Intelligent meeting dispatching method for two-way single lane in flat tunnel based on machine learning

Abstract

The present invention discloses a two-way single-lane intelligent passing scheduling method in a flat tunnel based on machine learning, which relates to the field of mine excavation project management and vehicle scheduling. The method steps are as follows: Step 1: First, a passing situation perception thread collects and preprocesses key information of vehicles in the tunnel, and caches the information; Step 2: The passing warning thread further polls the vehicle cache information in real time; Step 3: Calling a preset comprehensive decision model, and automatically issuing a passing scheduling instruction or prompting manual intervention in the passing scheduling at an appropriate time. The present invention utilizes the fusion of different types of learning models (empirical functions, supervised learning, and unsupervised learning) to complement each other. The empirical function can represent general rules, and is efficient and concise. At the same time, the comprehensive decision model can upgrade its accuracy through repeated training as the actual sample of the project increases. The method can adapt to the complex environment in the tunnel, has great application potential, and has significant effects of high scheduling efficiency and high scheduling accuracy.

Description

Technical Field

[0001] The present invention relates to the fields of mine excavation engineering management and vehicle dispatching, and in particular to a machine learning-based intelligent vehicle passing dispatching method for a two-way single lane in a flat tunnel. Background Art

[0002] In mining excavation projects, many tunnels have limited vehicle width due to their small cross-sections, often supporting only one-way traffic. However, during underground transportation operations, vehicles often travel in opposite directions. The two vehicles can only coordinate autonomously through empirical agreements (such as rules such as small vehicles giving way to large vehicles) and through signals such as lights and horns. These methods are severely ineffective at tunnel bends, greatly reducing transportation efficiency. Furthermore, the underground environment is poorly lit, the space is cramped, and the driver's vision is poor, making it easy for them to make decisions incorrectly, resulting in vehicle collisions and even casualties, posing a significant safety hazard to on-site vehicle dispatching and management. Currently, to ensure smooth traffic flow on two-way single-lane roads underground, avoidance chambers are often set up at regular intervals on the two-way single-lane roads. At the same time, dispatch room personnel manually determine the vehicle's position based on information such as vehicle location provided by the mine's comprehensive information system. They use signal lights, voice broadcasts, and other means to coordinate the two vehicles and reuse the single lane in a time-sharing manner. However, in on-site production operations, vehicle traffic demands are frequent, and the dispatching pressure on dispatch room personnel is too great, which makes it easy for command errors and confusion to occur, and the real-time and effectiveness of dispatching decisions are difficult to guarantee. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for intelligently dispatching vehicles on two-way single lanes in a flat tunnel based on machine learning, so as to solve the problem raised in the above-mentioned background technology. At present, in order to ensure the smooth passage of vehicles on two-way single lanes in the mine, avoidance chambers are often set at certain intervals on the two-way single lanes. At the same time, the dispatching room personnel make manual judgments based on the information such as the vehicle position provided by the mine comprehensive information system, and use signal light indications, voice broadcasts and other means to uniformly dispatch the two vehicles on the road, and time-share the single lane. However, in on-site production operations, the demand for vehicle passage is frequent, the dispatching room personnel are under too much dispatching pressure, command errors and confusion are prone to occur, and the real-time and effectiveness of dispatching decisions are difficult to guarantee.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A machine learning-based intelligent meeting dispatching method for a two-way single lane in a flat tunnel, the method steps are as follows:

[0006] Step 1: First, the oncoming vehicle situation awareness thread collects and pre-processes key information about vehicles in the tunnel and caches the information;

[0007] Step 2: The oncoming vehicle warning thread further polls the vehicle cache information in real time;

[0008] Step 3: Further call the preset comprehensive decision-making model and automatically issue the meeting scheduling instructions or prompt manual intervention in the meeting scheduling at the appropriate time.

[0009] As a further solution of the present invention: in the step 1, the meeting situation perception thread further processes and identifies the vehicle information before the meeting based on the vehicle information in the lane information collection area.

