Vehicle safety management intelligent scheduling system based on artificial intelligence

Abstract

The invention discloses a vehicle safety management intelligent scheduling system based on artificial intelligence, and relates to the technical field of vehicle management scheduling. The vehicle safety management intelligent scheduling system based on artificial intelligence comprises a time sequence data acquisition unit used for acquiring vehicle driving state time sequence data based on a set time period during vehicle driving, and a time sequence analysis network used for carrying out feature extraction on the vehicle driving state time sequence data based on pre-training, the timing sequence feature extraction unit is used for collecting vehicle driving state timing sequence data to obtain a vehicle driving state timing sequence feature set, and the timing sequence index analysis unit is used for comprehensively analyzing the vehicle driving state timing sequence feature set to obtain a vehicle driving state feature index set. The system completely records the change track of the interaction state of the vehicle power and the driver, the defect that the causal relationship is neglected in a traditional method is avoided, a more stable basic information flow is provided, and the monitoring continuity of the system on the dynamic driving state is enhanced.

Description

Technical Field

[0001] This invention relates to the field of vehicle management and dispatching technology, specifically to an intelligent dispatching system for vehicle safety management based on artificial intelligence. Background Technology

[0002] With the rapid development of the automotive industry, intelligent vehicle technologies have been widely applied. Vehicle safety has become a major concern, especially in the context of technologies such as autonomous driving, vehicle-to-everything (V2X) communication, and intelligent transportation. Traditional vehicle safety management methods are struggling to meet the increasingly complex road environments and driving demands. Therefore, more and more research and development is exploring artificial intelligence-based vehicle safety management systems to achieve more efficient and accurate vehicle safety monitoring and management.

[0003] Traditional vehicle safety management relies primarily on driver experience and human intervention, or on basic sensors monitoring vehicle status such as speed, fuel level, and tire pressure. However, this approach lacks the ability to respond in real-time to complex driving behaviors and environmental changes, and it cannot perform in-depth data mining and intelligent decision-making. In recent years, artificial intelligence technologies, especially machine learning and deep learning, have gradually become core technologies for improving vehicle safety management capabilities.

[0004] Although some existing intelligent vehicle safety management systems have made some progress, most systems still rely on relatively basic algorithms or sensors and lack deeper multimodal data fusion and intelligent decision-making capabilities.

[0005] The limitations of existing technologies include at least the following problems: traditional vehicle safety management systems usually rely on static data or simple real-time monitoring, but lack dynamic analysis of vehicle driving status. They often only focus on measuring the vehicle's status at a certain moment, such as speed or position, and find it difficult to capture dynamic changes during vehicle driving. In particular, when the vehicle's status changes abruptly, such as sudden acceleration or emergency braking, the system is unable to respond in time.

[0006] When a vehicle is in a complex traffic environment, especially in an emergency or under complex road conditions, the vehicle's driving state can change rapidly. Traditional systems often rely on static data collection or judgments based on fixed rules, making it difficult to accurately analyze the evolution of the vehicle's state and its interaction with the driver's behavior. This lack of flexibility makes it difficult to detect potential risks in real time at critical moments, such as when there is sudden braking or fluctuations in vehicle stability, and it is difficult to make timely safety interventions, thereby increasing the risk of traffic accidents. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an intelligent dispatch system for vehicle safety management based on artificial intelligence, which solves the problem of difficulty in real-time monitoring and analysis of vehicle dynamic driving status and changes in driver interaction behavior in existing technologies.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an intelligent dispatch system for vehicle safety management based on artificial intelligence, comprising: a time-series data acquisition unit, used to acquire time-series data of vehicle driving status based on a set time period during vehicle operation, including time-series data of vehicle power and driver interaction status; a time-series feature extraction unit, used to extract features from the vehicle driving status time-series data based on a pre-trained time-series analysis network to obtain a vehicle driving status time-series feature set, including time-series features of vehicle power and driver interaction status; a time-series index analysis unit, used to comprehensively analyze the vehicle driving status time-series feature set to obtain a vehicle driving status feature index set, including time-series indices of vehicle power and driver interaction status; and an intelligent dispatch unit, used to execute intelligent dispatch measures for the vehicle based on the vehicle driving status feature index set.

[0009] Furthermore, the temporal analysis network includes a vehicle dynamics temporal feature extraction network and a driving interaction feature extraction network. The vehicle dynamics temporal feature extraction network includes an input encoding layer, a multi-scale convolutional feature extraction layer, a bidirectional gated recurrent unit layer, and a statistical mapping layer. The driving interaction feature extraction network includes a signal alignment embedding layer, a sliding convolutional fluctuation detection layer, a Transformer self-attention analysis layer, and a statistical mapping and offset analysis layer.

[0010] Furthermore, the vehicle dynamics time series data includes vehicle linear velocity values, vehicle lateral acceleration values, and vehicle braking pressure values ​​at several time points. The vehicle dynamics time series characteristics include the vehicle average linear velocity change rate, the magnitude of extreme changes in vehicle linear velocity, the vehicle maximum lateral acceleration value, the frequency of sudden changes in vehicle lateral acceleration, and the vehicle maximum braking pressure value.

[0011] Further, the specific steps for obtaining vehicle power temporal features are as follows: In the input encoding layer of the vehicle power temporal feature extraction network, the vehicle power temporal data is processed to construct a temporal joint vector to obtain a standardized vehicle power input sequence; in the multi-scale convolutional feature extraction layer of the vehicle power temporal feature extraction network, the standardized vehicle power input sequence is processed by multi-scale convolution to obtain a local dynamic structure feature sequence; in the bidirectional gated recurrent unit layer of the vehicle power temporal feature extraction network, the local dynamic structure feature sequence is processed by time-dependent modeling to obtain a vehicle power temporal embedding feature sequence; in the statistical mapping layer of the vehicle power temporal feature extraction network, extreme value analysis, mutation frequency identification, and change rate analysis are performed on the vehicle power temporal embedding feature sequence to obtain vehicle power temporal features.

[0012] Furthermore, the specific steps to obtain the vehicle power timing index are as follows: read the vehicle power timing characteristics; perform a comprehensive analysis of the vehicle power timing characteristics to obtain the vehicle power timing index.

