Ski field transport vehicle emergency processing method based on deep learning and vehicle-mounted system

Abstract

The invention relates to the field of data processing, and particularly provides a deep learning-based ski field transport vehicle emergency processing method and a vehicle-mounted system, and the method comprises the steps: obtaining a real-time monitoring data sequence of a ski field transport vehicle, and carrying out the real-time monitoring of the real-time monitoring data sequence based on a mode matching neural network according to a state representation vector of the real-time monitoring data sequence; determining a target vehicle operation mode corresponding to the real-time monitoring data sequence in each first vehicle operation mode prepared in advance; determining the first vehicle mode matching data belonging to the target vehicle operation mode according to the target vehicle operation mode and the affiliation condition between the first vehicle operation mode prepared in advance and the first vehicle mode matching data; wherein the vehicle state characteristics of the first vehicle mode matching data belonging to the same first vehicle operation mode are similar; and determining target vehicle mode matching data corresponding to the real-time monitoring data sequence based on the first vehicle mode matching data belonging to the target vehicle operation mode.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an emergency handling method and vehicle-mounted system for a ski resort transport vehicle based on deep learning. Background Art

[0002] In the operation and management of ski resorts, the safe and stable operation of ski resort transport vehicles is crucial to ensuring the experience and safety of tourists. However, due to the special geographical environment and complex operating conditions of ski resorts, ski resort transport vehicles may face various emergencies during operation, which require timely and accurate emergency response.

[0003] At present, in the field of vehicle emergency response, although there are some monitoring and processing methods, there are still many shortcomings. Traditional monitoring methods can often only obtain limited vehicle operation data, such as focusing only on individual conventional indicators such as vehicle speed and temperature, which is difficult to fully reflect the actual operation status of the vehicle in a complex environment. This makes it impossible to identify and judge some potential safety hazards or emergencies in a timely and accurate manner, resulting in lag and inaccuracy in emergency response. In terms of data processing and analysis, most existing technologies use feasible threshold judgment or rule matching methods. For example, when a certain operating parameter exceeds a preset threshold, the corresponding alarm or processing mechanism is triggered. However, this method is too simple and crude, and does not fully consider the complexity and diversity of vehicle operation status. Under different operating scenarios and working conditions, the same parameter value may have different meanings, and relying solely on threshold judgment is prone to misjudgment and missed judgment. Summary of the invention

[0004] In view of this, the present application provides a ski resort transport vehicle emergency handling method and vehicle-mounted system based on deep learning. The technical solution of the present application is implemented as follows:

[0005] On the one hand, the present application provides a deep learning-based emergency handling method for ski resort transport vehicles, including: obtaining a real-time monitoring data sequence of a ski resort transport vehicle, and based on a pattern matching neural network, determining a target vehicle operation mode corresponding to the real-time monitoring data sequence in each first vehicle operation mode prepared in advance according to a state characterization vector of the real-time monitoring data sequence; determining the first vehicle pattern matching data belonging to the target vehicle operation mode according to the attribution between the target vehicle operation mode and the first vehicle operation mode prepared in advance and the first vehicle pattern matching data; wherein the vehicle state characteristics of the first vehicle pattern matching data belonging to the same first vehicle operation mode are similar; based on the first vehicle pattern matching data belonging to the target vehicle operation mode, determining the target vehicle pattern matching data corresponding to the real-time monitoring data sequence; based on a pre-set emergency management strategy mapping relationship, using the emergency management strategy corresponding to the target vehicle pattern matching data as the real-time emergency management strategy.

[0006] On the other hand, the present application provides a vehicle-mounted system, including a memory and a processor, wherein the memory stores a computer program that can be executed on the processor, and when the processor executes the program, the steps in the above-described method are implemented.

[0007] In an embodiment of the present application, a real-time monitoring data sequence of a ski resort transport vehicle is obtained, and based on a pattern matching neural network and a state characterization vector of the real-time monitoring data sequence, a target vehicle operating mode corresponding to the real-time monitoring data sequence is determined in each first vehicle operating mode prepared in advance, and then based on the attribution between the target vehicle operating mode and the first vehicle operating mode prepared in advance and the first vehicle pattern matching data, first vehicle pattern matching data belonging to the target vehicle operating mode is determined, wherein the vehicle state characteristics of the first vehicle pattern matching data belonging to the same first vehicle operating mode are similar, and based on the first vehicle pattern matching data belonging to the target vehicle operating mode, the target vehicle pattern matching data corresponding to the real-time monitoring data sequence is determined. Because the target vehicle operation mode corresponding to the real-time monitoring data sequence is determined based on the state characterization vector of the real-time monitoring data sequence based on the pattern matching neural network, and the target vehicle pattern matching data corresponding to the real-time monitoring data sequence is determined based on the first vehicle pattern matching data belonging to the target vehicle operation mode, the target vehicle pattern matching data can be made similar to the vehicle state characteristics of the real-time monitoring data sequence. In addition, because the first vehicle pattern matching data with similar vehicle state characteristics are pre-attributed to the same first vehicle operation mode, the vehicle state characteristics of the first vehicle pattern matching data belonging to the target vehicle operation mode are also similar, and the vehicle state characteristics of each target vehicle pattern matching data are also similar. Then, based on the present application, target vehicle pattern matching data with similar vehicle state characteristics of the real-time monitoring data sequence of the ski resort transport vehicle can be provided, and the vehicle state characteristics of each target vehicle pattern matching data are also similar to each other, which improves the accuracy of the target vehicle pattern matching data and provides a reliable reference for emergency handling of the ski resort transport vehicle.

[0008] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.

[0010] Figure 1 A schematic diagram of the implementation flow of a deep learning-based emergency handling method for a ski resort transport vehicle provided in an embodiment of the present application.

[0011] Figure 2 A schematic diagram of a hardware entity of a vehicle-mounted system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0013] The present application embodiment provides a method for emergency handling of ski resort transport vehicles based on deep learning, which can be executed by a processor of the vehicle system. Figure 1 As shown, the method comprises the following steps:

[0014] Step S10: Acquire the real-time monitoring data sequence of the ski resort transport vehicle, and determine the target vehicle operation mode corresponding to the real-time monitoring data sequence in each first vehicle operation mode prepared in advance based on the pattern matching neural network and the state characterization vector of the real-time monitoring data sequence.

[0015] The on-board system collects real-time monitoring data from various sensors installed on the transport vehicles on the ski resort. These sensors may include, but are not limited to, speed sensors, acceleration sensors, gyroscopes, temperature sensors, pressure sensors, etc. For example, the speed sensor can measure the vehicle's speed in real time, the acceleration sensor can detect the acceleration and deceleration of the vehicle, the gyroscope can be used to monitor the vehicle's posture changes, the temperature sensor can obtain the temperature information of the vehicle's key components, and the pressure sensor can measure the tire pressure, etc. These sensors collect data at a certain time interval (such as every 0.1 seconds) and send the data to the on-board system, which arranges these continuously collected data in chronological order to form a real-time monitoring data sequence. Assume that at a certain moment, the speed sensor collects a vehicle speed of 30 kilometers per hour, the acceleration sensor detects an acceleration of 0.5 meters per second squared, the gyroscope shows that the vehicle posture is normal, the temperature sensor measures the engine temperature of 80 degrees Celsius, and the pressure sensor shows that the tire pressure is 2.5 MPa. These data constitute a data point in the real-time monitoring data sequence at that moment.

[0016] The pattern matching neural network is an artificial intelligence model specially designed for processing and analyzing data patterns. It can recognize and classify new data by learning a large number of known data patterns. In this step, the neural network is used to analyze the real-time monitoring data sequence of the ski resort transport vehicle, find the patterns contained therein, and match them with the first vehicle operation mode prepared in advance. The pattern matching neural network is usually composed of a hierarchical structure composed of multiple neurons, including an input layer, a hidden layer, and an output layer. The input layer receives the real-time monitoring data sequence, the hidden layer extracts and transforms the data features, and the output layer gives the probability that the data belongs to each first vehicle operation mode. For example, a feasible pattern matching neural network may have 10 input neurons, corresponding to the data collected by different sensors; the hidden layer contains 20 neurons for extracting and transforming data features; and the output layer has 5 neurons, representing 5 different first vehicle operation modes.