[0010] As a further solution of the present invention: the vehicle information before the meeting includes: ① vehicle condition, which is divided into two categories, vehicle type and whether it is empty. The vehicle type is determined by the registration information entered in the vehicle's own positioning tag, and whether it is empty is set by the driver in real time according to the situation; ② the number of vehicles in the team, first based on the relative distance of adjacent vehicles with the same movement trend, adjacent vehicles below a certain relative threshold within a specified time are classified as a team, and a single vehicle is regarded as a team with a number of 1; ③ speed, obtained by taking the position difference value, determines whether the avoidance decision is issued, and stationary vehicles do not receive avoidance information by default; ④ direction, obtained by jointly judging the position and speed trends, taking the direction of entering or exiting the tunnel, or uphill or downhill; ⑤ the relative distance between the vehicle and the avoidance chamber, which may potentially affect the timing of issuing the avoidance instruction; ⑥ the relative distance between the vehicle and the tunnel entrance, generally relative to the tunnel entrance, providing a benchmark for vehicle sorting; ⑦ the number of avoidance chambers, bound to the tunnel number.

[0011] As a further solution of the present invention: the oncoming vehicle warning thread in step 2 needs to perform traversal pre-processing on the lane vehicle information updated in the cache.

[0012] As a further solution of the present invention: in step three, the meeting samples that require meeting scheduling are input into a comprehensive decision-making model. The comprehensive decision-making model includes three meeting scheduling models: a pre-defined rule function, unsupervised machine learning, and supervised machine learning. The three trained models will eventually make decisions on the meeting scenario simultaneously, and their respective decisions are integrated according to a voting mechanism to obtain the final result.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] 1. The present invention can automatically make decisions on the dispatch of vehicles passing each other in a two-way single lane in a tunnel, which can greatly reduce the workload of dispatchers, improve the efficiency of traffic in the tunnel, and reduce the risk of passing vehicles in the tunnel;

[0015] 2. The present invention utilizes the fusion of different types of learning models (empirical functions, supervised learning, and unsupervised learning) to complement each other. The empirical function can express general rules, which is efficient and concise, but it cannot express the potential information of complex scenes. Machine learning can tap into the considerations of drivers' decisions in complex scenes (such as lighting, under-dug branches as temporary shelters, etc.), while reducing the negative impact of system delays (the problem of inconsistency between the displayed position and the actual position of the vehicle. Since the samples used by machine learning come from the same time, the dispatcher is not in the feedback loop, so this problem will be improved). The selected algorithm takes into account real-time performance. As the actual samples of the project increase, the accuracy of this comprehensive decision-making model can be upgraded through repeated training. Overall, this method can adapt to the complex environment inside the tunnel, has great application potential, and has significant effects of high scheduling efficiency and high scheduling accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of the method for intelligent meeting scheduling of two-way single-lane vehicles in flat tunnels based on machine learning.

[0017] Figure 2 This is a schematic diagram of the on-site machine learning-based intelligent meeting scheduling method for two-way single-lane vehicles in a flat tunnel.

[0018] Figure 3 This is a diagram of the comprehensive decision-making model of the two-way single-lane intelligent meeting scheduling method in a flat tunnel based on machine learning. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] See also Figures 1 to 3 In an embodiment of the present invention, a machine learning-based intelligent meeting scheduling method for a two-way single lane in a flat tunnel is provided, and the method steps are as follows:

[0021] Step 1: First, the oncoming vehicle situation awareness thread collects and pre-processes key information about vehicles in the tunnel and caches the information;

[0022] Step 2: The oncoming vehicle warning thread further polls the vehicle cache information in real time;

[0023] Step 3: Further call the preset comprehensive decision-making model and automatically issue the meeting scheduling instruction or prompt manual intervention in the meeting scheduling;