[0013] Furthermore, the time-series data of the driver interaction state includes steering wheel grip force value, cabin carbon dioxide concentration value and driver heart rate value at several time points. The time-series characteristics of the driver interaction state include steering wheel grip force fluctuation amplitude value, steering wheel grip force disturbance frequency, cabin carbon dioxide concentration rise rate, cabin carbon dioxide concentration fluctuation average amplitude and driver heart rate average deviation.

[0014] Further, the specific steps for obtaining the temporal features of driver interaction states are as follows: In the signal alignment embedding layer of the driver interaction feature extraction network, the temporal data of driver interaction states is standardized to obtain a standardized driver interaction input sequence; in the sliding convolutional fluctuation detection layer of the driver interaction feature extraction network, the standardized driver interaction input sequence is processed by sliding window one-dimensional convolution to obtain a local interaction fluctuation feature sequence; in the Transformer self-attention analysis layer of the driver interaction feature extraction network, multi-head self-attention weight analysis is performed on the local interaction fluctuation feature sequence to obtain a driver interaction temporal embedding feature sequence; in the statistical mapping and offset analysis layer of the driver interaction feature extraction network, fluctuation amplitude analysis, disturbance frequency counting, rise rate and mean offset analysis are performed on the driver interaction temporal embedding feature sequence to obtain the temporal features of driver interaction states.

[0015] Furthermore, the specific steps to obtain the driver interaction state timing index are as follows: read the timing characteristics of the driver interaction state; perform comprehensive analysis on the timing characteristics of the driver interaction state to obtain the driver interaction state timing index.

[0016] Furthermore, the specific formula for calculating the driver interaction state timing index is as follows: ;in, , , , , , The parameters, in order, are: driver interaction state timing index, steering wheel grip force fluctuation amplitude, steering wheel grip force disturbance frequency, cabin carbon dioxide concentration rise rate, cabin carbon dioxide concentration fluctuation average amplitude, and driver heart rate average deviation. It is a natural constant. , , The coefficients are, in order, the carbon dioxide elevation adjustment coefficient, the carbon dioxide fluctuation adjustment coefficient, and the heart rate deviation adjustment coefficient stored in the database.

[0017] Furthermore, based on the vehicle driving state characteristic index set, the specific steps for implementing intelligent dispatching measures for vehicles are as follows: The vehicle driving state characteristic index set is compared with several preset sets of driving dispatch intervals for judgment and analysis. Each set of driving dispatch intervals includes a vehicle power interval and a driver interaction state interval, and each set of driving dispatch intervals corresponds to an intelligent dispatching strategy; based on the intelligent dispatching strategy corresponding to when the vehicle driving state characteristic index set is within the preset set of driving dispatch intervals, intelligent dispatching processing is performed on the vehicle.

[0018] The present invention has the following beneficial effects: (1) The intelligent dispatch system for vehicle safety management based on artificial intelligence, through the set time-series data acquisition unit, can continuously acquire time-series data related to vehicle power and driver interaction status at a set time period during vehicle driving, and form a multi-channel serialized input format. Unlike the traditional solution that judges the vehicle status based on a certain time point or non-continuous sampling data, this method can completely retain the evolution trajectory of vehicle acceleration and deceleration, direction change, driving operation fluctuations, etc., which helps to identify features such as sudden changes in status and dangerous trends in advance. In scenarios with complex road conditions or drastic changes in driving behavior, if only fragment data is relied upon, it is easy for the system to ignore the cause-and-effect relationship and misjudge the risk level. Periodic time-series acquisition can solve this defect, provide a more stable basic information flow for subsequent feature extraction and intelligent dispatch, enhance the system's expression of dynamic driving status and monitoring continuity, thereby improving the reliability of the overall judgment.

[0019] (2) This AI-based intelligent dispatch system for vehicle safety management achieves in-depth analysis and feature compression of vehicle driving status data through the set temporal feature extraction unit and temporal index analysis unit. The system not only extracts numerical features of vehicle dynamic status, such as speed changes, acceleration fluctuations, and braking intensity, but also extracts temporal features of driver interaction-related behaviors, such as grip strength fluctuations and attention changes, and finally analyzes the corresponding state index. Compared with traditional systems that rely on rule thresholds or single-variable logical judgments, this system is conducive to uniformly applying judgment criteria under different vehicle types and road scenarios. (3) The intelligent dispatch system for vehicle safety management based on artificial intelligence automatically executes dispatch response based on the generated vehicle driving status characteristic index set through the set intelligent dispatch unit. It does not rely on manual intervention or remote server instructions. Unlike traditional systems that require drivers to make their own judgments or wait for the platform to issue instructions, this system can directly judge whether the driving strategy needs to be adjusted based on local indices when the vehicle status changes suddenly or the driving behavior is abnormal. This realizes the closed-loop and localized processing of the dispatch control path. This mechanism is particularly effective on highways or mountain roads where communication conditions are limited. It can independently complete status perception and dispatch response, reducing the safety hazards caused by information transmission delay.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] Figure 1 This is a block diagram of the intelligent dispatch system for vehicle safety management based on artificial intelligence, as described in this invention.

[0022] Figure 2 This is a flowchart illustrating the specific steps involved in obtaining vehicle power timing characteristics in the AI-based intelligent scheduling system for vehicle safety management according to the present invention.

[0023] Figure 3 This is a flowchart illustrating the specific steps involved in obtaining the temporal characteristics of driver interaction states in the intelligent dispatch system for vehicle safety management based on artificial intelligence, as described in this invention. Detailed Implementation

[0024] Please see Figure 1 This invention provides a technical solution: an intelligent dispatch system for vehicle safety management based on artificial intelligence, comprising: a time-series data acquisition unit, used to acquire time-series data of vehicle driving status, including vehicle power and driver interaction status, based on a set time period (e.g., a time period of five seconds and a time interval of 0.2 seconds); a time-series feature extraction unit, used to extract features from the vehicle driving status time-series data based on a pre-trained time-series analysis network to obtain a vehicle driving status time-series feature set, including time-series features of vehicle power and driver interaction status; a time-series index analysis unit, used to comprehensively analyze the vehicle driving status time-series feature set to obtain a vehicle driving status feature index set, including time-series indices of vehicle power and driver interaction status; and an intelligent dispatch unit, used to execute intelligent dispatch measures for the vehicle based on the vehicle driving status feature index set.