[0017] The state representation vector is a mathematical representation of the characteristics of the real-time monitoring data sequence. The vehicle-mounted system extracts and transforms the real-time monitoring data sequence into a vector form. This vector contains key information related to the vehicle's operating status in the real-time monitoring data sequence. For example, for a real-time monitoring data sequence containing information such as speed, acceleration, posture and temperature, the vehicle-mounted system can extract the features that best represent the vehicle's operating status through certain algorithms (such as principal component analysis, feature selection algorithm, etc.), and combine these features into a vector, which is the state representation vector. Assume that after feature extraction, the state representation vector obtained is [30, 0.5, 0.8, 80], where 30 represents speed, 0.5 represents acceleration, 0.8 represents the normality of posture (value range 0-1), and 80 represents temperature.

[0018] The first vehicle operation mode is a mode predefined according to the operation status of the vehicle under different circumstances. These modes do not refer to the normal modes set by the vehicle (such as climbing mode, downhill mode, etc.), but are the operation status modes of the vehicle when it encounters emergency handling based on the vehicle's monitoring data analysis. For example, in the bumpy mode, when the vehicle is driving on an uneven snowy road, the real-time monitoring data sequence may show characteristics such as frequent changes in the vertical acceleration of the vehicle and unstable body posture; in the skidding mode, when the vehicle is driving on the snow, due to the reduction of ground friction, abnormal wheel speed and increased lateral acceleration of the vehicle may occur; in the equipment failure mode, when a key component of the vehicle (such as the engine, brake system, etc.) fails, the data of the relevant sensors will show abnormalities, such as excessive engine temperature and insufficient brake pressure.

[0019] After acquiring the state representation vector of the real-time monitoring data sequence, the vehicle-mounted system inputs it into the pattern matching neural network. The neural network analyzes and calculates the state representation vector to evaluate the degree of matching between the vector and each first vehicle operation mode. Specifically, the neural network can calculate the similarity or distance between the state representation vector and the feature vector of each first vehicle operation mode. Commonly used similarity calculation methods include cosine similarity, Euclidean distance, etc.

[0020] Step S20: Determine the first vehicle mode matching data belonging to the target vehicle operation mode according to the attribution between the target vehicle operation mode and the first vehicle mode matching data prepared in advance; wherein the vehicle state characteristics of the first vehicle mode matching data belonging to the same first vehicle operation mode are similar.

[0021] In step S10, the vehicle system has determined the target vehicle operation mode, which is a classification of the current operation state of the ski resort transport vehicle. For example, if the target vehicle operation mode is determined to be "slip mode", it means that the vehicle is currently in a specific operation state such as reduced friction, abnormal wheel speed, increased lateral acceleration, etc. when driving on snow. The target vehicle operation mode is an important basis for determining the relevant pattern matching data later.

[0022] There is a specific attribution relationship between the first vehicle operation mode and the first vehicle mode matching data, that is, which first vehicle mode matching data belong to which first vehicle operation modes. This attribution relationship is determined based on the similarity of vehicle state characteristics. For example, for the first vehicle operation mode of "slip mode", the first vehicle mode matching data belonging to it may include data such as a specific wheel speed range, a lateral acceleration value range, and a vehicle body posture angle change range. These data are closely related to and similar to the state characteristics of the vehicle in "slip mode". For another example, the first vehicle mode matching data corresponding to the "bump mode" may contain relevant data that can reflect the bump state, such as the fluctuation range of vertical acceleration and the up and down vibration frequency of the vehicle body.

[0023] When the vehicle-mounted system determines the first vehicle mode matching data belonging to the target vehicle operation mode, it needs to be based on the previously prepared attribution information. Assume that the vehicle-mounted system has established an attribution relationship database, which stores the association information of various first vehicle operation modes and their corresponding first vehicle mode matching data. When the target vehicle operation mode is determined, the vehicle-mounted system will query and match in this database. For example, when the target vehicle operation mode is "equipment failure mode", and specifically the engine failure, the vehicle-mounted system will search for the first vehicle mode matching data associated with the "engine failure" operation mode in the attribution relationship database. These data may include the specific numerical range of abnormal increase in engine temperature (such as more than 90 degrees Celsius), the interval of abnormal decrease in engine speed (such as less than 1000 rpm), etc. In this way, the vehicle-mounted system can accurately find the first vehicle mode matching data belonging to the target vehicle operation mode.

[0024] The vehicle state characteristics of the first vehicle mode matching data belonging to the same first vehicle operation mode are similar, and this similarity ensures that the determined first vehicle mode matching data can accurately reflect the vehicle state under the target vehicle operation mode. For example, in the "slip mode", the first vehicle mode matching data corresponding to different slip conditions (such as light slip and heavy slip) are similar in vehicle state characteristics, and both reflect the characteristics of reduced friction between the wheels and the ground and reduced vehicle stability. This similarity enables the vehicle-mounted system to analyze and judge the vehicle's operating state more accurately, providing reliable data support for subsequent emergency processing.

[0025] The vehicle-mounted system can establish a special database to store the attribution information between the first vehicle operation mode and the first vehicle pattern matching data. This database needs to have efficient query and retrieval functions so that the vehicle-mounted system can quickly find the corresponding first vehicle pattern matching data according to the target vehicle operation mode. For example, a relational database (such as MySQL) or a non-relational database (such as MongoDB) can be used to implement data storage and management. A pattern matching algorithm can be designed to search for the first vehicle pattern matching data that matches the target vehicle operation mode in the attribution relationship database. The algorithm can be implemented based on similarity calculation, rule matching and other methods. For example, by calculating the similarity between the feature vector of the target vehicle operation mode and the feature vectors of each first vehicle operation mode in the database, the matching result with the highest similarity is selected.

[0026] Since the vehicle's operating status and related data may change with time and environmental changes, it is necessary to establish a data update and maintenance mechanism. Regularly collect new vehicle monitoring data, update and optimize the attribution between the first vehicle operating mode and the first vehicle mode matching data, so as to ensure the accuracy and timeliness of the attribution relationship information based on the vehicle system. For example, the data in the database can be checked and updated at regular intervals (such as one month), outdated or inaccurate data can be deleted, and new valid data can be added.

[0027] Step S30: Based on the first vehicle mode matching data belonging to the target vehicle operation mode, determine the target vehicle mode matching data corresponding to the real-time monitoring data sequence.

[0028] In step S20, the vehicle-mounted system has determined the first vehicle mode matching data belonging to the target vehicle operation mode. These data are a set of data that are similar to the target vehicle operation mode in terms of vehicle state characteristics. For example, if the target vehicle operation mode is a "bumping mode", then the first vehicle mode matching data belonging to this mode may include a specific fluctuation range of vertical acceleration (such as between 1-3 m / s²), a specific interval of the up and down vibration frequency of the vehicle body (such as 5-10 Hz), etc. These data reflect the typical characteristics of the vehicle in a bumpy state and are an important basis for determining the target vehicle mode matching data.

[0029] The real-time monitoring data sequence is a data sequence that the on-board system continuously collects from various sensors on the ski resort transport vehicle. It contains various status information of the vehicle during operation, such as speed, acceleration, posture, temperature, etc. These data are continuously updated over time and reflect the real-time operating status of the vehicle. For example, within a certain period of time, the real-time monitoring data sequence may show that the vertical acceleration of the vehicle frequently fluctuates between 1.5-2.5 m / s², and the body vibrates up and down at a frequency of about 8 Hz, which has a certain correlation with the first vehicle mode matching data in the "bump mode" in terms of characteristics.

[0030] When determining the target vehicle pattern matching data, the on-board system conducts a comprehensive analysis and comparison of the first vehicle pattern matching data and the real-time monitoring data sequence belonging to the target vehicle operating mode. Specifically, the on-board system calculates the degree of matching between the real-time monitoring data sequence and each first vehicle pattern matching data. For example, for multiple first vehicle pattern matching data under the "bumping mode" (such as vertical acceleration and vibration frequency data under different degrees of bumping), the on-board system will respectively calculate the similarity between the vertical acceleration and vibration frequency in the real-time monitoring data sequence and these first vehicle pattern matching data. The similarity can be calculated using a variety of methods, such as the Euclidean distance formula. Assume that there are two data points P(x1, y1) and , the Euclidean distance d between them is calculated as: . In this scenario, P can represent a data point in the real-time monitoring data sequence (such as the vertical acceleration and vibration frequency at a certain moment), and Q represents a data point in the first vehicle pattern matching data. By calculating the Euclidean distance, the smaller the distance, the higher the similarity between the two. The on-board system will traverse all the first vehicle pattern matching data belonging to the target vehicle operation mode, find the first vehicle pattern matching data with the highest similarity to the real-time monitoring data sequence, and determine it as the target vehicle pattern matching data. For example, after calculation, the Euclidean distance between the real-time monitoring data sequence and a group of first vehicle pattern matching data (vertical acceleration fluctuation range is 1.8-2.2 m / s², vibration frequency is 7-9 Hz) is the smallest, then this group of data is determined as the target vehicle pattern matching data.