[0024] In the step 1, the vehicle situation perception thread further processes and identifies the vehicle information before the meeting based on the vehicle information in the lane information collection area (the part of the lane except the meeting area near the avoidance chamber), which is divided into seven types: ① vehicle condition, which is divided into two categories: vehicle type and whether it is empty. The vehicle type is determined by the registration information entered in the vehicle's own positioning tag, and whether it is empty is set by the driver in real time according to the situation (the on-board terminal needs to be equipped with an empty vehicle indication information function); ② the number of vehicles in the team, which is first determined based on the relative distance between adjacent vehicles with the same movement trend and the number of vehicles below a certain limit within a specified time. Adjacent vehicles within a relative threshold are grouped together as a platoon, with a single vehicle considered a platoon of 1. ③ Speed, derived from the position difference, determines whether an avoidance decision is issued. Stationary vehicles do not receive avoidance information by default. ④ Direction, determined by a combination of position and speed trends, is determined as the direction of entering or exiting a lane, or going up or downhill. ⑤ The relative distance between the vehicle and the avoidance chamber may potentially affect the timing of issuing the avoidance command. ⑥ The relative distance between the vehicle and the lane entrance, generally relative to the lane entrance, provides a benchmark for vehicle sorting. ⑦ The number of avoidance chambers is bound to the lane number.

[0025] The oncoming vehicle warning thread needs to traverse and pre-process the cached updated lane vehicle information, such as Figures 1 to 3 As shown in the figure, the samples requiring meeting scheduling are finally input into the comprehensive decision-making model. The comprehensive decision-making model includes three meeting scheduling models: a pre-defined rule function, unsupervised machine learning (fuzzy clustering algorithm), and supervised machine learning (XGBoost algorithm). The three trained models will eventually make decisions on the meeting scenario simultaneously. The respective decisions are integrated according to the voting mechanism to obtain the final result, which greatly improves the accuracy and rationality of meeting avoidance scheduling.

[0026] A passing sample consists of two objects: a convoy entering and exiting a lane. Each object has a feature vector. The dimensions of the vector represent the following features: vehicle type, driving direction, average speed, convoy type, number of vehicles in the convoy, vehicle type in the convoy, distance between the first vehicle and the nearest passing hole, and number of passing holes ahead.

[0027] A sample set of oncoming vehicle scenarios was extracted from a video of an inside cave recorded on the monitoring screen of the mine's comprehensive information system. The sample labels were the actual avoidance results in the video. A total of 1,000 samples were divided into training and test sets in a ratio of 7:3. After data preprocessing, model training and model verification were performed.

[0028] Data preprocessing: Based on the content and time of a video recording, the corresponding vehicle information within 100 meters of the edge of the meeting area is extracted from the comprehensive information management system database. The information is normalized, and the forward filling strategy is used to fill missing values, and outliers (often caused by jumps in vehicle position information) are eliminated. The information is then converted into the corresponding input format of the relevant algorithm package.

[0029] The rule function is shown in the figure. The function result DS takes values ​​of 1, 0, and -1, which correspond to labels 1, 3, and 2 respectively.

[0030] The number of convoys, vehicle condition parameters, driving direction, and the distance of the first vehicle in the convoy to the nearest avoidance chamber are used as rule variables. The goal set by the rule function is consistent with the daily avoidance regulations in the tunnel. It tends to give priority to full vehicles, large vehicles, long convoys, uphill vehicles, and vehicles exiting the tunnel. It also takes into account the distance to the first vehicle in the passing tunnel to improve avoidance efficiency. The parameters of the rule function do not need to be obtained through data training, and the expression weight coefficients are manually specified based on experience.

[0031]

[0032]

[0033]

[0034] Where: N-number of vehicles in the same direction;

[0035] M-Vehicle condition parameters (including vehicle type and whether it is empty)

[0036] D- Driving direction (including entering and exiting lanes and uphill and downhill)

[0037] S- Distance between the first vehicle in the convoy and the nearest avoidance chamber

[0038] a, b, c, d are linear weight functions

[0039] Fuzzy clustering, namely fuzzy k-means (FCM) clustering, uses Euclidean distance to measure similarity and uses the sum of squared errors (SSE) as the objective function to measure clustering quality. Given a dataset D containing n data objects, D = {x1, x2, ..., x n}, n = 1000. Define the category set generated by cluster analysis using the mean algorithm as C = {C1, C2, ..., C K The objective function SSE(C) of the FCM algorithm is as follows:

[0040]

[0041] in:

[0042]

[0043]

[0044] Where w xik Represents object x i The degree of membership to category k, i.e. the possibility. ck It is cluster C K The FCM algorithm updates w by iteratively xik and c k The goal of minimizing SSE is achieved. The specific steps of the FCM algorithm are given below. β is determined empirically and is taken as 2 here. The FCM algorithm is shown in the table of overall execution steps of the fuzzy clustering algorithm.