[0025] The temporal analysis network includes a vehicle dynamics temporal feature extraction network and a driving interaction feature extraction network. The vehicle dynamics temporal feature extraction network includes an input encoding layer, a multi-scale convolutional feature extraction layer, a bidirectional gated recurrent unit layer, and a statistical mapping layer. The driving interaction feature extraction network includes a signal alignment embedding layer, a sliding convolutional fluctuation detection layer, a Transformer self-attention analysis layer, and a statistical mapping and offset analysis layer.

[0026] Specifically, the vehicle dynamics time series data includes vehicle linear velocity values, vehicle lateral acceleration values, and vehicle braking pressure values ​​at several time points. The vehicle dynamics time series characteristics include the vehicle average linear velocity change rate, the magnitude of extreme changes in vehicle linear velocity, the vehicle maximum lateral acceleration value, the frequency of lateral acceleration changes, and the vehicle maximum braking pressure value.

[0027] Among them, the vehicle linear velocity value refers to the instantaneous linear velocity of the vehicle's center of gravity along the driving direction. It is used to characterize the current driving speed and power output level. It can be obtained by measuring the wheel speed sensors installed on the four wheels and combined with the wheel radius to obtain the linear velocity (i.e., the wheel speed encoder signal of any wheel multiplied by the wheel radius to obtain the linear velocity).

[0028] The lateral acceleration value of a vehicle refers to the instantaneous acceleration generated by the vehicle under the action of lateral inertia. It is used to reflect the stability state during lane changes, sharp turns, or sideslips. It can be measured and obtained by a triaxial acceleration sensor installed at the center of gravity of the vehicle, and its Y-axis component is the lateral acceleration value of the vehicle.

[0029] The vehicle braking pressure value refers to the current actual hydraulic pressure value in the brake master cylinder, which reflects the force input intensity during the braking process. It can be measured and obtained by a resistive hydraulic pressure sensor installed on the brake hydraulic manifold.

[0030] like Figure 2 As shown, the specific steps to obtain the vehicle power temporal features are as follows: In the input encoding layer of the vehicle power temporal feature extraction network, the vehicle power temporal data is processed to construct a temporal joint vector to obtain a standardized vehicle power input sequence. Specifically, the vehicle linear velocity value, vehicle lateral acceleration value, and vehicle braking pressure value are read sequentially at each time point according to the time order, and the three are concatenated to form a three-dimensional temporal vector. Then, the interval normalization method is used to map the linear velocity value to the preset maximum speed, the lateral acceleration value to ±1g, and the braking pressure value to the maximum rated pressure of the brake master cylinder to the [0,1] interval, and after normalization, the entire sequence is normalized to zero mean-unit variance to obtain a standardized vehicle power input sequence. In the multi-scale convolutional feature extraction layer of the vehicle dynamic temporal feature extraction network, the standardized vehicle dynamic input sequence is processed by multi-scale convolution to obtain the local dynamic structure feature sequence. Specifically, three sets of one-dimensional convolutional kernels with kernel sizes of 3, 5, and 7 are used in parallel to perform convolution operations on each channel to extract the speed, acceleration, and braking change patterns in the short-term, sub-short-term, and medium-term domains. After applying ReLU activation to the convolutional output, max pooling is performed to retain abrupt peaks and valleys. Then, the three sets of convolution-pooling results are concatenated according to the channel dimension to form the local dynamic structure feature sequence. In the bidirectional gated recurrent unit layer of the vehicle power temporal feature extraction network, the local dynamic structure feature sequence is processed by time-dependent modeling to obtain the vehicle power temporal embedding feature sequence. Specifically, the local dynamic structure feature sequence is first input into the forward GRU and backward GRU branches to calculate the hidden state vectors in the forward and backward time directions respectively. Then, the forward and backward hidden states are concatenated step by step to form a bidirectional gated recurrent output containing complete context information. The output is then subjected to time-dimensional global average pooling to obtain the vehicle power temporal embedding feature sequence representing the entire sequence. In the statistical mapping layer of the vehicle dynamics temporal feature extraction network, extreme value analysis, mutation frequency identification, and change rate analysis are performed on the vehicle dynamics temporal embedded feature sequence to obtain the vehicle dynamics temporal features. Specifically, the following steps are taken: First, the absolute value of the time gradient component of the corresponding linear velocity channel in the embedded feature is taken and then averaged to obtain the vehicle's average linear velocity change rate; second, the difference between the maximum and minimum values ​​of the original values ​​of the linear velocity channel in the embedded feature is taken to obtain the extreme mutation amplitude of the vehicle's linear velocity; third, the absolute value of the lateral acceleration channel in the embedded feature is taken and the maximum value is obtained to obtain the vehicle's maximum lateral acceleration value; then, the peak points in the lateral acceleration channel that exceed the 0.3g threshold are counted and normalized to the frequency per unit time to obtain the frequency of lateral acceleration mutations; finally, the maximum value of the braking pressure channel is taken directly to obtain the vehicle's maximum braking pressure value.

[0031] The pre-training steps for the vehicle dynamic temporal feature extraction network are as follows: First, training samples were collected based on a large-scale labeled dataset. This dataset contains a large number of time-series driving data samples under normal vehicle operation and potential risk conditions. Each sample consists of a time series of vehicle linear velocity, lateral acceleration, and braking pressure values, and is simultaneously labeled with its corresponding driving state classification label or dynamic characteristic index reference value. To improve the model's generalization ability, the training data collected during the collection phase covered various vehicle speed ranges, different braking levels, turning radii, and road types. All time-series data were sampled synchronously at set time steps and underwent uniform normalization and outlier removal operations.

[0032] The standardized training data is then input into the vehicle dynamics temporal feature extraction network, sequentially passing through an input encoding layer, a multi-scale convolutional feature extraction layer, a bidirectional gated recurrent unit layer, and a statistical mapping layer. During training, a supervised learning strategy is employed, using a loss function to constrain the error between the extracted temporal features and the true labels or reference feature values. If the labels are categorical information, a cross-entropy loss function is used to train the feature classification ability; if the labels are continuous feature values, mean squared error (MSE) is used as the regression loss function. Simultaneously, to prevent overfitting, a Dropout layer and L2 regularization mechanism are introduced between the convolutional and recurrent unit layers. Training uses the Adam optimizer for gradient descent, and the network's recognition accuracy under multiple typical dynamic states is monitored in real time using a validation set.