[0031] Since the vehicle's operating status and environment may change, the first vehicle mode matching data also needs to be continuously updated and optimized. A data update and feedback mechanism is established to adjust and update the first vehicle mode matching data based on actual monitoring data and emergency handling results to improve the accuracy and reliability of matching. For example, if a set of first vehicle mode matching data is found to deviate from the actual vehicle operating status in actual application, it can be corrected based on new data.

[0032] Step S40: Based on a preset emergency management strategy mapping relationship, the emergency management strategy corresponding to the target vehicle pattern matching data is used as a real-time emergency management strategy.

[0033] The pre-set emergency management strategy mapping relationship establishes a clear correspondence between the target vehicle pattern matching data and the corresponding emergency management strategy. This mapping relationship is summarized through a large amount of historical data, professional experience, and analysis of various possible vehicle operating conditions. For example, for the target vehicle pattern matching data in the "skid mode", based on previous experience and analysis of vehicle skidding, the possible mapped emergency management strategy is "reduce vehicle speed, avoid sudden braking and acceleration, and turn on the anti-skid device"; for the target vehicle pattern matching data related to engine failure in the "equipment failure mode", the corresponding emergency management strategy may be "immediately pull over, contact maintenance personnel, and turn on the hazard warning lights at the same time." These mapping relationships are organized and stored in a specific database or data structure of the vehicle system for quick query and call.

[0034] The target vehicle pattern matching data is determined by the vehicle-mounted system in the previous step based on the real-time monitoring data sequence, and it represents a characteristic description of the current actual operating state of the vehicle. After obtaining the target vehicle pattern matching data, the vehicle-mounted system will search according to the pre-set emergency management strategy mapping relationship. This search process is similar to performing an exact matching operation in a large index table. For example, assuming that the currently determined target vehicle pattern matching data indicates that the vehicle is in a "bumping mode" and the parameter corresponding to the degree of bumping is in a certain range, the vehicle-mounted system will quickly locate the emergency management strategy associated with the "bumping mode" and the corresponding parameter range in the emergency management strategy mapping relationship data structure.

[0035] Once the corresponding emergency management strategy is found, the on-board system will use it as a real-time emergency management strategy. This means that the system will guide subsequent operations based on this strategy to ensure the safe operation of the ski resort transport vehicle. For example, if the real-time emergency management strategy is "reduce vehicle speed, avoid sudden braking and acceleration, and turn on anti-skid devices" for "skid mode", the on-board system will send instructions through the vehicle's electronic control system to automatically reduce the vehicle speed, restrict the driver from performing sudden braking and acceleration operations, and turn on the anti-skid devices equipped on the vehicle (such as the automatic anti-skid chain installation system or the anti-skid control system).

[0036] In step S10, the vehicle-mounted system obtains the real-time monitoring data sequence of the ski resort transport vehicle, and based on the pattern matching neural network, according to the state characterization vector of the real-time monitoring data sequence, determines the target vehicle operation mode corresponding to the real-time monitoring data sequence in each first vehicle operation mode prepared in advance. By extracting the state characterization vector and calculating its similarity with the characteristic vector of each first vehicle operation mode, the current operation state of the vehicle can be accurately identified, for example, accurately judging whether the vehicle is in a bumpy mode, a skidding mode, or an equipment failure mode. Therefore, the target vehicle operation mode can be accurately determined, providing an accurate and reliable basis for subsequent emergency handling, and avoiding inappropriate emergency measures due to misjudgment of the vehicle operation mode.

[0037] In step S20, the vehicle system determines the first vehicle mode matching data belonging to the target vehicle operation mode according to the attribution between the target vehicle operation mode and the first vehicle operation mode prepared in advance and the first vehicle mode matching data, and the vehicle state characteristics of the first vehicle mode matching data belonging to the same first vehicle operation mode are similar. This means that the system can find matching mode matching data with similar state characteristics according to the current operation mode of the vehicle. For example, when it is determined that the vehicle is in a slipping mode, corresponding matching data such as the abnormal range of wheel speed and the range of lateral acceleration changes can be found. Therefore, the vehicle mode matching data can be accurately matched, providing strong support for the subsequent formulation of more targeted emergency management strategies, so that emergency measures are more in line with the actual state of the vehicle.

[0038] In step S10, the vehicle-mounted system determines the target vehicle operation mode corresponding to the real-time monitoring data sequence based on the state characterization vector of the real-time monitoring data sequence based on the pattern matching neural network, so that the target vehicle operation mode can accurately reflect the current real state of the vehicle. In step S20, the first vehicle pattern matching data belonging to the target vehicle operation mode is determined based on the target vehicle operation mode and the attribution situation prepared in advance. Since the first vehicle pattern matching data with similar vehicle state characteristics are attributed to the same first vehicle operation mode in advance, the vehicle state characteristics of the first vehicle pattern matching data belonging to the target vehicle operation mode are similar. Then in step S30, the target vehicle pattern matching data corresponding to the real-time monitoring data sequence is determined based on these first vehicle pattern matching data, so that the target vehicle pattern matching data can be made similar to the vehicle state characteristics of the real-time monitoring data sequence, and the vehicle state characteristics of each target vehicle pattern matching data are also similar to each other. Therefore, the accuracy of the target vehicle pattern matching data is improved. Furthermore, in step S40, based on the pre-set emergency management strategy mapping relationship, the ski resort transport vehicle can be provided with an emergency management strategy corresponding to the target vehicle pattern matching data that is similar to the vehicle status characteristics of the real-time monitoring data sequence, thereby enhancing the adaptability of the emergency strategy to the actual status of the vehicle and making emergency handling more scientific and effective.

[0039] In step S40, the vehicle-mounted system uses the emergency management strategy corresponding to the target vehicle pattern matching data as a real-time emergency management strategy based on the pre-set emergency management strategy mapping relationship. This mapping relationship is established based on a large amount of historical data and professional experience, and can provide scientific and reasonable emergency measures according to the specific status of the vehicle. For example, when it is determined that the vehicle is in equipment failure mode and the engine is faulty, the system will adopt emergency strategies such as "immediately pull over, contact maintenance personnel, and turn on the hazard warning lights" according to the mapping relationship. Therefore, it can quickly generate and execute effective emergency management strategies in emergency situations, maximize the safe operation of ski resort transport vehicles, and reduce accident risks and losses.

[0040] Furthermore, in the embodiment of the present application, the vehicle state characteristics of multiple first vehicle pattern matching data under a target vehicle operation mode are similar. At the same time, because each target vehicle operation mode is determined based on the state characterization vector of the real-time monitoring data sequence, the multiple first vehicle pattern matching data under each target vehicle operation mode are similar to the vehicle state characteristics of the real-time monitoring data sequence, and the vehicle state characteristics of multiple first vehicle pattern matching data under each target vehicle operation mode are also similar. Based on this, the embodiment of the present application can determine multiple first vehicle pattern matching data similar to the vehicle state characteristics of the real-time monitoring data sequence for the ski resort transport vehicle based on the vehicle pattern matching data of the ski resort transport vehicle, and the vehicle state characteristics between the multiple first vehicle pattern matching data determined at the same time are also similar, completing the determination of the same state data and improving the accuracy of the vehicle pattern matching data. Because when constructing the attribution between the first vehicle operation mode and the first vehicle pattern matching data, the attribution is obtained by clustering the first vehicle pattern matching data based on the state characterization vector of the second representative data with a larger data capacity of the first vehicle pattern matching data, and when determining the target vehicle pattern matching data, the corresponding target vehicle operation mode is determined based on the pattern matching neural network based on the real-time monitoring data sequence input by the ski resort transport vehicle. The performance of mapping the smaller-scale data (real-time monitoring data sequence) input by the ski resort transport vehicle to the larger-scale data (vehicle pattern matching data) is completed based on the pattern matching neural network, so that the real-time monitoring data sequence input by the ski resort transport vehicle can accurately determine the vehicle pattern matching data with the same pattern, thereby improving the determination performance of the vehicle pattern matching data. Eliminate the problem of poor accuracy when matching large-scale data, as well as the precision error in the matching process of long features and short features.

