An electric tricycle safety collision warning method, the method includes:

[0009] Collect real-time environmental data and vehicle state data of the electric tricycle through a multi-source sensor module, and the multi-source sensor module includes a millimeter-wave radar, a binocular vision sensor, an ultrasonic sensor, and an inertial measurement unit;

[0010] Perform multi-scale feature extraction and fusion on the real-time environmental data based on a convolutional neural network to generate an environmental feature vector; input the environmental feature vector into a pre-trained Bayesian network model, and the Bayesian network model adopts a dynamic node update mechanism to adjust the conditional probability distribution based on real-time observation data and output a collision risk probability value;

[0011] Construct a multi-objective path optimization model according to the collision risk probability value, and the multi-objective path optimization model takes the highest obstacle avoidance priority and the minimum path curvature as optimization objectives, and uses an improved particle swarm optimization algorithm to globally search for a feasible path, where the improved particle swarm optimization algorithm introduces an adaptive inertia weight and a dynamic learning factor; output the optimal obstacle avoidance path data based on the multi-objective path optimization model;

[0012] Establish a hierarchical warning control model according to the optimal obstacle avoidance path data, and the hierarchical warning control model includes a data layer, a decision layer, and an execution layer. The data layer updates the risk level based on the environmental feature vector, the decision layer generates a multi-level warning strategy based on the collision risk probability value, and the execution layer maps the warning instruction into a braking pressure and an audible and visual alarm signal through a fuzzy logic controller.

[0013] Preferably, the dynamic node update mechanism adopted by the Bayesian network model includes:

[0014] Construct a Bayesian network structure, which includes an environmental observation node, a vehicle state node, and a collision risk node. The environmental observation node and the vehicle state node are parent nodes, and the collision risk node is a child node;

[0015] Update the evidence probability of the environmental observation node based on real-time sensor data, and perform state estimation on the vehicle state node through the Kalman filter algorithm to obtain the posterior probability distributions of vehicle position, speed, and acceleration;

[0016] Introduce a dynamic time window mechanism to dynamically adjust the inference time interval of the Bayesian network according to the vehicle movement trend. The length of the time window is negatively correlated with the vehicle speed;

[0017] Use the variational inference algorithm to perform online parameter update on the Bayesian network, and optimize the conditional probability table parameters by minimizing the evidence lower bound.

[0018] Preferably, the improved particle swarm optimization algorithm introduces an adaptive inertia weight and a dynamic learning factor, including:

[0019] Construct a particle swarm optimization objective function, which includes a path obstacle avoidance safety term and a path smoothness term. The path obstacle avoidance safety term is calculated by the minimum distance between the path points and the obstacles, and the path smoothness term is calculated by the curvature change amount of adjacent path points;

[0020] Design an adaptive inertia weight update rule. The inertia weight decays linearly in segments as the number of iterations increases. In the initial stage, a higher global search ability is retained, and in the later stage, the local search accuracy is enhanced;

[0021] Design a dynamic learning factor adjustment mechanism. The learning factor is dynamically adjusted according to the population diversity index. When the population diversity is lower than the threshold, the weight of the social learning factor is increased, and vice versa, the weight of the individual learning factor is increased.

[0022] Preferably, the fuzzy logic controller maps the warning instruction into a braking pressure and an audible and visual alarm signal, including:

[0023] Construct an input fuzzification module. The inputs include the collision risk probability value, the relative distance to the obstacle, and the current vehicle speed. The triangular membership function is used to perform fuzzification processing on the input variables;

[0024] Construct a fuzzy rule base, which contains IF-THEN rules;

[0025] Design a defuzzification module, and use the centroid method to convert the fuzzy output into an exact control quantity. The control quantity includes the braking pressure gradient and the audible and visual alarm frequency;

[0026] Introduce a feedback correction mechanism to dynamically adjust the confidence of fuzzy rules according to the actual braking effect and the change rate of the distance to the obstacle.

[0027] Preferably, the data fusion processing of the multi-source sensor module includes:

[0028] Adopt a constant false alarm rate detection algorithm for millimeter-wave radar data to extract obstacle distance and speed information;

[0029] Adopt a stereo matching algorithm for binocular vision data to generate a depth map, and remove noise through morphological filtering;

[0030] Adopt a moving average filter for ultrasonic sensor data to eliminate short-term interference;

[0031] Adopt a quaternion complementary filtering algorithm for inertial measurement unit data to fuse gyroscope and accelerometer data, and output vehicle attitude angle and angular velocity.