[0033] After completing the basic training, multiple rounds of fine-tuning training are performed to retrain the network on datasets with different vehicle models, driving styles, and road types to improve the generalization robustness and transferability of the feature extraction network. Finally, the parameters of the trained model are fixed, and this model is used as a pre-trained model for the vehicle dynamics temporal feature extraction network.

[0034] The specific steps to obtain the vehicle power timing index are as follows: read the vehicle power timing characteristics; perform a comprehensive analysis of the vehicle power timing characteristics to obtain the vehicle power timing index.

[0035] The specific formula for calculating the vehicle power timing index is as follows: ;in, , , , , , In order, they are: vehicle dynamics time series index, vehicle average linear velocity change rate, vehicle linear velocity extreme value abrupt change amplitude, vehicle maximum lateral acceleration value, vehicle lateral acceleration abrupt change frequency, and vehicle maximum braking pressure value. , , , , The coefficients stored in the database are, in order: speed change adjustment coefficient, speed change adjustment coefficient, lateral acceleration adjustment coefficient, lateral acceleration frequency adjustment coefficient, and braking pressure adjustment coefficient.

[0036] It should be explained that the specific steps for obtaining the speed change adjustment coefficient, speed change adjustment coefficient, lateral acceleration adjustment coefficient, lateral acceleration frequency adjustment coefficient, and braking pressure adjustment coefficient stored in the database are as follows: For the speed change adjustment coefficient: at least 5000 vehicle speed time series segments, each with a duration of 20 seconds, are randomly selected from the historical database. All data are sampled at a uniform frequency of 10Hz. For each segment, the speed change between adjacent sampling points (i.e., the speed at the current moment minus the speed at the previous moment) is calculated sequentially, and the absolute values ​​of all speed changes are averaged to obtain the average speed change rate of the sample segment. The average speed change rates of all samples are divided into five level intervals according to their numerical values ​​from smallest to largest, labeled as "extremely slow", "slowly changing", "medium changing", "fast changing", and "extremely fast". Then, the proportion of samples in each level interval that are manually labeled or marked as "vehicle stable operation" is counted and recorded as the stability ratio of that level. At the same time, the proportion of that level in all samples is counted and recorded as the base proportion. The stability ratio of each level is divided by its base proportion to obtain the relative stability weight of that level. All weights are then normalized to make their average value 1. Finally, the normalized weight corresponding to each level is used as the value of the speed change adjustment coefficient.

[0037] For the speed mutation adjustment coefficient: at least 5000 vehicle speed time series segments, each with a duration of 20 seconds, are randomly selected from the historical database. The sampling frequency is uniformly 10Hz. For each segment, the difference between the maximum and minimum speed values ​​is first calculated and recorded as the speed extreme value change amplitude. Then, all samples are divided into five level intervals according to the amplitude value from smallest to largest, and labeled as "minimal change", "small change", "medium change", "large change" and "maximum change" respectively. For each level, the proportion of samples marked as "vehicle stable operation" by manual or rule system is counted and recorded as the stability ratio of that level. At the same time, the frequency of occurrence of this level in all samples is counted and recorded as the basic proportion. The stability ratio of each level is divided by its basic proportion to obtain the relative stability weight. Then, the weights of all levels are uniformly normalized so that the average weight of all levels is 1. The normalized weight is the value of the speed mutation adjustment coefficient.

[0038] For the lateral acceleration adjustment coefficient: Extract no less than 5000 segments of vehicle lateral acceleration time-series data, each segment lasting 20 seconds, from the historical database. The sampling frequency is uniformly set to 10Hz. For each data segment, calculate the maximum value of the lateral acceleration of the entire sequence and take its absolute value as the peak value of the lateral acceleration of that segment. Based on the peak value, divide the samples into five levels: "extremely low", "low", "medium", "high" and "extremely high". Calculate the proportion of samples marked as "stable operation" within each level and calculate its proportion relative to the total number of samples in that level (i.e., the stability proportion). Then divide it by the proportion of that level in the overall sample to obtain the relative stability weight. Normalize the relative stability weights of all levels to make their mean value 1, and obtain the value of the lateral acceleration adjustment coefficient.

[0039] For the transverse acceleration frequency adjustment coefficient: Extract no less than 5000 transverse acceleration data sequences, each lasting 20 seconds, from the historical database. The sampling frequency is 10Hz. For each sequence, set 0.3g as the mutation threshold and count the number of times the absolute value of acceleration continuously crosses this threshold to obtain the transverse acceleration mutation frequency of the sequence (in times / second). Divide all frequency values ​​into five levels according to their numerical value: "extremely low frequency", "low frequency", "medium frequency", "high frequency" and "extremely high frequency". Then, count the proportion of samples marked as "stable state" in each level and calculate the ratio between the proportion of samples in each level and the proportion of samples in that level to obtain the relative stability weight. Normalize the relative stability weights of all levels to make the mean 1. The normalized value is the transverse acceleration frequency adjustment coefficient.

[0040] For the brake pressure adjustment coefficient: at least 5,000 brake pressure time series segments, each 20 seconds long, were extracted from the historical database. The sampling frequency was uniformly set to 10Hz. For each data segment, the maximum brake pressure value was extracted as the representative value of that segment. All samples were sorted according to the pressure value and divided into five levels: "extremely light", "light", "medium", "heavy" and "extremely heavy". For each level, the proportion of samples marked "vehicle stable operation" was calculated and compared with the proportion of that level in the entire sample to obtain the relative stability weight. All weights were normalized and mapped to a value with a mean of 1, which was used as the brake pressure adjustment coefficient corresponding to that level.

[0041] The specific implementation example for calculating the vehicle dynamics timing index is as follows, with the following parameters: The average linear velocity of the vehicle changes at a rate of approximately 0.354.

[0042] The extreme change range of the vehicle's linear velocity is approximately 0.612.

[0043] The maximum lateral acceleration of the vehicle is approximately 0.475.

[0044] The frequency of sudden changes in the vehicle's lateral acceleration is approximately 0.268.

[0045] The maximum braking pressure of the vehicle is approximately 0.587.