[0041] As an implementation mode, step S10, based on a pattern matching neural network and according to a state characterization vector of the real-time monitoring data sequence, determines a target vehicle operation mode corresponding to the real-time monitoring data sequence from each first vehicle operation mode prepared in advance, including:

[0042] Step S11: extracting a state representation vector of a real-time monitoring data sequence based on a first state representation component in a pattern matching neural network;

[0043] Step S12: determining the first mode classification probability of the real-time monitoring data sequence being classified into each first vehicle operation mode based on the first probability prediction component in the pattern matching neural network and the state characterization vector of the real-time monitoring data sequence;

[0044] Step S13: determining the target vehicle operating mode corresponding to the real-time monitoring data sequence in each first vehicle operating mode according to the first mode classification probability.

[0045] In step S11, the real-time monitoring data sequence is processed by relying on the first state representation component in the pattern matching neural network. The first state representation component is a part of the neural network specifically used for feature extraction and conversion. It converts the original real-time monitoring data sequence into a state representation vector that can represent its key features through a series of algorithms and calculations. This state representation vector is a multidimensional vector, each dimension of which corresponds to a specific feature in the real-time monitoring data sequence.

[0046] For example, suppose the real-time monitoring data sequence contains information such as the speed, acceleration, body posture angle, and engine temperature of a ski resort transport vehicle. The first state representation component may use convolutional neural network (CNN) technology to process this data. For speed and acceleration data, it may use a series of convolutional layers and pooling layers to extract the changing trends and fluctuation characteristics in the data; for body posture angle data, a specific algorithm may be used to analyze the stability and changes of the posture; for engine temperature data, the rate of change of temperature and outliers may be focused on. After these processes, all these features are combined together to form a state representation vector of the real-time monitoring data sequence.

[0047] Assume that the real-time monitoring data sequence at a certain moment is: speed 30 km / h, acceleration 0.5 m / s², body posture angle normal (represented by the value 1), engine temperature 80 degrees Celsius. After being processed by the first state representation component, the extracted state representation vector may be [0.8, 0.6, 0.9, 0.7], where 0.8 represents the extracted value of speed-related features, 0.6 represents the extracted value of acceleration-related features, 0.9 represents the extracted value of body posture angle-related features, and 0.7 represents the extracted value of engine temperature-related features.

[0048] In step S12, after obtaining the state characterization vector of the real-time monitoring data sequence, the vehicle system will input it into the first probability prediction component in the pattern matching neural network. The main task of the first probability prediction component is to calculate the probability that the real-time monitoring data sequence belongs to each pre-set first vehicle operation mode, that is, the first mode classification probability, based on the state characterization vector.

[0049] The first probability prediction component usually includes a plurality of neurons, which process the input state representation vector through complex calculations and activation functions and output corresponding probability values. The process of calculating the probability involves evaluating the degree of match between the state representation vector and the characteristics of each first vehicle operation mode.

[0050] For example, suppose there are three first vehicle operation modes: bump mode, slip mode, and equipment failure mode. For the input state representation vector [0.8, 0.6, 0.9, 0.7], the first probability prediction component will calculate the degree of match between the vector and the characteristics of the bump mode, slip mode, and equipment failure mode, and convert them into probability values. Assume that after calculation, the first mode classification probability of the real-time monitoring data sequence being classified into the bump mode is 0.3, the first mode classification probability of being classified into the slip mode is 0.2, and the first mode classification probability of being classified into the equipment failure mode is 0.1.

[0051] In step S13, after obtaining the first mode classification probability of the real-time monitoring data sequence being classified into each first vehicle operation mode, the vehicle-mounted system determines the final target vehicle operation mode based on these probability values. Typically, the first vehicle operation mode with the highest first mode classification probability is selected as the target vehicle operation mode.

[0052] For example, in the previous example, the probability of the real-time monitoring data sequence being classified into the first mode of the bumping mode is 0.3, the probability of the first mode of the real-time monitoring data sequence being classified into the first mode of the slipping mode is 0.2, and the probability of the first mode of the real-time monitoring data sequence being classified into the first mode of the equipment failure mode is 0.1. Since 0.3 is the largest of the three probability values, the vehicle-mounted system will determine that the target vehicle operation mode corresponding to the real-time monitoring data sequence is the bumping mode.

[0053] In order to determine the target vehicle operation mode, the vehicle-mounted system can set a probability threshold. When the probability of the first mode classification of a certain first vehicle operation mode exceeds the threshold, it is used as the target vehicle operation mode. This can avoid misjudging the vehicle operation mode when the probability value is low. For example, the probability threshold is set to 0.25. Only when the probability of a certain mode is greater than 0.25, the vehicle is considered to be in this mode. Secondly, when there are multiple first mode classification probabilities of the first vehicle operation mode that are similar and high, further judgment can be made in combination with other information, such as certain key features or historical data in the real-time monitoring data sequence. For example, if the probabilities of the bump mode and the slip mode are close and high, you can check the data such as acceleration and body posture angle. If the acceleration fluctuates greatly and the body posture is unstable, it is more likely to be judged as a bump mode.

[0054] Steps S11 to S13 perform feature extraction and probability calculation on the real-time monitoring data sequence through the first state characterization component and the first probability prediction component in the pattern matching neural network, and finally determine the target vehicle operation mode corresponding to the real-time monitoring data sequence, providing an accurate basis for subsequent emergency processing.

[0055] As an implementation manner, in step S12, determining the first mode classification probability of the real-time monitoring data sequence being classified into each first vehicle operation mode according to the state characterization vector of the real-time monitoring data sequence includes:

[0056] Step S121: determining a first commonality metric coefficient between a state characterization vector of a real-time monitoring data sequence and a state characterization vector of each first vehicle operation mode; the state characterization vector of the first vehicle operation mode is set in a first probability prediction component;

[0057] Step S122: determining the first mode classification probability of the real-time monitoring data sequence being classified into each first vehicle operation mode according to the first commonality metric coefficient.

[0058] In step S121, the focus is on calculating the first commonality measurement coefficient between different vectors to measure the similarity between the real-time monitoring data sequence and each first vehicle operation mode. The first commonality measurement coefficient is a quantitative value that reflects the similarity or proximity of two vectors in the feature space.

[0059] The state representation vector of the real-time monitoring data sequence is extracted by the first state representation component of the pattern matching neural network in the previous step S11, which condenses the key characteristic information of the current vehicle operation state. The state representation vector of each first vehicle operation mode is determined in advance through a large amount of data learning and analysis, and is set in the first probability prediction component, representing the typical characteristics of each specific operation mode.

[0060] For example, assuming that the state characterization vector A=[0.8, 0.6, 0.9, 0.7] of the real-time monitoring data sequence may represent the characteristic values ​​of vehicle speed, acceleration, body posture and engine temperature. Assume that there are three first vehicle operation modes: the state characterization vector B=[0.7, 0.7, 0.5, 0.8] of the bump mode, the state characterization vector C=[0.4, 0.3, 0.6, 0.7] of the slip mode, and the state characterization vector D=[0.1, 0.2, 0.3, 0.9] of the equipment failure mode.

[0061] The vehicle-mounted system may use a variety of methods to calculate the first commonality metric coefficient, such as cosine similarity.

[0062] In step S122, after obtaining the first commonality measurement coefficients between the state characterization vector of the real-time monitoring data sequence and the state characterization vector of each first vehicle operation mode, the vehicle-mounted system then determines the first mode classification probability of the real-time monitoring data sequence being classified into each first vehicle operation mode based on these coefficients.

[0063] The first commonality measurement coefficient reflects the similarity between the real-time monitoring data sequence and each first vehicle operation mode, and the first mode classification probability is based on this similarity to quantitatively estimate the possibility of the real-time monitoring data sequence belonging to each mode.

[0064] For example, the first commonality metric coefficients of the state characterization vector of the real-time monitoring data sequence and the turbulence mode, the slip mode, and the equipment failure mode calculated by step S121 are 0.93, 0.65, and 0.32, respectively. These coefficients indicate that the real-time monitoring data sequence has the highest similarity with the turbulence mode and the lowest similarity with the equipment failure mode.

[0065] A feasible method to determine the probability of the first mode classification is to use the softmax function. The softmax function can convert a set of numerical values ​​into a vector representing the probability of each category. Its formula is:

[0066] ;

[0067] Where P(i) represents the first mode classification probability of the real-time monitoring data sequence being classified into the i-th first vehicle operation mode, z i is the first commonality metric coefficient corresponding to the i-th first vehicle operating mode, and n is the total number of first vehicle operating modes.