[0032] Preferably, the decision-making layer of the hierarchical early warning control model generates multi-level early warning strategies including:

[0033] Divide the risk levels into three levels: low, medium, and high. The low risk level corresponds to audible and visual warnings, the medium risk level corresponds to pre-pressurization of braking, and the high risk level triggers full braking;

[0034] Construct a state machine model. The state machine performs state transitions according to continuous risk level changes. If the risk level rises for two consecutive frames of data, it jumps to a higher early warning level;

[0035] Introduce a delay trigger mechanism. When the high risk level lasts for more than a preset time threshold, start the emergency obstacle avoidance protocol.

[0036] Preferably, the constraint conditions of the multi-objective path optimization model include:

[0037] Path curvature constraint, limiting the maximum curvature of path points not to exceed the curvature threshold corresponding to the minimum turning radius of the vehicle;

[0038] Dynamic obstacle avoidance constraint, predicting the motion trajectory of obstacles through the velocity obstacle method and generating an avoidance corridor;

[0039] Actuator dynamics constraint, limiting the physical upper limits of the braking pressure change rate and the steering angular velocity.

[0040] Preferably, the system further includes:

[0041] Construct an online self-check module to detect sensor faults through cross-verification of redundant sensor data. If the data of a certain sensor deviates from the mean by more than three standard deviations, it is marked as abnormal;

[0042] Adopt a hot backup mechanism. When the main control unit fails, switch to the standby controller, and maintain the synchronization of the communication status of the dual machines through heartbeat packets.

[0043] Preferably, the present invention further includes an electric tricycle safety collision warning system, and the system includes:

[0044] A multi-source sensor module: used to collect real-time environmental data and vehicle status data of the electric tricycle, including a millimeter-wave radar, a binocular vision sensor, an ultrasonic sensor, and an inertial measurement unit;

[0045] A feature extraction and fusion module: perform multi-scale feature extraction and fusion on the real-time environmental data collected by the multi-source sensor module based on a convolutional neural network to generate an environmental feature vector;

[0046] A collision risk assessment module: input the environmental feature vector into a pre-trained Bayesian network model. This Bayesian network model adopts a dynamic node update mechanism, adjusts the conditional probability distribution based on real-time observation data, and outputs a collision risk probability value;

[0047] A path optimization module: construct a multi-objective path optimization model according to the collision risk probability value, take the highest obstacle avoidance priority and the minimum path curvature as the optimization objectives, and use an improved particle swarm optimization algorithm introducing an adaptive inertia weight and a dynamic learning factor to globally search for a feasible path, and output the optimal obstacle avoidance path data;

[0048] A hierarchical warning control module: established according to the optimal obstacle avoidance path data, including a data layer, a decision layer, and an execution layer. Among them, the data layer updates the risk level based on the environmental feature vector, the decision layer generates a multi-level warning strategy based on the collision risk probability value, and the execution layer maps the warning instruction into a braking pressure and an audible and visual alarm signal through a fuzzy logic controller.

[0049] Preferably, the present invention further includes an electronic device, including:

[0050] A processor;

[0051] A memory for storing instructions executable by the processor;

[0052] Wherein, the processor is configured to call the instructions stored in the memory to execute the functions of the above-mentioned electric tricycle safety collision warning method.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] The present invention adopts a multi-source sensor module, integrating millimeter-wave radar, binocular vision sensor, ultrasonic sensor and inertial measurement unit. Millimeter-wave radar can accurately measure the distance and speed of obstacles, and is not affected by lighting conditions; binocular vision sensor can identify the shape and category of obstacles and provide rich visual information; ultrasonic sensor has high accuracy in close-range detection; inertial measurement unit monitors vehicle posture and motion state in real time. Multi-source sensor data is processed through a specific fusion algorithm, such as using a constant false alarm rate detection algorithm for millimeter-wave radar data, and generating a depth map through a stereo matching algorithm for binocular vision data, which effectively makes up for the limitations of a single sensor, improves the accuracy and reliability of environmental perception, and provides a solid data foundation for subsequent collision risk assessment and early warning.