[0046] The speed change adjustment coefficient stored in the database is approximately 1.025.

[0047] The rate mutation adjustment coefficient stored in the database is approximately 1.080.

[0048] The lateral acceleration adjustment coefficient stored in the database is approximately 1.102.

[0049] The lateral acceleration frequency adjustment coefficient stored in the database is approximately 1.011.

[0050] The brake pressure adjustment coefficient stored in the database is approximately 1.034.

[0051] Substituting the above data into the specific formula for calculating the vehicle power timing index, we get: Vehicle power timing index = ((0.354)^1.025) + ((0.612)^1.080) + ((0.475)^1.102)) × ((0.268)^1.011) × ((0.587)^1.034) ≈ 0.209.

[0052] In this implementation scheme, a vehicle dynamic temporal feature extraction network is constructed to achieve temporal modeling and deep feature extraction of vehicle linear velocity, lateral acceleration, and braking pressure. All data are acquired in real time through wheel speed sensors, centroid accelerometers, and hydraulic pressure sensors, ensuring the physical accuracy and real-time performance of the collected results. The input encoding layer unifies the three data formats and performs normalization and standardization processing to improve feature consistency; multi-scale convolutional layers introduce convolutional kernels with different receptive fields to identify various dynamic fluctuation patterns, adapting to the changing characteristics of different driving time domains; a bidirectional gated recurrent network models temporal dependencies to capture the impact of asymmetric or sudden driving behaviors on the vehicle's dynamic state; finally, the statistical mapping layer outputs the average speed change rate, speed range, maximum lateral acceleration, mutation frequency, and maximum braking pressure. These temporal features can realistically depict the vehicle's response intensity and mutation patterns under different operating conditions. Based on the corresponding temporal indices derived from feature analysis, the subsequent scheduling system can automatically identify risk trends such as excessive acceleration, frequent lane changes, or abnormal braking, thereby enhancing the intelligent judgment and intervention capabilities for driving behavior in complex traffic situations.

[0053] Specifically, the time-series data of the driver interaction status includes steering wheel grip force value, cabin carbon dioxide concentration value and driver heart rate value at several time points. The time-series characteristics of the driver interaction status include steering wheel grip force fluctuation amplitude value, steering wheel grip force disturbance frequency, cabin carbon dioxide concentration rise rate, cabin carbon dioxide concentration fluctuation average amplitude and driver heart rate average deviation.

[0054] Among them, the steering wheel grip force value refers to the average grip force applied by the driver's hands while holding the steering wheel. It can be measured and obtained by a flexible piezoelectric thin film sensor array embedded in the inner ring of the steering wheel. That is, the grip force value can be obtained by collecting its charge output, amplifying it and averaging it.

[0055] The carbon dioxide concentration value in the vehicle cabin refers to the volume fraction of carbon dioxide in the air inside the driver's cabin. It is used to reflect the driver's respiratory load and fatigue risk. It can be measured by a non-dispersive infrared (NDIR) carbon dioxide sensor installed near the air vent on the dashboard.

[0056] The driver's heart rate value refers to the frequency of the driver's heartbeat per unit time, which can be measured by a photoplethysmography (PPG) sensor embedded in the steering wheel spokes.

[0057] like Figure 3 As shown, the specific steps to obtain the temporal features of driver interaction states are as follows: In the signal alignment embedding layer of the driving interaction feature extraction network, the temporal data of driver interaction states are standardized to obtain a standardized driving interaction input sequence. Specifically, the steering wheel grip force value, cabin carbon dioxide concentration value, and driver heart rate value are read in chronological order, and each parameter is standardized. Specifically, the steering wheel grip force value, cabin carbon dioxide concentration value, and driver heart rate value are normalized according to their maximum and minimum values, respectively, so that their values ​​are mapped to the [0,1] interval. Then, zero-mean unit variance standardization (Z-score standardization) is used to obtain the standardized driving interaction input sequence, ensuring that the data from different channels have the same scale. In the sliding convolutional fluctuation detection layer of the driving interaction feature extraction network, a sliding window one-dimensional convolution is performed on the standardized driving interaction input sequence to obtain a local interaction fluctuation feature sequence. Specifically, the standardized driving interaction input sequence is fed into the convolutional layer, and different sizes of convolutional kernels (such as kernels of sizes 3, 5, and 7) are applied to perform a sliding convolution operation on the sequence of each channel to extract local fluctuation features in the time series. Specifically, the convolutional layer calculates the local patterns (such as short-term fluctuations, sub-short-term fluctuations, and medium-term fluctuations) within each time window and performs a nonlinear transformation using the ReLU activation function. Then, max pooling is used to retain the most significant fluctuation features. Finally, the results of the convolution and pooling processes are concatenated along the channel dimension to obtain the local interaction fluctuation feature sequence. In the Transformer self-attention analysis layer of the driving interaction feature extraction network, multi-head self-attention weight analysis is performed on the local interaction fluctuation feature sequence to obtain the driving interaction temporal embedding feature sequence. Specifically, the local interaction fluctuation feature sequence is input into the Transformer self-attention analysis layer, where a multi-head self-attention mechanism is used to calculate the dependencies between time steps. Specifically, the feature vector of each input time step is weighted and averaged with the features of other time steps to capture the global dependencies between time points in the sequence. By calculating the attention weights for each time step, the network can identify important fluctuations and patterns, integrate this information into the global context, and finally output the driving interaction temporal embedding feature sequence. In the statistical mapping and offset analysis layer of the driving interaction feature extraction network, fluctuation amplitude analysis, disturbance frequency counting, rise rate and mean offset analysis are performed on the driving interaction time-series embedded feature sequence to obtain the time-series features of the driver interaction state. Specifically, the following statistical analysis is performed on the driving interaction time-series embedded feature sequence: The difference between the maximum and minimum values ​​of the steering wheel grip force channel is calculated to obtain the steering wheel grip force fluctuation amplitude; The frequency of steering wheel grip force disturbance is obtained by counting the fluctuation points in the grip force channel that exceed the threshold (e.g., 0.7). The carbon dioxide concentration channel in the cabin was subjected to time difference analysis to obtain the rate of concentration increase. The average amplitude of concentration fluctuations is obtained by performing a sliding calculation on the standard deviation of carbon dioxide concentration. For the driver's heart rate channel, calculate the deviation between the heart rate and the resting heart rate at each time point to obtain the heart rate deviation.