[0068] Assume that n=3 here, corresponding to the turbulence mode, slip mode and equipment failure mode, z1=0.93 (the first commonality metric coefficient of the turbulence mode), z2=0.65 (the first commonality metric coefficient of the slip mode), and z3=0.32 (the first commonality metric coefficient of the equipment failure mode).

[0069] First calculate the denominator: ;

[0070] Then the first mode classification probability of the real-time monitoring data sequence being classified into the turbulence mode is:

[0071] ;

[0072] The first mode classification probability classified as slip mode is:

[0073] ;

[0074] The probability of the first mode classification classified as the equipment failure mode is:

[0075] ;

[0076] In this way, the vehicle-mounted system can reasonably determine the first mode classification probability of classifying the real-time monitoring data sequence into each first vehicle operation mode based on the first commonality measurement coefficient using methods such as the softmax function.

[0077] In addition to using the softmax function, other probability conversion methods can also be used, such as logistic regression models. In practical applications, in order to improve the accuracy and stability of probability calculations, the first commonality metric coefficient can be properly normalized to avoid calculation errors caused by values ​​that are too large or too small. At the same time, data is continuously collected and updated, and the model is optimized and adjusted to adapt to changes in vehicle operation modes in different scenarios, ensuring the accuracy and reliability of the first mode classification probability.

[0078] As an implementation mode, step S13, determining the target vehicle operation mode corresponding to the real-time monitoring data sequence in each first vehicle operation mode according to the first mode classification probability, includes:

[0079] Step S131: normalizing the first mode classification probability;

[0080] Step S132: Based on the first mode classification probability and the first critical mode classification probability after the normalization process, determine the target vehicle operation mode corresponding to the real-time monitoring data sequence in each first vehicle operation mode.

[0081] In step S131, after obtaining the first mode classification probability of the real-time monitoring data sequence classified into each first vehicle operation mode, the vehicle-mounted system performs standardization processing on these probabilities. The purpose of the standardization processing is to map these probability values ​​to a specific range so that they are comparable and meet some characteristics of probability distribution, so as to facilitate subsequent accurate decision making.

[0082] The first mode classification probability is obtained by calculating the relationship between the state characterization vector of the real-time monitoring data sequence and the state characterization vector of each first vehicle operation mode in step S12. These probability values ​​may have different value ranges and magnitudes. For example, assuming that the first mode classification probabilities of the real-time monitoring data sequence classified into the bump mode, the slip mode and the equipment failure mode calculated in step S12 are P1=0.4, P2=0.3, and P3=0.2 respectively, the probability values ​​here are only an example, and these probability values ​​may vary in actual situations due to different data and calculation methods.

[0083] One of the feasible normalization methods is to use a normalization operation so that the sum of all first mode classification probabilities is 1. A feasible normalization formula is:

[0084] ;

[0085] in, is the normalized probability of the i-th first mode classification, is the original i-th first mode classification probability, and n is the total number of first vehicle operation modes.

[0086] In the above example, n=3, .

[0087] Then the normalized bump mode probability is:

[0088] ;

[0089] The normalized slip mode probability is:

[0090] ;

[0091] The normalized probability of equipment failure mode is:

[0092] ;

[0093] Through this standardization process, the probability values ​​of all modes are adjusted to a reasonable range, and their sum is 1, which meets the basic requirements of probability distribution. This processing method enables the vehicle system to more intuitively compare the probability of different modes, providing a more reliable basis for the subsequent determination of the target vehicle operation mode.

[0094] In step S132, after completing the normalization processing of the first mode classification probability, the vehicle-mounted system will then combine the first critical mode classification probability to finally determine the target vehicle operation mode corresponding to the real-time monitoring data sequence. The first critical mode classification probability is a pre-set threshold value, which is determined based on actual application requirements and experience, and is used to determine whether a first vehicle operation mode is significant enough to be determined as the target vehicle operation mode.

[0095] The vehicle system compares each normalized first mode classification probability with the first critical mode classification probability. In this example, the bump mode probability 0.44 > 0.3, the skid mode probability 0.33 > 0.3, and the equipment failure mode probability 0.22 < 0.3.

[0096] When a standardized first mode classification probability is greater than the first critical mode classification probability, it means that the real-time monitoring data sequence has a greater probability of belonging to the first vehicle operation mode. When there are multiple cases where the probability of the classification of the mode is greater than the first critical mode, the vehicle-mounted system usually selects the mode with the largest probability value as the target vehicle operation mode. In the above example, the standardized probability of the bumpy mode is the largest, so the vehicle-mounted system will determine that the target vehicle operation mode corresponding to the real-time monitoring data sequence is the bumpy mode.

[0097] As an implementation method, the method provided in this application may further include:

[0098] Step S50: acquiring first representative data of the first vehicle pattern matching data; the data capacity of the first representative data is smaller than the data capacity of the first reference data;

[0099] Step S60: determining the attribution between the first vehicle operation mode and the first representative data according to the attribution between the first vehicle operation mode and the first vehicle mode matching data;

[0100] Step S70: Debug the first machine learning model according to the attribution between the first representative data, the first vehicle operation mode and the first representative data. The first machine learning model obtained by debugging is a pattern matching neural network.

[0101] In step S50, the vehicle-mounted system obtains first representative data of the first vehicle pattern matching data when performing pattern matching neural network training. The first vehicle pattern matching data is data related to various vehicle operating modes, such as speed, acceleration, posture and other data under different operating modes. However, since the actual amount of data may be very large, directly using all the data for training may result in a waste of computing resources and excessive training time. Therefore, it is necessary to select a portion of representative data from the first vehicle pattern matching data as the first representative data, and its data capacity is smaller than the first reference data capacity.

[0102] For example, suppose the first vehicle mode matching data contains monitoring data of all transport vehicles in various operating modes at the ski resort in the past year, and the amount of data is very large. The first reference data capacity is set to 100,000 data records, and the vehicle system selects 10,000 representative data from these data as the first representative data through a certain sampling method. These first representative data should be able to reflect the characteristics and changing laws of different vehicle operating modes as much as possible.

[0103] For example, the vehicle system can use random sampling, stratified sampling and other methods to obtain the first representative data. Random sampling refers to randomly selecting a certain amount of data from the first vehicle mode matching data as the first representative data; stratified sampling refers to stratifying the data according to different vehicle operation modes or other relevant characteristics, and then selecting data from each layer in a certain proportion to form the first representative data. This ensures that the first representative data has good representativeness in different modes and characteristics.

[0104] In step S60, after acquiring the first representative data, the vehicle-mounted system determines the attribution relationship between the first vehicle operation mode and the first representative data according to the attribution between the known first vehicle operation mode and the first vehicle mode matching data. This attribution reflects the association between different vehicle operation modes and the corresponding mode matching data.

[0105] For example, it is known that in the first vehicle pattern matching data, when the vehicle is in the "bumping mode", the corresponding pattern matching data may include features such as frequent changes in vertical acceleration and unstable body posture. Then, for the selected first representative data, if the vertical acceleration and body posture features of a certain part of the data match the features of the "bumping mode", the vehicle system will determine that this part of the first representative data belongs to the "bumping mode".

[0106] The vehicle-mounted system can determine this attribution through data analysis and feature matching methods. First, the first representative data is feature extracted and analyzed to extract features related to the vehicle operation mode, such as speed, acceleration, attitude angle, etc. Then, these features are matched with the features of the known first vehicle operation mode, and the vehicle operation mode to which the first representative data belongs is determined based on the degree of matching. For example, a similarity calculation method (such as cosine similarity, Euclidean distance, etc.) can be used to measure the similarity between the features of the first representative data and the features of each first vehicle operation mode, and the first representative data is attributed to the vehicle operation mode with the highest similarity.

[0107] In step S70, after determining the first representative data and their attribution to the first vehicle operating mode, the on-board system will use this information to debug and train the first machine learning model to ultimately obtain a pattern matching neural network.

[0108] The core of the training process is to let the first machine learning model learn the characteristics of the first representative data and the relationship between them and the corresponding vehicle operation mode. For example, for a feasible first machine learning model with a multi-layer perceptron (MLP) structure, it consists of an input layer, a hidden layer, and an output layer. The input layer receives the first representative data, the hidden layer extracts and transforms the data through a series of neurons and activation functions, and the output layer outputs the predicted vehicle operation mode.

[0109] During the training process, the vehicle system will provide the first representative data as input to the first machine learning model, and the model will calculate and output the prediction result based on the current parameters. Then, the prediction result is compared with the vehicle operation mode to which the first representative data actually belongs (i.e., the known attribution situation) to calculate the prediction error. Common error calculation methods include mean square error (MSE) and so on.