[0045] Table of overall execution steps of fuzzy clustering algorithm

[0046]

[0047]

[0048] The convergence condition of the FCM algorithm is set to the difference in SSE calculated during two consecutive iterations, as shown in the following formula:

[0049] E(t)=||SSE t -SSE t-1 ||<ε

[0050] Where ε is a pre-set tolerance and t is the number of iterations. When the change in SSE between two iterations is less than this preset value, the algorithm is considered to have converged.

[0051] The FCM model was trained using a sample set of meeting scenarios. The number of clusters, K, was set to 3, resulting in three clusters and cluster centers, corresponding to three decision outcomes (entry convoy avoidance, exit convoy avoidance, and submission for manual intervention, with labels of 1, 2, and 3, respectively). The cluster label was the value that best matched the actual avoidance outcome after classification. Actual samples were fed into the trained model in real time, and their distance to the center point was calculated to generate the results. For each outcome, the object sample was assigned a membership degree in the range [0, 1]. This value indicates the rationality of the object following the outcome. This value was compared with a preset threshold. If it exceeded the threshold, the decision was made based on the cluster label.

[0052] XGBoost classification model training and validation. The XGBoost-based classifier is evaluated through a Bayesian network using 5-fold cross-validation for parameter optimization. The dataset is divided into five equal-sized parts (folds). One fold is discarded, and the model is trained on the other four combined folds. Finally, the model is trained on the discarded fold. This process is repeated five times, with each fold used as a test set. The AUC score is used as the evaluation result of the optimized parameters. The evaluation results of each fold are then combined and averaged to form the basis for hyperparameter adjustment optimization. The parameters with the best overall effect are ultimately selected.

[0053] Bayesian optimization was performed using the Python-based BayesianOptimization package. The optimization function was first defined, and parameters and bounds were passed in. Random exploration was then performed in the initialization phase. Bayesian optimization was then iterated to maximize the return value of the optimization function (i.e., to find the parameter combination with the highest AUC). Once the optimal parameter combination was found, it was used as the XGBoost parameters for retraining the model. The final XgBoost classification algorithm parameters were set as follows: 300 iterations, 3 classes, gbtree classifier type, multi:softmax objective function, learning rate eta = 0.1, and maximum tree depth max_depth = 6. The trained model was used for decision classification in meeting scheduling (labels of 1, 2, and 3 for lane entry avoidance, lane exit avoidance, and manual intervention submission, respectively).

[0054] Model fusion uses a hard voting mechanism. The classification output values ​​of the three models, counting rule function, fuzzy clustering, and XGBoost, are determined by the majority rule. Only when at least two classifiers output the same category can the final meeting decision result output the corresponding category (1, 2, or 3). Otherwise, manual intervention command and dispatch will be prompted through the display terminal.

[0055] The original technical solution of the present invention:

[0056] The present invention provides a two-way single-lane intelligent meeting dispatching method in an underground tunnel based on machine learning. The method abstracts and digitizes the key influencing parameters of the meeting scene, inputs the relevant scene data into a pre-trained model, and outputs the intelligent judgment results in real time, providing auxiliary decision-making for dispatchers, greatly reducing the pressure on personnel dispatching, and improving the real-time and effectiveness of dispatching decisions.

[0057] As shown in the attached figure, the method involves the oncoming vehicle situation awareness thread collecting and preprocessing key information of vehicles in the tunnel, caching the information, and then the oncoming vehicle warning thread polling the cached information of the vehicles in real time, calling the preset comprehensive decision-making model, and automatically issuing oncoming vehicle scheduling instructions or prompting manual intervention in the oncoming vehicle scheduling.