[0058] The pre-training steps for the driving interaction feature extraction network are as follows: First, training samples were collected based on a large-scale labeled dataset. This dataset contains a large number of time-series data samples of driver interaction states. Each sample consists of a time series of steering wheel grip force, cabin carbon dioxide concentration, and driver heart rate values, and is simultaneously labeled with its corresponding driving behavior category or health status label. To improve the model's generalization ability, the training data collection phase covered driver interaction data under different driving styles, speed ranges, driving times, and environmental conditions. All time-series data were sampled synchronously at set time steps, and outliers were removed.

[0059] Next, the standardized training data is input into the driving interaction feature extraction network, where features are extracted sequentially through a signal alignment embedding layer, a sliding convolutional fluctuation detection layer, a Transformer self-attention analysis layer, and a statistical mapping and offset analysis layer. During the training phase, a supervised learning strategy is employed, using a loss function to constrain the error between the extracted temporal features and the true labels or reference feature values. If the labels are categorical information, the cross-entropy loss function is used for training; if the labels are continuous feature values, the mean squared error (MSE) is used as the regression loss function.

[0060] Meanwhile, to prevent overfitting during training, a Dropout layer and L2 regularization mechanism are introduced between the convolutional layers and the self-attention analysis layer. During optimization, the Adam optimizer is used for gradient descent, and the accuracy of feature extraction under different driving conditions is monitored in real time using a validation set to ensure that the network can stably extract driver interaction state features.

[0061] After completing the basic training, a fine-tuning training phase is performed, retraining the model on datasets with different vehicle types, driving styles, and road environments to improve its adaptability and generalization performance. Finally, the trained model parameters are frozen, and the pre-trained model is applied to a real-world system.

[0062] The specific steps to obtain the driver interaction state timing index are as follows: read the timing characteristics of the driver interaction state; perform comprehensive analysis on the timing characteristics of the driver interaction state to obtain the driver interaction state timing index.

[0063] The specific formula for calculating the timing index of driver interaction states is as follows: ;in, , , , , , The parameters, in order, are: driver interaction state timing index, steering wheel grip force fluctuation amplitude, steering wheel grip force disturbance frequency, cabin carbon dioxide concentration rise rate, cabin carbon dioxide concentration fluctuation average amplitude, and driver heart rate average deviation. This is a natural constant, and in this embodiment, it is taken as 2.71. , , The coefficients are, in order, the carbon dioxide elevation adjustment coefficient, the carbon dioxide fluctuation adjustment coefficient, and the heart rate deviation adjustment coefficient stored in the database.

[0064] It should be explained that the specific steps for obtaining the carbon dioxide elevation adjustment coefficient, carbon dioxide fluctuation adjustment coefficient, and heart rate deviation adjustment coefficient stored in the database are as follows: For the carbon dioxide elevation adjustment coefficient: at least 5000 time-series data points of cabin carbon dioxide concentration, each lasting 20 seconds, were randomly selected from the historical database, with a sampling frequency of 1Hz. For each data point, linear least squares was used to perform linear fitting to obtain the slope of the carbon dioxide concentration change in that segment, denoted as the elevation rate (unit: ppm / s). All elevation rates were divided into five level intervals according to their magnitude, named "virtually no increase", "slow increase", "moderate increase", "rapid increase", and "sharp increase". Subsequently, the proportion of samples labeled "stable driver mental state" or "no attention abnormality" in each level was counted as the stability proportion of that level. Then, the frequency of each level in the entire sample was counted, and the ratio of the stability proportion to the frequency of occurrence was calculated to obtain the relative stability weight of that level. Finally, the relative stability weights of all levels were normalized to a mean of 1, yielding the carbon dioxide elevation adjustment coefficient corresponding to each level.

[0065] For the carbon dioxide fluctuation adjustment coefficient: Randomly extract no less than 5000 segments of carbon dioxide concentration time series data (sampling frequency 1Hz) of 20 seconds each from the historical database. Calculate the mean absolute deviation (MAD) or standard deviation of the concentration change for each segment as the "fluctuation amplitude index" for that segment. Sort all fluctuation amplitude values ​​by size and divide them into five levels: "extremely low fluctuation", "low fluctuation", "medium fluctuation", "high fluctuation", and "severe fluctuation". Count the proportion of samples labeled "normal pilot condition" or "stable cockpit environment" in each level and calculate the ratio with the proportion of the corresponding level in the total sample to obtain the relative stability weight. Normalize the relative stability weights of all levels to make the mean 1. The normalized value is the carbon dioxide fluctuation adjustment coefficient corresponding to each level.

[0066] For the heart rate deviation adjustment coefficient: At least 5000 segments of driver heart rate time-series data, each 20 seconds long, were extracted from the historical database at a sampling frequency of 1Hz. For each data segment, the mean heart rate of that segment was first calculated. Then, the individual resting heart rate range was extracted from historical heart rate samples of drivers in resting or non-driving scenarios. The deviation (i.e., the difference between the mean heart rate of that segment and the individual resting heart rate mean, in bpm) was calculated. All deviations were divided into five levels according to their absolute values: "No significant deviation," "Slight deviation," "Moderate deviation," "Significant deviation," and "Severe deviation." The proportion of samples labeled "Concentrated" or "Stable Driving" in each level was counted and divided by the proportion of that level in the entire sample to obtain the relative stability weight. Finally, the weights of all levels were normalized and mapped to a value with a mean of 1, which served as the specific value of the heart rate deviation adjustment coefficient.

[0067] The specific implementation example for calculating the timing index of driver interaction state is as follows, with the following parameters: The fluctuation range of steering wheel grip force is approximately 0.132.

[0068] The frequency of steering wheel grip force disturbance is approximately 0.056.

[0069] The rate of increase in carbon dioxide concentration in the cabin was approximately 0.045.

[0070] The average fluctuation amplitude of carbon dioxide concentration in the cabin is approximately 0.120.

[0071] The average deviation of the driver's heart rate is approximately 0.356.

[0072] The natural constant is 2.71.

[0073] The carbon dioxide elevation adjustment coefficient stored in the database is approximately 0.875.

[0074] The carbon dioxide fluctuation adjustment coefficient stored in the database is approximately 1.054.