[0110] Based on the calculated error, the vehicle system uses an optimization algorithm (such as a gradient descent algorithm) to adjust the parameters of the first machine learning model to reduce the error. The basic idea of ​​the gradient descent algorithm is to update the parameters of the model along the negative gradient direction of the error function so that the error gradually decreases.

[0111] The on-board system continuously repeats the above process, i.e. inputting the first representative data, calculating the prediction results and errors, and updating the model parameters, until the prediction error of the model reaches an acceptable range or the predetermined training stop condition (such as reaching the maximum number of training rounds, etc.) is met. At this point, the first machine learning model obtained after debugging and training can be used as a pattern matching neural network for subsequent analysis and processing of the real-time monitoring data sequence of the ski resort transport vehicle.

[0112] In order to improve training efficiency and model performance, optimization strategies such as batch gradient descent and stochastic gradient descent can be used. Batch gradient descent uses all the first representative data to calculate the gradient each time the model parameters are updated; stochastic gradient descent randomly selects one or a small batch of data each time to calculate the gradient and update the parameters. In addition, regularization methods (such as L1 regularization and L2 regularization) can be used to prevent model overfitting and improve the generalization ability of the model. At the same time, the training process and performance indicators of the model can be recorded during the training process for analysis and adjustment. For example, plot the curves of training error and verification error as the number of training rounds changes, observe whether the model is overfitting or underfitting, and adjust the structure and parameters of the model according to the trend of the curve.

[0113] As an implementation manner, step S70, debugging the first machine learning model according to the attribution between the first representative data, the first vehicle operation mode and the first representative data, includes:

[0114] Step S71: extracting a state representation vector of first representative data based on a first state representation component in a first machine learning model;

[0115] Step S72: determining, based on the first probability prediction component in the first machine learning model and according to the state representation vector of the first representative data, a second mode classification probability of the first representative data being classified into each first vehicle operation mode;

[0116] Step S73: Debug the parameters in the first state characterization component and the first probability prediction component according to the second mode classification probability until the matching vehicle operation mode corresponding to the first representative data represented by the second mode classification probability conforms to the attribution between the first vehicle operation mode and the first representative data.

[0117] In step S71, when debugging the first machine learning model, the vehicle-mounted system first uses the first state representation component in the model to process the first representative data. The first state representation component is a part of the first machine learning model that is specifically responsible for extracting key features from input data and converting it into a vector form that can effectively represent the data state.

[0118] The first representative data contains various information related to the operation mode of the ski resort transport vehicle, such as speed, acceleration, vehicle posture, equipment parameters, etc. The first state representation component analyzes and processes this data and extracts the features that best reflect the vehicle's operating state.

[0119] For example, suppose the first representative data records the information of a ski resort transporter at a certain moment: the speed is 35 km / h, the acceleration is 0.3 m / s², the body posture angle is within the normal range (set to a value of 1), and the engine temperature is 75 degrees Celsius. The first state representation component may use specific algorithms and calculation processes to extract features from these data. For example, it may process the speed and acceleration data through convolution operations (similar to convolution in image processing, which is used to extract local features in the data) to capture the changing trends and fluctuation characteristics of the data; for the body posture angle, some mathematical transformations may be used to highlight its stability characteristics; for the engine temperature, the relative changes in temperature may be considered. After these processes, these features are combined to form a state representation vector.

[0120] Assume that after being processed by the first state representation component, the state representation vector obtained is [0.7, 0.5, 0.9, 0.6]. Among them, 0.7 represents the extracted value of the speed-related feature, reflecting the characteristics of the speed within a specific range; 0.5 represents the extracted value of the acceleration-related feature, reflecting the change of acceleration; 0.9 represents the extracted value of the body posture angle-related feature, indicating the stability of the body posture; 0.6 represents the extracted value of the engine temperature-related feature, showing the relative state of the engine temperature.

[0121] The first state representation component can be implemented based on a variety of technologies, such as convolutional neural networks (CNN), recurrent neural networks (RNN) and their variants (such as long short-term memory networks LSTM, gated recurrent units GRU), etc. Taking CNN as an example, it contains multiple convolutional layers, pooling layers, and fully connected layers. The convolution layer performs convolution operations by sliding the convolution kernel on the data to extract local features of the data; the pooling layer is used to downsample the data to reduce the data dimension while retaining the main features; the fully connected layer integrates the data processed by convolution and pooling, and outputs the final state representation vector. In practical applications, the first state representation component needs to be parameterized and optimized to ensure that it can accurately extract valuable features. The backpropagation algorithm can be used to calculate the gradient and update the parameters in the component according to the gradient, so that the component can generate a more effective state representation vector when processing the first representative data.

[0122] In step S72, after obtaining the state representation vector of the first representative data, the vehicle-mounted system inputs it into the first probability prediction component in the first machine learning model. The function of the first probability prediction component is to calculate the probability of the first representative data being classified into each first vehicle operation mode, that is, the second mode classification probability, based on the input state representation vector.

[0123] The first probability prediction component is usually composed of a series of neurons and connection weights, and processes the input state representation vector through complex mathematical operations and nonlinear transformations. It analyzes the state representation vector, evaluates the degree of match between it and the characteristics of each first vehicle operation mode, and converts this degree of match into a probability value.

[0124] For example, suppose there are three first vehicle operation modes: bump mode, skidding mode, and equipment failure mode. After receiving the state representation vector [0.7, 0.5, 0.9, 0.6], the first probability prediction component will process it. The neurons in the component will perform weighted summation on each element in the vector according to the pre-set weights, and perform nonlinear transformation through activation functions (such as Sigmoid function, ReLU function, etc.).

[0125] In step S73, after obtaining the second mode classification probability of the first representative data being classified into each first vehicle operation mode, the vehicle-mounted system will compare these probabilities with the known attribution between the first vehicle operation mode and the first representative data. If there is a deviation between the two, the vehicle-mounted system will debug the parameters in the first state representation component and the first probability prediction component according to the second mode classification probability to reduce the deviation until the matching vehicle operation mode corresponding to the first representative data represented by the second mode classification probability meets the attribution between the first vehicle operation mode and the first representative data.

[0126] For example, suppose it is known that a certain first representative data actually belongs to the "bumping mode", and among the second mode classification probabilities calculated by the first probability prediction component, the probability of "bumping mode" is 0.61, the probability of "slipping mode" is 0.30, and the probability of "equipment failure mode" is 0.09. Although the probability of "bumping mode" is the highest, it may not match the actual belonging situation well enough, or there may be a large deviation when processing other first representative data.

[0127] The vehicle system adjusts the parameters in the first state representation component and the first probability prediction component according to the difference between the second mode classification probability and the actual attribution situation. This adjustment process can be based on the gradient descent algorithm or its variants. Taking the gradient descent algorithm as an example, it is first necessary to define a loss function to measure the difference between the prediction result (second mode classification probability) and the actual attribution situation. Commonly used loss functions such as the cross entropy loss function are not described here.

[0128] The vehicle system calculates the gradient of the loss function with respect to the parameters in the first state representation component and the first probability prediction component, and then adjusts the parameters according to the update rules of the gradient descent algorithm based on the direction and magnitude of the gradient.

[0129] The vehicle system continuously repeats this process, i.e., calculating the loss function, calculating the gradient, and updating the parameter variables, until the matching vehicle operation mode corresponding to the first representative data represented by the second mode classification probability is consistent with the attribution between the first vehicle operation mode and the first representative data. For example, after multiple adjustments, for various first representative data, the vehicle operation mode corresponding to the highest probability predicted by the model is consistent with the actual attribution. At this time, it can be considered that the parameter variables in the first state representation component and the first probability prediction component have been adjusted to appropriate values, and the first machine learning model has become a pattern matching neural network that can accurately identify and classify vehicle operation modes after debugging.

[0130] As an implementation manner, step S72, determining the second mode classification probability of the first representative data being classified into each first vehicle operation mode according to the state characterization vector of the first representative data, includes:

[0131] Step S721: determining a second commonality metric coefficient between the state characterization vector of the first representative data and the state characterization vector of each first vehicle operation mode; the state characterization vector of the first vehicle operation mode is pre-set in the first probability prediction component; the state characterization vector of the first vehicle operation mode is optimized in the debugging link of the neural network;

[0132] Step S722: Determine the second mode classification probability of the first representative data being classified into each first vehicle operation mode according to the second commonality metric coefficient.