[0058] The on-road situation awareness thread further processes and identifies the vehicle information before the on-road meeting based on the vehicle information in the lane information collection area (the part of the lane except the on-road meeting area near the avoidance chamber), which includes seven types of vehicle information: ① Vehicle condition, which is divided into two categories: vehicle type and whether it is empty. The vehicle type is determined by the registration information entered in the vehicle's own positioning tag, and whether it is empty is set by the driver in real time according to the situation (the on-board terminal needs to be equipped with an empty vehicle indication information function); ② The number of vehicles in the convoy, which is first determined based on the relative distance between adjacent vehicles with the same movement trend and is below a certain relative threshold within a specified time. Adjacent vehicles with the same or similar values ​​are grouped as a team, and a single vehicle is considered a team with a quantity of 1; ③ Speed, derived from the position difference, determines whether an avoidance decision is issued. Stationary vehicles do not receive avoidance information by default; ④ Direction, determined by the combined judgment of position and speed trends, and is taken as the direction of entering or exiting the lane, or uphill or downhill; ⑤ The relative distance between the vehicle and the avoidance chamber may potentially affect the timing of issuing the avoidance command; ⑥ The relative distance between the vehicle and the lane entrance, generally relative to the lane entrance, provides a benchmark for vehicle sorting; ⑦ The number of avoidance chambers is bound to the lane number.

[0059] The oncoming vehicle warning thread needs to traverse and pre-process the cached updated lane vehicle information, as shown in the attached figure. Finally, the oncoming vehicle samples that need to be dispatched are input into the comprehensive decision-making model. The comprehensive decision-making model includes three oncoming vehicle dispatch models: artificial pre-defined rule functions, unsupervised machine learning (fuzzy clustering algorithm) and supervised machine learning (XGBoost algorithm). The three trained models will eventually make decisions on the oncoming vehicle scenario at the same time. The respective decisions are integrated according to the voting mechanism to obtain the final result, which greatly improves the accuracy and rationality of oncoming vehicle avoidance scheduling.

[0060] A passing sample consists of two objects: a convoy entering and exiting a lane. Each object has a feature vector. The dimensions of the vector represent the following features: vehicle type, driving direction, average speed, convoy type, number of vehicles in the convoy, vehicle type in the convoy, distance between the first vehicle and the nearest passing hole, and number of passing holes ahead.

[0061] A sample set of oncoming vehicle scenarios was extracted from a video of an inside cave recorded on the monitoring screen of the mine's comprehensive information system. The sample labels were the actual avoidance results in the video. A total of 1,000 samples were divided into training and test sets in a ratio of 7:3. After data preprocessing, model training and model verification were performed.

[0062] Data preprocessing: Based on the content and time of a video recording, the corresponding vehicle information within 100 meters of the edge of the meeting area is extracted from the comprehensive information management system database. The information is normalized, and the forward filling strategy is used to fill missing values, and outliers (often caused by jumps in vehicle position information) are eliminated. The information is then converted into the corresponding input format of the relevant algorithm package.

[0063] The rule function is shown below. The function result DS takes values ​​of 1, 0, and -1, corresponding to labels 1, 3, and 2, respectively. The number of convoys, vehicle condition parameters, driving direction, and the distance of the first vehicle in the convoy to the nearest avoidance chamber are used as rule variables. The rule function sets objectives consistent with routine avoidance regulations within the tunnel. It prioritizes full vehicles, large vehicles, long convoys, vehicles going uphill, and vehicles exiting the tunnel, while also taking into account the distance to the first vehicle in the tunnel to improve avoidance efficiency. The parameters of the rule function do not require data training; the expression weight coefficients are manually specified based on experience.