[0055] Based on the multi-scale feature extraction and fusion of real-time environmental data by convolutional neural network, more representative features can be extracted from complex environmental data to generate environmental feature vectors. The vectors are input into the Bayesian network model with dynamic node update mechanism. The model constructs a reasonable network structure, including environmental observation nodes, vehicle status nodes and collision risk nodes, updates node status based on real-time sensor data, uses Kalman filter algorithm to accurately estimate vehicle status, and adjusts the inference time interval and optimizes the conditional probability table parameters through dynamic time window mechanism and variational inference algorithm. It can accurately and real-timely output collision risk probability values, and buy more time for early warning and taking measures.

[0056] A multi-objective path optimization model is constructed based on the collision risk probability value, with the highest obstacle avoidance priority and the smallest path curvature as the optimization goal. An improved particle swarm optimization algorithm that introduces adaptive inertia weights and dynamic learning factors is adopted. By constructing an objective function that includes path obstacle avoidance safety terms and path smoothness terms, and rationally designing the adjustment mechanism of inertia weights and learning factors, the optimal obstacle avoidance path can be quickly searched in a complex environment. At the same time, the model takes into account path curvature constraints, dynamic obstacle avoidance constraints, and actuator dynamics constraints to ensure that the planned path is both safe and in line with the actual driving capabilities of the vehicle, effectively avoiding secondary accidents caused by unreasonable obstacle avoidance paths.

[0057] The hierarchical early warning control model includes a data layer, a decision-making layer, and an execution layer. The data layer updates the risk level in real time according to the environmental feature vector. The decision-making layer divides the risk level based on the collision risk probability value and generates a multi-level early warning strategy. For example, when the risk is low, an audible and visual warning is given; when the risk is medium, pre-pressure is applied to the brakes; when the risk is high, full braking is triggered. At the same time, the accuracy and timeliness of the early warning are ensured through the state machine model and the delay trigger mechanism. The execution layer uses a fuzzy logic controller. By constructing an input fuzzification module, a fuzzy rule base, and a defuzzification module, and introducing a feedback correction mechanism, the early warning instruction is accurately mapped into a braking pressure and an audible and visual alarm signal, realizing precise control of the vehicle and improving driving safety.

[0058] The present invention also includes an on-line self-checking module and a hot backup mechanism. The on-line self-checking module detects sensor faults through cross-verification of redundant sensor data, discovers and marks abnormal sensors in a timely manner, and ensures the reliability of data acquisition. The hot backup mechanism quickly switches to a standby controller when the main control unit fails, maintains the synchronization of the communication status of the dual machines through heartbeat packets, ensures the uninterrupted operation of the system, improves the stability and reliability of the entire system, and reduces potential safety hazards caused by system failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a working principle diagram of the electric tricycle safety collision warning method described in the present invention;

[0060] Figure 2 is a working flow chart of the improved particle swarm optimization algorithm;

[0061] Figure 3 is a working flow chart of the fuzzy logic controller. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0063] Please refer to Figures 1-3 , the present invention provides a technical solution: an electric tricycle safety collision warning system and method, and the method includes:

[0064] Utilize a multi-source sensor module, including a millimeter-wave radar, a binocular vision sensor, an ultrasonic sensor, and an inertial measurement unit, to collect real-time environmental data and vehicle status data of the electric tricycle. The millimeter-wave radar can obtain the distance and speed information of surrounding obstacles; the binocular vision sensor can identify the shape and category of obstacles, etc.; the ultrasonic sensor plays a high-precision advantage in close-range detection; the inertial measurement unit is used to monitor the attitude angle, angular velocity and other status data of the vehicle.

[0065] Based on a convolutional neural network, perform multi-scale feature extraction and fusion on the collected real-time environmental data to generate an environmental feature vector. The convolutional neural network has a powerful image feature extraction ability. Through the combination of different convolutional layers and pooling layers, it can extract rich feature information from the environmental data. Input the environmental feature vector into a pre-trained Bayesian network model. This model adopts a dynamic node update mechanism, adjusts the conditional probability distribution according to real-time observation data, and outputs the collision risk probability value. The Bayesian network model can comprehensively consider the probability relationships among various factors and accurately evaluate the collision risk.