[0075] The heart rate deviation adjustment coefficient stored in the database is approximately 0.921.

[0076] Substituting the above data into the specific formula for calculating the driver interaction state timing index, we get: Driver interaction state timing index = ((0.132×0.056) / (1+(2.71^(-(0.875×0.045+1.054×0.120)))))+((0.921×0.356)^2)≈0.112.

[0077] This implementation scheme achieves precise monitoring and quantification of multidimensional physiological and behavioral characteristics of drivers during driving by extracting and analyzing time-series data on driver interaction states. Specifically, in-depth analysis of time-series data on steering wheel grip force, cabin carbon dioxide concentration, and driver heart rate comprehensively reflects the driver's physiological load, emotional fluctuations, and potential driving risks such as fatigue and distraction. Through a multi-layered feature extraction network, including signal alignment, convolutional fluctuation detection, Transformer self-attention analysis, and statistical mapping analysis, subtle changes in driver interaction states are deeply explored. By calculating features such as steering wheel grip force fluctuation amplitude, carbon dioxide concentration rise rate, and heart rate deviation, driver fatigue and distraction states can be identified in a timely manner. Furthermore, by using data adjustment coefficients based on historical databases, this method not only adapts to different drivers and environments but also improves the system's generalization ability and stability. Ultimately, it achieves precise quantification of driver interaction states, greatly improving the intelligence level of vehicle safety management and effectively reducing the risk of traffic accidents.

[0078] Specifically, the steps for implementing intelligent dispatching measures for vehicles based on a set of vehicle driving state characteristic indices are as follows: The set of vehicle driving state characteristic indices is compared with several preset sets of driving dispatch intervals. Each set of driving dispatch intervals includes a vehicle power interval and a driver interaction state interval, and each set of driving dispatch intervals corresponds to an intelligent dispatching strategy. Based on the intelligent dispatching strategy corresponding to when the set of vehicle driving state characteristic indices falls within a preset set of driving dispatch intervals, intelligent dispatching processing is performed on the vehicle, including but not limited to the following examples: Scheduling interval set 1: Vehicle power range: [0.150, 0.300] indicates that the vehicle is in a moderate acceleration or braking state, with relatively smooth power output, suitable for normal driving.

[0079] The driver interaction state range is [0.100, 0.200], which indicates that the driver is in a relatively stable state with small fluctuations in heart rate and grip strength, and the driver's state is relatively concentrated.

[0080] Intelligent Dispatch Strategy 1: The vehicle is in normal driving condition, and the driver's interaction is stable. The standard driving mode is executed, and the system provides smooth driving suggestions based on road conditions, traffic flow, and other factors.

[0081] Intelligent dispatch measure 1: Maintain the current speed. It is recommended that drivers maintain a constant speed for an extended period of time, provided it is safe to do so.

[0082] Scheduling interval set 2: Vehicle power range: [0.300, 0.500] indicates that the vehicle is in the acceleration phase and has strong power output.

[0083] Driver interaction state range: [0.200, 0.400] indicates that the driver's heart rate deviates significantly, possibly indicating a state of high tension, such as during rapid acceleration or driving.

[0084] Intelligent Dispatch Strategy 2: The vehicle's acceleration is quite strong, which may cause some psychological stress for the driver. The system automatically adjusts to a smooth acceleration strategy to avoid excessive acceleration that could lead to driver discomfort.

[0085] Intelligent dispatching measure 2: The system suggests that drivers slow down appropriately and provides real-time reminders to prevent excessive acceleration and avoid excessive psychological pressure on drivers.

[0086] Scheduling interval set 3: Vehicle power range: [0.500, 0.700], indicating high-speed driving state, where the vehicle is at a high speed and accelerating or maintaining a high speed.

[0087] The driver interaction state range is [0.400, 0.600], which indicates that the driver may be in a relatively fatigued or tense state, with a large deviation in heart rate and frequent fluctuations in grip strength.

[0088] Intelligent Dispatch Strategy 3: When a vehicle is traveling at high speed, the driver may be fatigued or stressed, which could affect driving safety. The intelligent dispatch system recommends an automatic deceleration mode and provides safety reminders.

[0089] Intelligent dispatching measure 3: The system actively reduces vehicle speed, controls acceleration, and reminds the driver to rest or adjust their state to prevent fatigue driving.

[0090] Scheduling interval set 4: Vehicle power range: [0.700, 1.000], indicating rapid acceleration or high-intensity driving, when the vehicle enters a state of rapid acceleration or deceleration.

[0091] The driver interaction state range is [0.600, 0.800], which indicates that the driver is in a state of high tension or alertness, possibly accompanied by strong physiological reactions.

[0092] Intelligent Dispatch Strategy 4: The vehicle is undergoing high-intensity acceleration or sudden braking, and the driver is in a state of tension. The system suggests entering emergency mode to reduce driving intensity.

[0093] Intelligent dispatching measure 4: Automatically adjust acceleration and braking force to remind the driver to slow down the driving pace, maintain an appropriate speed, and ensure safe driving.

[0094] Scheduling interval set 5: Vehicle power range: [0.000, 0.150], indicating low-speed driving or stationary state, where the vehicle is stopped or at low speed.

[0095] The driver interaction state range is [0.000, 0.100], which indicates that the driver's heart rate deviates slightly and is in a relatively relaxed and focused state.

[0096] Intelligent dispatch strategy 5: When the vehicle is at low speed or stationary and the driver is relaxed, the system recommends a rest mode to ensure the driver can recover their energy.

[0097] Intelligent dispatch measure 5: The system automatically switches to parking mode and provides rest suggestions to remind drivers to relax appropriately and maintain a good mental state.

[0098] In this implementation plan, by monitoring the vehicle's power performance and the driver's interaction status in real time, the system can dynamically adjust the vehicle's driving strategy. For example, when the driver is under stress or fatigued, the system will automatically slow down and remind the driver to rest, thereby effectively reducing the risk of traffic accidents caused by the driver's physical and mental discomfort. At the same time, the intelligent dispatch system can provide corresponding dispatch strategies according to different driving scenarios (such as acceleration, emergency braking, high-speed driving, etc.) to ensure that the vehicle operates in the most suitable driving state, improving the driver's comfort and sense of security.