[0064] DS=sign(f(1)-f(2))

[0065]

[0066]

[0067] Where: N-number of vehicles in the same direction;

[0068] M-Vehicle condition parameters (including vehicle type and whether it is empty)

[0069] D- Driving direction (including entering and exiting lanes and uphill and downhill)

[0070] S- The distance between the first vehicle in the convoy and the nearest avoidance chamber

[0071] a, b, c, d are linear weight functions

[0072] Fuzzy clustering, namely fuzzy k-means (FCM) clustering, uses Euclidean distance to measure similarity and uses the sum of squared errors (SSE) as the objective function to measure clustering quality. Given a dataset D containing n data objects, D = {x1, x2, ..., x n}, n = 1000. Define the category set generated by cluster analysis using the mean algorithm as C = {C1, C2, ..., C K The objective function SSE(C) of the FCM algorithm is as follows:

[0073]

[0074] in:

[0075]

[0076]

[0077] Where w xik Represents object x i The degree of membership to category k, i.e. the possibility. c k It is cluster C KThe FCM algorithm updates w by iteratively xik and c k The goal of minimizing SSE is achieved. The specific steps of the FCM algorithm are given below. β is determined empirically and is taken as 2 here.

[0078] Fuzzy clustering algorithm

[0079]

[0080]

[0081] The convergence condition of the FCM algorithm is set to the difference in SSE calculated during two consecutive iterations, as shown in the following formula:

[0082] E(t)=||SSE t -SSE t-1 ||<ε

[0083] Where ε is a pre-set tolerance and t is the number of iterations. When the change in SSE between two iterations is less than this preset value, the algorithm is considered to have converged.

[0084] The FCM model was trained using a sample set of meeting scenarios. The number of clusters, K, was set to 3, resulting in three clusters and cluster centers, corresponding to three decision outcomes (entry convoy avoidance, exit convoy avoidance, and submission for manual intervention, with labels of 1, 2, and 3, respectively). The cluster label was the value that best matched the actual avoidance outcome after classification. Actual samples were fed into the trained model in real time, and their distance to the center point was calculated to generate the results. For each outcome, the object sample was assigned a membership degree in the range [0, 1]. This value indicates the rationality of the object following the outcome. This value was compared with a preset threshold. If it exceeded the threshold, the decision was made based on the cluster label.

[0085] XGBoost classification model training and validation. The XGBoost-based classifier is evaluated through a Bayesian network using 5-fold cross-validation for parameter optimization. The dataset is divided into five equal-sized parts (folds). One fold is discarded, and the model is trained on the other four combined folds. Finally, the model is trained on the discarded fold. This process is repeated five times, with each fold used as a test set. The AUC score is used as the evaluation result of the optimized parameters. The evaluation results of each fold are then combined and averaged to form the basis for hyperparameter adjustment optimization. The parameters with the best overall effect are ultimately selected.

[0086] Bayesian optimization was performed using the Python-based BayesianOptimization package. The optimization function was first defined, and parameters and bounds were passed in. Random exploration was then performed in the initialization phase. Bayesian optimization was then iterated to maximize the return value of the optimization function (i.e., to find the parameter combination with the highest AUC). Once the optimal parameter combination was found, it was used as the XGBoost parameters for retraining the model. The final XgBoost classification algorithm parameters were set as follows: 300 iterations, 3 classes, gbtree classifier type, multi:softmax objective function, learning rate eta = 0.1, and maximum tree depth max_depth = 6. The trained model was used for decision classification in meeting scheduling (labels of 1, 2, and 3 for lane entry avoidance, lane exit avoidance, and manual intervention submission, respectively).

[0087] Model fusion uses a hard voting mechanism. The classification output values ​​of the three models, counting rule function, fuzzy clustering, and XGBoost, are determined by the majority rule. Only when at least two classifiers output the same category can the final meeting decision result output the corresponding category (1, 2, or 3). Otherwise, manual intervention command and dispatch will be prompted through the display terminal.