[0066] Construct a multi-objective path optimization model according to the collision risk probability value, with the highest obstacle avoidance priority and the minimum path curvature as the optimization objectives. Use an improved particle swarm optimization algorithm to perform a global search for feasible paths. The improved particle swarm optimization algorithm introduces an adaptive inertia weight and a dynamic learning factor, and finally outputs the optimal obstacle avoidance path data. The improved algorithm can improve the search efficiency and find a safe and smooth obstacle avoidance path faster.

[0067] Establish a hierarchical early warning control model according to the optimal obstacle avoidance path data. This model includes a data layer, a decision layer, and an execution layer. The data layer updates the risk level based on the environmental feature vector; the decision layer generates multi-level early warning strategies based on the collision risk probability value; the execution layer maps the early warning instructions into braking pressure and sound and light alarm signals through a fuzzy logic controller to achieve effective early warning and control of the vehicle.

[0068] The following further illustrates the present invention in combination with Embodiments 1 to 5:

[0069] Embodiment 1:

[0070] This embodiment details the specific implementation method of the Bayesian network model adopting the dynamic node update mechanism, and improves the accuracy and real-time performance of collision risk assessment by more accurately updating the model node state and adjusting the inference time interval. The specific steps include:

[0071] Construct a Bayesian network structure that includes environmental observation nodes, vehicle state nodes, and collision risk nodes. The environmental observation nodes are used to receive environmental data collected by millimeter-wave radars, binocular vision sensors, etc., such as obstacle distance, azimuth, etc.; the vehicle state nodes receive data such as vehicle position, speed, and acceleration collected by inertial measurement units; the collision risk nodes calculate the collision risk probability based on the data of the environmental observation nodes and the vehicle state nodes. The environmental observation nodes and the vehicle state nodes are the parent nodes, and the collision risk nodes are the child nodes. Their connection relationships are determined according to the actual physical meaning and causal relationships.

[0072] Update the evidence probability of the environmental observation nodes based on real-time sensor data. For example, when the millimeter-wave radar detects new obstacle distance and speed data, update the probability distributions of the corresponding obstacle distance and speed in the environmental observation nodes according to these data. Perform state estimation on the vehicle state nodes through the Kalman filter algorithm. The Kalman filter algorithm is an optimal linear recursive estimation algorithm. Through this algorithm, the posterior probability distributions of vehicle position, speed, and acceleration are obtained, so as to more accurately reflect the actual state of the vehicle.

[0073] Introduce a dynamic time window mechanism to dynamically adjust the inference time interval of the Bayesian network according to the vehicle movement trend. Let the vehicle speed be v and the time window length be T. Define their negative correlation relationship as (where C is a constant and ∈ is an extremely small positive number used to avoid the denominator being zero). When the vehicle speed is fast, shorten the inference time interval to update the collision risk assessment more timely; when the vehicle speed is slow, appropriately extend the inference time interval to reduce the consumption of computing resources.

[0074] Adopt the variational inference algorithm to perform online parameter update on the Bayesian network. The core idea of variational inference is to find an approximate distribution q(θ) to approximate the true posterior distribution p(θ|X), and optimize the conditional probability table parameters by minimizing the evidence lower bound (ELBO). The calculation formula of the evidence lower bound is:

[0075] ELBO = E q(θ) [logp(X|θ)] - KL(q(θ)||p(θ))

[0076] where, E q(θ) [logp(X|θ)] is the expectation of the log-likelihood, and KL(q(θ)||p(θ)) is the KL divergence between q(θ) and p(θ). By continuously iteratively optimizing q(θ), the evidence lower bound is maximized, so as to realize the optimization of the conditional probability table parameters of the Bayesian network and improve the accuracy of the model.

[0077] Example 2:

[0078] This embodiment details the specific implementation of the improved particle swarm optimization algorithm by introducing an adaptive inertia weight and a dynamic learning factor to improve the efficiency and quality of searching for the optimal obstacle avoidance path in the multi-objective path optimization model. The specific steps are as follows:

[0079] Construct a particle swarm optimization objective function that includes a path obstacle avoidance safety term and a path smoothness term. Let the set of path points be {P 1 , P 2 , …, P n}, and the set of obstacles be {O 1 , O 2 , …, O m}. The path obstacle avoidance safety term is calculated by the minimum distance between the path point and the obstacle, and the formula is:

[0080]

[0081] where f safe is the evaluation value of the path obstacle avoidance safety term, and d(P i , O j ) represents the distance between the path point P i and the obstacle O j . The path smoothness term is calculated by the change in curvature between adjacent path points. Let the adjacent path points be P i , P i+1 , P i+2 , and the formula for the curvature k i is:

[0082]

[0083] The formula for the path smoothness term f smooth is:

[0084]

[0085] Then the particle swarm optimization objective function is F = w 1 f safe + w 2 f smooth , where w 1 , w 2 are weight coefficients, and their values are adjusted according to the actual situation to balance the importance of obstacle avoidance safety and path smoothness.