[0133] In step S721, the main task is to calculate the second commonality measurement coefficient between the state characterization vector of the first representative data and the state characterization vector of each first vehicle operation mode. This coefficient is used to measure the similarity between the two and is an important basis for subsequently determining the second mode classification probability.

[0134] The state representation vector of the first representative data is a vector obtained by extracting and transforming the features of the first representative data through the first state representation component. It contains key feature information related to the vehicle operation state in the first representative data. For example, assuming that the first representative data records the speed of the ski resort transport vehicle at a certain moment as 40 km / h, the acceleration as 0.2 m / s², the body posture angle as 1.2 (within the normal range), and the engine temperature as 80 degrees Celsius, after being processed by the first state representation component, the state representation vector obtained may be [0.6, 0.4, 0.8, 0.7], where each value corresponds to the extraction value of different features.

[0135] The state characterization vector of the first vehicle operation mode is pre-set in the first probability prediction component, and it represents the typical characteristics of each vehicle operation mode. For example, for the "bumping mode", its state characterization vector may be [0.8, 0.6, 0.4, 0.5], indicating the typical characteristic values ​​of vehicle speed, acceleration, body posture angle and engine temperature in the bumping mode; for the "slipping mode", its state characterization vector may be [0.3, 0.1, 0.7, 0.6]. These state characterization vectors are obtained based on the analysis and summary of a large amount of historical data, and will be continuously optimized in the debugging link of the neural network to improve its representativeness and accuracy of the vehicle operation mode.

[0136] There are many methods for calculating the second commonality measure coefficient, and a feasible method is cosine similarity.

[0137] In step S722, after calculating the second commonality measurement coefficients between the state characterization vector of the first representative data and the state characterization vector of each first vehicle operation mode, the vehicle system determines the second mode classification probability of the first representative data being classified into each first vehicle operation mode based on these coefficients.

[0138] The second commonality measurement coefficient reflects the similarity between the first representative data and each first vehicle operation mode, and the second mode classification probability quantitatively estimates the possibility that the first representative data belongs to each mode based on the similarity.

[0139] One feasible method is to convert the second commonality measure coefficient into a probability value through the softmax function. The softmax function can convert a set of real numbers into a vector representing the probability of each category, and the sum of these probability values ​​is 1. I will not go into details here.

[0140] As an implementation mode, step S73, debugging the parameter variables in the first state representation component and the first probability prediction component according to the second mode classification probability until the matching vehicle operation mode corresponding to the first representative data represented by the second mode classification probability conforms to the attribution between the first vehicle operation mode and the first representative data, includes:

[0141] Step S731: normalizing the second mode classification probability;

[0142] Step S732: determining a matching vehicle operating mode corresponding to the first representative data in each first vehicle operating mode based on the second mode classification probability and the second critical mode classification probability after the normalization process;

[0143] Step S733: Debug the parameters in the first state characterization component and the first probability prediction component according to the matching vehicle operation mode corresponding to the first representative data until the matching vehicle operation mode corresponding to the first representative data conforms to the attribution between the first vehicle operation mode and the first representative data.

[0144] In step S731, after obtaining the second mode classification probability of the first representative data classified into each first vehicle operation mode, the vehicle system firstly performs standardization processing on these probabilities. This is because in actual calculation, the value range and distribution of the second mode classification probability may not meet the requirements of subsequent decision-making or model optimization. Standardization processing can convert these probability values ​​into a more appropriate range, making them more comparable and usable.

[0145] The second mode classification probability is calculated by step S72, reflecting the probability that the first representative data belongs to each first vehicle operation mode. For example, assuming that after calculation in step S72, the second mode classification probability of a first representative data being classified into "bumping mode", "slipping mode" and "equipment failure mode" is P_1=0.4, P_2=0.3, and P_3=0.2 respectively. The probability values ​​here are only examples, and in actual situations these probability values ​​will vary depending on the data and calculation methods.

[0146] One of the feasible standardization methods is to use a normalization operation so that the sum of all second mode classification probabilities is 1.

[0147] In step S732, after completing the normalization process of the second mode classification probability, the vehicle system will then determine the matching vehicle operation mode corresponding to the first representative data in combination with the second critical mode classification probability. The second critical mode classification probability is a pre-set threshold value, which is determined based on actual application requirements and experience, and is used to determine whether a first vehicle operation mode is significant enough to be determined as a matching mode for the first representative data.

[0148] In step S733, after determining the matching vehicle operation mode corresponding to the first representative data, the vehicle system needs to compare this result with the known attribution between the first vehicle operation mode and the first representative data. If the two are inconsistent, it means that the parameters in the first state representation component and the first probability prediction component may need to be adjusted to make the prediction result of the model more consistent with the actual situation.

[0149] For example, it is known that a certain first representative data actually belongs to the "equipment failure mode", but the matching vehicle operation mode determined in the previous step is the "bumping mode", which results in a mismatch. The vehicle system will debug the parameters in the first state representation component and the first probability prediction component based on this difference.

[0150] This debugging process is usually based on optimization algorithms, such as the gradient descent algorithm and its variants. Taking the gradient descent algorithm as an example, a loss function is first defined to measure the difference between the prediction result (the matching vehicle operation mode corresponding to the first representative data) and the actual attribution situation. For example, the cross entropy loss function. The vehicle-mounted system calculates the gradient of the loss function with respect to the parameter variables in the first state representation component and the first probability prediction component. Then, according to the direction and size of the gradient, the parameter variables are adjusted according to the update rule of the gradient descent algorithm. The vehicle-mounted system will continue to repeat this process, that is, calculate the loss function, calculate the gradient, and update the parameter variables until the matching vehicle operation mode corresponding to the first representative data is consistent with the attribution situation between the first vehicle operation mode and the first representative data. For example, after multiple adjustments, for various first representative data, the matching vehicle operation modes predicted by the model are consistent with the actual attribution situation. At this time, it can be considered that the parameter variables in the first state representation component and the first probability prediction component have been adjusted to appropriate values, and the performance of the first machine learning model in this regard has been optimized.

[0151] As an implementation method, the method provided in this application may further include:

[0152] Step S101: acquiring second representative data of the first vehicle pattern matching data; the data capacity of the second representative data is greater than the data capacity of the second reference data;

[0153] Step S102: extracting a state representation vector of the second representative data based on a vehicle state feature extraction network;

[0154] Step S103: performing cluster analysis on the first vehicle mode matching data according to the state characterization vector of the second representative data to obtain the attribution between the first vehicle operation mode and the first vehicle mode matching data.

[0155] In step S101, the second representative data of the first vehicle pattern matching data is obtained. The first vehicle pattern matching data includes data related to various operating states of the ski resort transport vehicle, such as the speed, acceleration, attitude angle, temperature of each component, etc. of the vehicle under different road conditions. In order to more comprehensively and accurately analyze and determine the attribution between the vehicle operating mode and the pattern matching data, it is necessary to obtain sufficient and representative data, namely the second representative data, and its data capacity must be greater than the second reference data capacity.

[0156] For example, suppose the first vehicle pattern matching data is the monitoring data of all transport vehicles at the ski resort in the past year, which contains detailed information under various operating scenarios. The second reference data capacity is set to 5,000 data records, then the vehicle system will select a part of the data with a data capacity greater than 5,000 from these data as the second representative data. For example, 8,000 data are selected. These data should cover as many different operating conditions as possible, such as data under different weather conditions, road conditions, vehicle loads, etc., to ensure the accuracy and comprehensiveness of subsequent analysis.

[0157] In step S102, after obtaining the second representative data, the vehicle system uses the vehicle state feature extraction network to extract the state representation vector of these data. The vehicle state feature extraction network is a neural network model specially designed to extract key features from raw data. It can convert complex, high-dimensional raw data into low-dimensional, representative state representation vectors for subsequent analysis and processing.

[0158] For example, a record in the second representative data contains information such as the vehicle speed of 50 km / h, acceleration of 0.3 m / s², body posture angle of 1.1 (within the normal range), and engine temperature of 78 degrees Celsius. The vehicle state feature extraction network processes this data. It may contain multiple convolutional layers, pooling layers, and fully connected layers. The convolution layer performs convolution operations by sliding the convolution kernel on the data to extract local features in the data, such as capturing the changing trend of speed and acceleration data, and the stability characteristics of the body posture angle; the pooling layer is used to downsample the data to reduce the data dimension while retaining the main features; the fully connected layer integrates the data after convolution and pooling, and finally outputs a state representation vector. Assume that after being processed by the vehicle state feature extraction network, the obtained state representation vector is [0.75, 0.45, 0.85, 0.65], where each value corresponds to the extraction value of a different feature, such as 0.75 represents the extraction value of speed-related features, 0.45 represents the extraction value of acceleration-related features, 0.85 represents the extraction value of body posture angle-related features, and 0.65 represents the extraction value of engine temperature-related features.