[0088] The present invention provides a machine learning-based intelligent meeting scheduling method for tunnels with two-way single lanes, which has the following beneficial effects: 1) It can automatically make meeting scheduling decisions for tunnels with two-way single lanes, which can greatly reduce the workload of dispatchers, improve tunnel traffic efficiency, and reduce the risk of meeting in tunnels; 2) It utilizes the integration of different types of learning models (empirical functions, supervised learning, and unsupervised learning) to complement each other. Empirical functions can express general rules efficiently and concisely, but cannot express the potential information of complex scenarios. Machine learning can exploit and utilize factors that drivers consider in complex scenarios (such as lighting, undercut branches used as temporary avoidance areas, etc.), while also reducing the negative impact of system latency (the problem of inconsistency between vehicle displayed position and actual position is improved because machine learning uses samples from the same time and dispatchers are not in the feedback loop). The selected algorithm takes into account real-time performance. As the number of actual project samples increases, the accuracy of the comprehensive decision model can be upgraded through repeated training. Overall, this method can adapt to the complex environment of tunnels and has great application potential, high scheduling efficiency, and high scheduling accuracy.

[0089] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A machine learning-based intelligent meeting dispatching method for a two-way single-lane in a flat tunnel, the method steps are as follows: Step 1: First, the oncoming vehicle situation awareness thread collects and pre-processes key information about vehicles in the tunnel and caches the information; Step 2: The oncoming vehicle warning thread further polls the vehicle cache information in real time; Step 3: Further call the preset comprehensive decision-making model and automatically issue the meeting scheduling instruction or prompt manual intervention in the meeting scheduling; In the step 1, the oncoming vehicle situation awareness thread further processes and identifies the vehicle information in the lane information collection area to obtain the vehicle information before the oncoming vehicle; The vehicle information before the meeting includes: ① Vehicle condition, which is divided into two categories: vehicle type and whether it is empty. The vehicle type is determined by the registration information entered in the vehicle's own positioning tag, and whether it is empty is set by the driver in real time according to the situation; ② The number of vehicles in the platoon is determined by the relative distance between adjacent vehicles with the same movement trend. Adjacent vehicles below a certain relative threshold within a specified time are classified as one platoon, and a single vehicle is considered to be a platoon with a number of 1. ③ Speed ​​is derived from the position difference and determines whether an avoidance decision is issued. Stationary vehicles do not receive avoidance information by default. ④ Direction is determined by the combined judgment of position and speed trends, and is determined by the direction of entering or exiting the lane, or going up or downhill. ⑤ The relative distance between the vehicle and the avoidance chamber may potentially affect the timing of the avoidance command. ⑥ The relative distance between the vehicle and the lane entrance is generally relative to the lane entrance and provides a benchmark for vehicle sorting. ⑦ The number of avoidance chambers is bound to the lane number. In the step 2, the oncoming vehicle warning thread needs to perform traversal pre-processing on the cached updated lane vehicle information; In step 3, the meeting samples requiring meeting scheduling are input into a comprehensive decision-making model. The comprehensive decision-making model includes three meeting scheduling models: a pre-defined rule function, unsupervised machine learning, and supervised machine learning. The three trained models will ultimately make decisions for the meeting scenario simultaneously, and their respective decisions are integrated according to a voting mechanism to obtain the final result. This is characterized by: A passing example consists of two objects: a convoy entering and exiting a lane. Each object has a feature vector. The dimensions of the vector represent the following features: vehicle type, driving direction, average speed, convoy type, number of vehicles in the convoy, vehicle type in the convoy, distance between the first vehicle and the nearest passing hole, and number of passing holes ahead. A sample set of oncoming vehicles was extracted from a video of a tunnel recorded on the monitoring screen of the mine's comprehensive information system. The sample labels were the actual avoidance results in the video. A total of 1,000 samples were divided into training and test sets in a ratio of 7:

3. After data preprocessing, model training and model verification were performed. Data preprocessing: Based on the content and time of a video recording, the corresponding vehicle information within 100 meters of the edge of the meeting area is extracted from the integrated information management system database. The information is normalized, missing values ​​are filled using a forward filling strategy, outliers are removed, and the information is converted into the corresponding input format of the relevant algorithm package; The rule function is shown in the figure. The function result DS takes values ​​of 1, 0, and -1, which correspond to labels 1, 3, and 2 respectively. The number of convoys, vehicle condition parameters, driving direction, and the distance of the first vehicle in the convoy to the nearest avoidance chamber are used as rule variables. The rule function sets the goal consistent with the daily avoidance regulations in the tunnel, tending to give priority to full vehicles, large vehicles, long convoys, vehicles going uphill, and vehicles exiting the tunnel, while also taking into account the distance to the first vehicle in the passing tunnel. The parameters of the rule function do not need to be obtained through data training, and the expression weight coefficients are manually specified based on experience. Where: N-number of vehicles in the same direction, M-vehicle condition parameters; D-travel direction; S-the distance between the first vehicle in the convoy and the nearest avoidance chamber; a, b, c, and d are linear weight functions; The Euclidean distance is used to measure similarity, and the sum of squared errors / SSE is used as the objective function to measure clustering quality. Given a dataset D containing n data objects, D = {x1, x2, L, x n }, n = 1000; define the category set generated by cluster analysis using the mean algorithm as C = {C1, C2, L, C K The objective function SSE(C) of the FCM algorithm is as follows: in: Where w xik Represents object x i The degree of membership to category k, i.e. the possibility; c k It is cluster C K The FCM algorithm updates w by iteratively xik and c k Achieve the goal of minimizing SSE; the specific steps of the FCM algorithm are given below; β is specified by experience, here it is 2, and the FCM algorithm is shown in the overall execution steps of the fuzzy clustering algorithm; Input: all data points A, number of clusters k; Output: k cluster center points; 1. Initialize the membership matrix W = [w xik ]; 2.Repeat; 3. Calculate the cluster center point c k ; 4. Calculate w xik Update the membership matrix; 5. until convergence conditions are met; The convergence condition of the FCM algorithm is set to the difference in SSE calculated during two iterations, as shown in the following formula: E(t)=||SSE t -SSE t-1 ||<e Where ε is a preset tolerance error and t is the number of iterations. When the change in SSE between two iterations is less than this preset value, the algorithm is considered to have converged. The FCM model was trained using a sample set of meeting scenarios. The number of clusters, K, was set to 3, resulting in three clusters and cluster centers, corresponding to three decision outcomes: lane entry convoy avoidance, lane exit convoy avoidance, and manual intervention submission. The cluster labels were assigned the values ​​1, 2, and 3, respectively. The cluster label was the value that matched the actual avoidance outcome most closely after classification. The actual collected samples were fed into the trained model in real time, and their distances to the center point were calculated to produce the results. For each outcome, the object sample was assigned a membership degree in the range [0, 1]. This value indicated the rationality of the object executing the outcome. This value was compared with a preset threshold. If it exceeded the threshold, the decision was made according to the cluster label. XGBoost classification model, training and validation; XGBoost-based classifiers are evaluated through Bayesian networks and parameter optimization is performed using a 5-fold cross-validation method. The dataset is divided into 5 equal-sized parts / folds, one fold is discarded, and the model is trained on the other 4 combined folds. Finally, the model is trained on the discarded fold. This process is repeated 5 times, and each fold is used as a test set. The AUC score is used as the evaluation result of the optimized parameters. The evaluation results of each fold are then combined and averaged as the basis for hyperparameter adjustment optimization; finally, the parameters with the best overall effect are selected; Bayesian optimization was performed using the Python-based BayesianOptimization package. The optimization function was first defined, and parameters and bounds were passed in. Random exploration was then performed in the initialization phase. Bayesian optimization was then iterated to maximize the return value of the optimization function, attempting to find the parameter combination with the highest AUC. Once the optimal parameter combination was found, it was used as the XGBoost parameter for retraining the model. The final XgBoost classification algorithm parameters were set as follows: 300 iterations, 3 categories, gbtree classifier type, multi:softmax objective function, learning rate eta = 0.1, and maximum tree depth max_depth = 6. The trained model was used for decision classification in meeting scheduling: entering lane avoidance, exiting lane avoidance, and submitting for manual intervention, with labels taking values ​​of 1, 2, and 3, respectively. Model fusion uses a hard voting mechanism. The classification output values ​​of the three models, counting rule function, fuzzy clustering, and XGBoost, are determined by the majority rule. Only when at least two classifiers output the same category can the final meeting decision result output the corresponding category 1, 2, or 3. Otherwise, manual intervention command and dispatch will be prompted through the display terminal.

Images