[0086] Design an adaptive inertia weight update rule: Design an adaptive inertia weight update rule where the inertia weight w decays linearly in segments as the number of iterations t increases. Let the maximum number of iterations be T max , and divide the iteration process into three stages. The first stage is 0 ≤ t ≤ 0.3T max , The second stage is 0.3T max<t ≤ 0.7T max , w = w mid ; The third stage is 0.7T max <t ≤ T max , where w max , w mid , w min are the initial, intermediate, and final inertia weight values respectively. In the initial stage, a larger inertia weight enables the particles to have a higher global search ability and can quickly explore the solution space; in the later stage, a smaller inertia weight enhances the local search accuracy and enables the particles to focus more on mining the local optimal solution.

[0087] Design a dynamic learning factor adjustment mechanism: Design a dynamic learning factor adjustment mechanism. The learning factors include the individual learning factor c 1 and the social learning factor c 2 . The population diversity index is measured by the fitness variance where N is the population size, F i is the fitness value of the i-th particle, is the population average fitness value. Set a threshold σ thresh , when σ 2 < σ thresh , increase the weight of the social learning factor; when σ 2 ≥ σ thresh , increase the weight of the individual learning factor. In this way, the learning factors are dynamically adjusted according to the population diversity to avoid the algorithm falling into local optimum.

[0088] Example 3:

[0089] This example details the specific process of mapping the warning instruction to the braking pressure and the audible and visual alarm signals by the fuzzy logic controller, realizing more reasonable and accurate warning control.

[0090] Construct an input fuzzification module. The inputs include the collision risk probability value P, the relative obstacle distance d, and the current vehicle speed v. The triangular membership function is used to fuzzify the input variables. Taking the collision risk probability value as an example, assuming its value range is [0, 1], it is fuzzified into three fuzzy sets: "low", "medium", and "high". For the "low" fuzzy set, the triangular membership function is:

[0091]

[0092] For the "medium" fuzzy set:

[0093]

[0094] For the "high" fuzzy set:

[0095]

[0096] Similarly, corresponding fuzzification processing is also carried out on the relative obstacle distance and the current vehicle speed, and the number of fuzzy sets and the membership function parameters are determined according to the actual situation.

[0097] Construct a fuzzy rule base containing IF-THEN rules. For example, Rule 1: IF the collision risk probability value is "high", the relative obstacle distance is "near", and the current vehicle speed is "fast", THEN the braking pressure gradient is "large" and the audible and visual alarm frequency is "high"; Rule 2: IF the collision risk probability value is "medium", the relative obstacle distance is "medium", and the current vehicle speed is "medium", THEN the braking pressure gradient is "medium" and the audible and visual alarm frequency is "medium", etc. According to the actual driving experience and safety requirements, a series of reasonable fuzzy rules are established to achieve accurate control of different situations.

[0098] Design a defuzzification module, and use the centroid method to convert the fuzzy output into an accurate control quantity. Let the fuzzy output variable be y, and its fuzzy set be {A 1 , A 2 , …, A n}, and the corresponding membership values are The calculation formula of the centroid method is:

[0099]

[0100] where y i is the central value of the fuzzy set A i . Through this formula, the fuzzy braking pressure gradient and audible and visual alarm frequency are converted into accurate control quantities for actual vehicle control.

[0101] Introduce a feedback correction mechanism to dynamically adjust the confidence of fuzzy rules according to the actual braking effect and the change rate of the obstacle distance. Let the actual braking acceleration be a actual , the expected braking acceleration be a desired , and the change rate of the obstacle distance be Define the confidence adjustment coefficient α of the fuzzy rule. When |a actual - a desired | > δ 1 and (δ 1 , δ 2 are set thresholds), reduce the confidence of the corresponding fuzzy rule; conversely, when |a actual - a desired | ≤ δ 1 and When it is, increase the confidence of the corresponding fuzzy rule. By continuously adjusting the confidence of the fuzzy rule, the fuzzy logic controller can better adapt to the actual situation and improve the control effect.