[0099] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0100] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent dispatch system for vehicle safety management based on artificial intelligence, characterized in that: include: The time-series data acquisition unit is used to acquire time-series data of vehicle driving status based on a set time period when the vehicle is in motion, including time-series data of vehicle power and driver interaction status. The temporal feature extraction unit is used to extract features from the vehicle driving state temporal data based on the pre-trained temporal analysis network, and obtain the vehicle driving state temporal feature set, including the vehicle power and driver interaction state temporal features. The time series index analysis unit is used to comprehensively analyze the time series feature set of vehicle driving status to obtain the vehicle driving status feature index set, including the time series index of vehicle power and driver interaction status. The intelligent scheduling unit is used to implement intelligent scheduling measures for vehicles based on a set of vehicle driving status characteristic indices.

2. The intelligent dispatch system for vehicle safety management based on artificial intelligence according to claim 1, characterized in that, The temporal analysis network includes a vehicle dynamics temporal feature extraction network and a driving interaction feature extraction network. The vehicle dynamics temporal feature extraction network includes an input encoding layer, a multi-scale convolutional feature extraction layer, a bidirectional gated recurrent unit layer, and a statistical mapping layer. The driving interaction feature extraction network includes a signal alignment embedding layer, a sliding convolutional fluctuation detection layer, a Transformer self-attention analysis layer, and a statistical mapping and offset analysis layer.

3. The intelligent dispatch system for vehicle safety management based on artificial intelligence according to claim 2, characterized in that, Vehicle dynamic time series data includes vehicle linear velocity values, vehicle lateral acceleration values, and vehicle braking pressure values ​​at several time points. Vehicle dynamic time series characteristics include the vehicle average linear velocity change rate, the magnitude of extreme changes in vehicle linear velocity, the vehicle maximum lateral acceleration value, the frequency of sudden changes in vehicle lateral acceleration, and the vehicle maximum braking pressure value.

4. The intelligent dispatch system for vehicle safety management based on artificial intelligence according to claim 3, characterized in that, The specific steps to obtain the vehicle dynamic timing characteristics are as follows: In the input encoding layer of the vehicle power temporal feature extraction network, the vehicle power temporal data is processed to construct a temporal joint vector to obtain a standardized vehicle power input sequence. In the multi-scale convolutional feature extraction layer of the vehicle dynamic temporal feature extraction network, the standardized vehicle dynamic input sequence is subjected to multi-scale convolution processing to obtain the local dynamic structure feature sequence. In the bidirectional gated recurrent unit layer of the vehicle power temporal feature extraction network, the local dynamic structural feature sequence is processed by time-dependent modeling to obtain the vehicle power temporal embedded feature sequence. In the statistical mapping layer of the vehicle power temporal feature extraction network, extreme value analysis, mutation frequency identification and change rate analysis are performed on the vehicle power temporal embedded feature sequence to obtain vehicle power temporal features.

5. The intelligent dispatch system for vehicle safety management based on artificial intelligence according to claim 3, characterized in that, The specific steps to obtain the vehicle power timing index are as follows: Read vehicle power timing characteristics; A comprehensive analysis of the vehicle's power timing characteristics yields the vehicle power timing index.

6. The intelligent dispatch system for vehicle safety management based on artificial intelligence according to claim 2, characterized in that, The time-series data of driver interaction status includes steering wheel grip force value, cabin carbon dioxide concentration value and driver heart rate value at several time points. The time-series characteristics of driver interaction status include steering wheel grip force fluctuation amplitude value, steering wheel grip force disturbance frequency, cabin carbon dioxide concentration rise rate, cabin carbon dioxide concentration fluctuation average amplitude and driver heart rate average deviation.

7. The intelligent dispatch system for vehicle safety management based on artificial intelligence according to claim 6, characterized in that, The specific steps to obtain the temporal features of the driver's interaction state are as follows: In the signal alignment embedding layer of the driving interaction feature extraction network, the temporal data of the driver's interaction state is standardized to obtain a standardized driving interaction input sequence. In the sliding convolutional fluctuation detection layer of the driving interaction feature extraction network, the standardized driving interaction input sequence is processed by sliding window one-dimensional convolution to obtain the local interaction fluctuation feature sequence. In the Transformer self-attention analysis layer of the driving interaction feature extraction network, multi-head self-attention weight analysis is performed on the local interaction fluctuation feature sequence to obtain the driving interaction temporal embedding feature sequence. In the statistical mapping and offset analysis layer of the driving interaction feature extraction network, fluctuation amplitude analysis, disturbance frequency counting, rise rate and mean offset analysis are performed on the driving interaction time-series embedded feature sequence to obtain the time-series features of the driver interaction state.

8. The intelligent dispatch system for vehicle safety management based on artificial intelligence according to claim 6, characterized in that, The specific steps to obtain the timing index of the driver interaction state are as follows: Read the timing characteristics of driver interaction status; A comprehensive analysis of the temporal characteristics of driver interaction states yields the driver interaction state temporal index.

9. The intelligent dispatch system for vehicle safety management based on artificial intelligence according to claim 6, characterized in that, The specific formula for calculating the timing index of driver interaction states is as follows: ； in, , , , , , The parameters, in order, are: driver interaction state timing index, steering wheel grip force fluctuation amplitude, steering wheel grip force disturbance frequency, cabin carbon dioxide concentration rise rate, cabin carbon dioxide concentration fluctuation average amplitude, and driver heart rate average deviation. It is a natural constant. , , The coefficients are, in order, the carbon dioxide elevation adjustment coefficient, the carbon dioxide fluctuation adjustment coefficient, and the heart rate deviation adjustment coefficient stored in the database.

10. The intelligent dispatch system for vehicle safety management based on artificial intelligence according to claim 1, characterized in that, Based on the vehicle driving state characteristic index set, the specific steps for implementing intelligent vehicle scheduling measures are as follows: The vehicle driving state characteristic index set is compared with several preset driving dispatch interval sets for judgment and analysis. Each driving dispatch interval set includes the vehicle power interval and the driver interaction state interval, and each driving dispatch interval set corresponds to an intelligent dispatch strategy. Intelligent scheduling strategies are implemented for vehicles based on the intelligent scheduling strategy corresponding to the vehicle driving state characteristic index set falling within a preset driving scheduling interval set.

Images