[0102] Example 4:

[0103] This embodiment details the data fusion processing method of the multi-source sensor module, improves the accuracy and reliability of sensor data, and provides more reliable data support for subsequent analysis and decision-making. The specific steps include:

[0104] For millimeter-wave radar data, a constant false alarm rate detection algorithm is used to extract obstacle distance and speed information. The constant false alarm rate detection algorithm performs statistical processing on the received signal by setting reference cells and guard cells. Let the number of reference cells be N 1 , and the number of guard cells be N 2 , the received signal power be x, and the estimated noise power be Using the mean type constant false alarm rate detection algorithm, the detection threshold T is:

[0105]

[0106] where λ is a constant determined according to the false alarm probability, and x i is the signal power of the reference cell. When the received signal power x is greater than the detection threshold T, it is determined that there is an obstacle, and the obstacle distance and speed are calculated according to the radar signal processing algorithm.

[0107] For binocular vision data, a stereo matching algorithm is used to generate a depth map. The stereo matching algorithm calculates depth information by finding the disparity of corresponding pixel points in the left and right images. The block-based matching algorithm is used. Taking a certain pixel point as the center, windows of the same size are selected in the left and right images, and the matching cost (such as the sum of absolute differences SAD, normalized cross-correlation NCC, etc.) between the windows is calculated to determine the matching points. For example, using the SAD algorithm, the matching cost calculation formula is:

[0108]

[0109] where I l and I r are the pixel values of the left and right images respectively, (x, y) is the current pixel point coordinate, d is the disparity, and w, h are the window sizes. By finding the disparity corresponding to the minimum matching cost, the depth value of each pixel point is calculated to generate a depth map. Then, noise is removed through morphological filtering, and morphological filtering includes erosion and dilation operations. The erosion operation can eliminate small noise points in the image, and the dilation operation can restore the object edges that have been eroded.

[0110] For ultrasonic sensor data, a moving average filter is used to eliminate short-term interference. Let the data sequence collected by the ultrasonic sensor be x1 , x 2 , …, x n , the window size of the moving average filter is m (m < n). At time k, the filtered output value y k is:

[0111]

[0112] For example, when the window size m = 5, for the ultrasonic data at the current time k, the data at the 5 time instants from k - 4 to k are averaged to obtain the filtered result. As time goes by, the window slides continuously, and the data is continuously smoothed, effectively reducing the short-term interference caused by environmental noise and other factors, making the data more stable and reliable.

[0113] The quaternion complementary filtering algorithm is used to fuse the gyroscope and accelerometer data of the inertial measurement unit, and the vehicle attitude angle and angular velocity are output.

[0114] Example 5:

[0115] This example elaborates in detail the specific method for the decision-making layer of the hierarchical early warning control model to generate multi-level early warning strategies. The specific method includes:

[0116] The risk levels are divided into three levels: low, medium, and high. The low risk level corresponds to audible and visual warnings. When the collision risk probability value is in the low risk interval, the system only reminds the driver to pay attention to the surrounding environment through audible and visual alarms, such as emitting a low-frequency beeping sound and flashing warning lights. The medium risk level corresponds to pre-pressurization of the brakes. At this time, the system pre-pressurizes the braking system to a certain extent according to the collision risk probability and vehicle state, shortening the braking response time and improving braking safety. The high risk level triggers full braking. When the collision risk is extremely high, the full braking operation is immediately started to reduce the vehicle speed to the greatest extent and avoid collision accidents.

[0117] A state machine model is constructed, and the state machine transfers states according to continuous risk level changes. Let the risk level of the current frame be R k , and the risk level of the previous frame be R k-1 . If R k > R k-1 and the data of two consecutive frames satisfy this condition, then it jumps to a higher early warning level. For example, if the current frame is a low risk, the next frame becomes a medium risk, and the frame after that is still a medium risk or a high risk, then the early warning level is raised to a medium risk or a high risk.

[0118] A delay trigger mechanism is introduced. When the high risk level lasts for more than the preset time threshold T highWhen the high - risk level is detected, start the emergency obstacle avoidance protocol. During the duration of the high - risk level, the system continuously monitors the vehicle status and environmental data. If the risk has not been lifted after a preset time, the emergency obstacle avoidance protocol will be started, such as automatically controlling the vehicle's steering, braking, etc. to avoid collisions.

[0119] Build an online self - checking module to detect sensor failures through cross - verification of redundant sensor data. Assume there are multiple sensors of the same type in the system (such as multiple millimeter - wave radars). For the data x collected by a certain sensor i , calculate the mean x and standard deviation σ of all sensor data of this type. If |x i - x|>3σ, then mark the sensor data as abnormal. For example, at a certain moment, multiple millimeter - wave radars measure the distance to the same obstacle. If the measurement value of one radar deviates too much from the mean of the measurement values of other radars, exceeding three times the standard deviation, the system determines that this radar may have a fault, marks it in time and takes corresponding measures, such as switching to a backup sensor or sending a fault alarm.

[0120] Adopt a hot - backup mechanism to switch to a backup controller when the main control unit fails. The main control unit and the backup controller run simultaneously. The main control unit sends its own status information to the backup controller through a heartbeat packet. A heartbeat packet is a short message sent regularly, containing information such as the operating status and data - processing progress of the main control unit. If the backup controller does not receive the heartbeat packet from the main control unit within a preset time, it determines that the main control unit has failed and immediately switches to the working mode of the backup controller. During the switching process, through a synchronization mechanism, ensure that the backup controller obtains system status data consistent with that of the main control unit, such as the vehicle's current position, speed, warning status, etc., to ensure the uninterrupted operation of the system and improve the reliability and stability of the system.

[0121] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non - exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0122] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A safety collision warning method for an electric tricycle, characterized in that: include: The real-time environmental data and vehicle status data of the electric tricycle are collected through a multi-source sensor module, wherein the multi-source sensor module includes a millimeter wave radar, a binocular vision sensor, an ultrasonic sensor, and an inertial measurement unit; Based on a convolutional neural network, multi-scale feature extraction and fusion are performed on the real-time environmental data to generate an environmental feature vector; the environmental feature vector is input into a pre-trained Bayesian network model, and the Bayesian network model adopts a dynamic node update mechanism to adjust the conditional probability distribution based on the real-time observation data and output a collision risk probability value; A multi-objective path optimization model is constructed according to the collision risk probability value, wherein the multi-objective path optimization model takes the highest obstacle avoidance priority and the smallest path curvature as optimization goals, and uses an improved particle swarm optimization algorithm to perform a global search for feasible paths, wherein the improved particle swarm optimization algorithm introduces an adaptive inertia weight and a dynamic learning factor; Outputting optimal obstacle avoidance path data based on the multi-objective path optimization model; A hierarchical warning control model is established based on the optimal obstacle avoidance path data, and the hierarchical warning control model includes a data layer, a decision layer and an execution layer, wherein the data layer updates the risk level based on the environmental feature vector, the decision layer generates a multi-level warning strategy based on the collision risk probability value, and the execution layer maps the warning instructions into braking pressure and sound and light alarm signals through a fuzzy logic controller.

2. The method according to claim 1, characterized in that The Bayesian network model adopts a dynamic node update mechanism including: Constructing a Bayesian network structure, wherein the Bayesian network structure includes an environment observation node, a vehicle status node, and a collision risk node, wherein the environment observation node and the vehicle status node are parent nodes, and the collision risk node is a child node; Based on the real-time sensor data, the evidence probability of the environment observation node is updated, and the state of the vehicle state node is estimated by the Kalman filter algorithm to obtain the posterior probability distribution of the vehicle position, speed and acceleration; A dynamic time window mechanism is introduced to dynamically adjust the inference time interval of the Bayesian network according to the vehicle movement trend. The length of the time window is negatively correlated with the vehicle speed. The variational inference algorithm is used to perform online parameter update on the Bayesian network, and the conditional probability table parameters are optimized by minimizing the lower bound of evidence.

3. The method according to claim 1, characterized in that The improved particle swarm optimization algorithm introduces adaptive inertia weight and dynamic learning factor including: Constructing a particle swarm optimization objective function, the objective function includes a path obstacle avoidance safety term and a path smoothness term, the path obstacle avoidance safety term is calculated by the minimum distance between the path point and the obstacle, and the path smoothness term is calculated by the curvature change of adjacent path points; Design an adaptive inertia weight update rule, where the inertia weight decays piecewise linearly with the number of iterations, retains a high global search capability in the initial stage, and enhances local search accuracy in the later stage; A dynamic learning factor adjustment mechanism is designed. The learning factor is dynamically adjusted according to the population diversity index. When the population diversity is lower than the threshold, the weight of the social learning factor is increased, otherwise the weight of the individual learning factor is increased.

4. The method according to claim 1, characterized in that: The fuzzy logic controller maps the warning instruction into brake pressure and sound and light alarm signals, including: Construct an input fuzzification module, wherein the input includes a collision risk probability value, a relative obstacle distance, and a current vehicle speed, and a triangle membership function is used to perform fuzzification processing on the input variables; Building a fuzzy rule base, wherein the rule base includes IF-THEN rules; Design a defuzzification module and use the centroid method to convert the fuzzy output into precise control quantities, including the brake pressure gradient and the sound and light alarm frequency; A feedback correction mechanism is introduced to dynamically adjust the fuzzy rule confidence according to the actual braking effect and the rate of change of obstacle distance.

5. The method according to claim 1, characterized in that The data fusion processing of the multi-source sensor module includes: A constant false alarm rate detection algorithm is used to extract obstacle distance and speed information from millimeter wave radar data; A stereo matching algorithm is used to generate a depth map for binocular vision data, and morphological filtering is used to remove noise; Sliding average filtering is used to eliminate short-term interference on ultrasonic sensor data; The quaternion complementary filtering algorithm is used to fuse the gyroscope and accelerometer data of the inertial measurement unit to output the vehicle attitude angle and angular velocity.

6. The system according to claim 1, characterized in that The decision layer of the hierarchical early warning control model generates a multi-level early warning strategy including: The risk level is divided into three levels: low, medium and high. The low risk level corresponds to sound and light warning, the medium risk level corresponds to brake pre-pressurization, and the high risk level triggers full braking; Construct a state machine model, wherein the state machine performs state transition according to continuous risk level changes, and jumps to a higher warning level if the risk level of two consecutive frames of data increases; A delayed trigger mechanism is introduced to initiate the emergency obstacle avoidance protocol when the high risk level continues to exceed the preset time threshold.

7. The system according to claim 1, characterized in that The constraints of the multi-objective path optimization model include: Path curvature constraint, which limits the maximum curvature of a path point to a curvature threshold corresponding to the minimum turning radius of the vehicle; Dynamic obstacle avoidance constraints, predicting obstacle trajectories and generating avoidance corridors through the speed obstacle method; Actuator dynamics constraints define the physical upper limits of the brake pressure change rate and steering angle velocity.

8. The system according to claim 1, characterized in that The system further comprises: An online self-check module is built to detect sensor failures through cross-validation of redundant sensor data. If a sensor data deviates from the mean by more than three times the standard deviation, it is marked as abnormal. A hot backup mechanism is adopted. When the main control unit fails, it switches to the backup controller and maintains the synchronization of the communication status of the two machines through heartbeat packets.

9. An electric tricycle safety collision warning system, characterized in that: include: Multi-source sensor module: used to collect real-time environmental data and vehicle status data of the electric tricycle, including millimeter-wave radar, binocular vision sensor, ultrasonic sensor and inertial measurement unit; Feature extraction and fusion module: Based on the convolutional neural network, multi-scale feature extraction and fusion are performed on the real-time environmental data collected by the multi-source sensor module to generate an environmental feature vector; Collision risk assessment module: The environmental feature vector is input into the pre-trained Bayesian network model, which adopts a dynamic node update mechanism to adjust the conditional probability distribution based on real-time observation data and output the collision risk probability value; Path optimization module: A multi-objective path optimization model is constructed based on the collision risk probability value, with the highest obstacle avoidance priority and the smallest path curvature as the optimization goal. An improved particle swarm optimization algorithm that introduces adaptive inertia weights and dynamic learning factors is used to perform a global search for feasible paths, and the optimal obstacle avoidance path data is output; Hierarchical warning control module: established according to the optimal obstacle avoidance path data, including data layer, decision layer and execution layer. The data layer updates the risk level based on the environmental feature vector, the decision layer generates a multi-level warning strategy based on the collision risk probability value, and the execution layer maps the warning instructions into braking pressure and sound and light alarm signals through the fuzzy logic controller.

10. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the functions of the method described in any one of claims 1 to 8.

Images