[0159] During the training process, a large amount of labeled data is required, that is, a data set with known vehicle operating states and corresponding features, so that the network can learn how to extract effective features from the data. The training process usually uses a back-propagation algorithm to adjust the parameters in the network to minimize the error between the predicted results and the true annotations. For example, a loss function (such as a mean square error loss function) is defined to measure the difference between the predicted state representation vector and the true labeled state representation vector, and then the network parameters are updated through a gradient descent algorithm to gradually reduce the value of the loss function.

[0160] In step S103, after obtaining the state representation vectors of the second representative data, the vehicle-mounted system performs cluster analysis (clustering) on ​​the first vehicle mode matching data based on these vectors to determine the attribution between the first vehicle operation mode and the first vehicle mode matching data. Clustering is an unsupervised learning method that divides data objects into different classes or clusters, so that data objects within the same cluster have high similarity, while data objects between different clusters have large differences.

[0161] For example, suppose there are three feasible first vehicle operation modes: bump mode, slip mode, and equipment failure mode. The vehicle-mounted system will cluster the first vehicle mode matching data according to the characteristics of the state characterization vector of the second representative data. For the bump mode, its corresponding state characterization vector may have obvious characteristics in terms of body posture angle and vertical acceleration, such as large fluctuations in body posture angle and frequent changes in vertical acceleration; for the slip mode, it may have unique performance in terms of wheel speed and lateral acceleration; for the equipment failure mode, abnormal values ​​may appear in terms of engine temperature and operating parameters of key components. The vehicle-mounted system can use a variety of clustering algorithms to implement cluster analysis, such as K-Means algorithm, DBSCAN algorithm, etc. Taking the K-Means algorithm as an example, its basic idea is to divide the data into K clusters, where K is the number of pre-set clusters. The algorithm first randomly selects K initial cluster centers, and then assigns each data point to the cluster with the closest cluster center. Then, the cluster center is recalculated based on the data points in each cluster, and this process is repeated until the cluster center no longer changes or the predetermined number of iterations is reached. Assume that the state representation vector of the second representative data is clustered by the K-Means algorithm, and the first vehicle mode matching data is divided into three clusters. After analysis, it is found that the characteristics of the data in one cluster in terms of body posture angle and vertical acceleration match the typical characteristics of the bump mode, so it can be considered that the first vehicle mode matching data in this cluster belongs to the bump mode; the characteristics of the data in another cluster in terms of wheel speed and lateral acceleration meet the characteristics of the slip mode, so the data in this cluster belongs to the slip mode; the data in the remaining cluster is abnormal in terms of engine temperature, etc., which is consistent with the characteristics of the equipment failure mode, so the data in this cluster is classified as the equipment failure mode.

[0162] Through this cluster analysis, the vehicle-mounted system obtains the attribution between the first vehicle operation mode and the first vehicle mode matching data.

[0163] In addition to the K-Means algorithm, other appropriate clustering algorithms can be selected according to the characteristics of the data and actual needs. In practical applications, in order to improve the accuracy and stability of clustering, the data can be preprocessed, such as normalization, to map the values ​​of each dimension of the data to a specific interval to avoid the impact of the magnitude difference of the data on the clustering results. At the same time, the appropriate clustering parameters and algorithms can be selected by evaluating the indicators of the clustering results (such as the silhouette coefficient, etc.) to obtain the best clustering effect.

[0164] As an implementation method, the method provided in this application may further include:

[0165] Step S201: acquiring vehicle pattern matching training data and training representative data of the vehicle pattern matching training data; the data capacity of the training representative data is greater than the third reference data capacity;

[0166] Step S202: Debugging the second machine learning model according to the training representative data of the vehicle pattern matching training data, and using the debugged second machine learning model to extract the training state representation vector of the training representative data and classify the training representative data of the same vehicle pattern matching training data into the same vehicle pattern matching training data according to the training state representation vector;

[0167] Step S203: Generate a vehicle state feature extraction network based on the second machine learning model obtained through debugging.

[0168] In step S201, the vehicle system obtains vehicle pattern matching training data and related training representative data. Vehicle pattern matching training data is a data set used to train models to identify and classify different vehicle operating modes. It contains rich vehicle operating status information, such as speed, acceleration, body posture, equipment parameters, etc. under different road conditions. In order to make the trained model have better generalization ability and accuracy, it is necessary to select enough and representative data from the vehicle pattern matching training data as training representative data, and its data capacity must be greater than the third reference data capacity.

[0169] For example, suppose the vehicle pattern matching training data is the detailed monitoring data of all transport vehicles at the ski resort in various operating scenarios in the past two years. The amount of data is very large. The third reference data capacity is set to 10,000 data records, then the vehicle system will select a part of the data with a data capacity greater than 10,000 from these massive data as training representative data. For example, 15,000 data are selected. These data should cover as many possible operating conditions as possible, including different weather conditions, different driving sections, different vehicle loads, etc., to ensure that the trained model can adapt to various actual situations.

[0170] In order to ensure the representativeness and randomness of the data, random sampling or stratified sampling methods can be used. Random sampling is to randomly select a certain number of data from all the data; stratified sampling is to first stratify the data according to certain characteristics (such as season, time period, type of transportation task, etc.), and then select data from each layer according to a certain proportion to form representative training data.

[0171] In step S202, after obtaining the training representative data, the vehicle system will debug the second machine learning model based on the data. The goal of the second machine learning model is to learn how to extract effective features from the training representative data, namely, training state representation vectors, and classify the training representative data with the same features into the same vehicle pattern matching training data category based on these vectors.

[0172] For example, suppose the training representative data contains data of vehicles in different situations such as normal driving, bumpy driving, skidding driving, and equipment failure. For normal driving data, its speed, acceleration, body posture and other parameters may be relatively stable; while in bumpy driving, the body posture angle and vertical acceleration may fluctuate greatly; in skidding driving, the wheel speed and lateral acceleration may change abnormally; in equipment failure, the parameters of related equipment (such as engine temperature, oil pressure, etc.) may exceed the normal range. The second machine learning model needs to learn the characteristics of the data in these different modes and be able to accurately extract the training state representation vector.

[0173] During the debugging process, the vehicle system inputs the training representative data into the second machine learning model, and the model processes the data according to the current parameters, trying to extract the training state representation vector and classify it. The vehicle system then compares the output of the model with the known true category label and calculates the error. Depending on the size of the error, an optimization algorithm (such as the gradient descent algorithm) is used to adjust the parameters of the model to reduce the error. This process is repeated until the performance of the model reaches a satisfactory level.

[0174] For example, in one iteration, after the model processed a portion of the training representative data, the training state representation vector obtained incorrectly classified some data that should belong to the bumpy driving mode as normal driving mode. After the vehicle system discovered this problem by calculating the error, it would adjust the model parameters according to the gradient descent algorithm, so that the model can more accurately identify the characteristics of the bumpy driving mode in the next iteration, thereby improving the accuracy of classification.

[0175] Second, the machine learning model can adopt a variety of deep learning architectures, such as convolutional neural networks (CNN), recurrent neural networks (RNN) and their variants. Taking CNN as an example, it extracts data features through structures such as convolutional layers, pooling layers, and fully connected layers. The convolution layer performs convolution operations by sliding the convolution kernel on the data to capture local features in the data; the pooling layer is used to downsample the data to reduce the data dimension while retaining the main features; the fully connected layer integrates the data processed by convolution and pooling, and outputs the training state representation vector. During the training process, the back propagation algorithm can be used to calculate the gradient and update the parameters of the model based on the gradient. At the same time, in order to prevent the model from overfitting, some regularization methods can be used, such as L1 regularization, L2 regularization, etc.

[0176] In step S203, after the debugging of the second machine learning model is completed, the vehicle system generates a vehicle state feature extraction network based on the debugged model. This network is trained and optimized, and can accurately extract the training state representation vectors of the vehicle pattern matching training data, and effectively classify and categorize the data based on these vectors.

[0177] For example, after multiple iterations of debugging, the second machine learning model can accurately identify the features of the training representative data under different vehicle operation modes and correctly classify them into the corresponding modes. At this point, the vehicle system will use this debugged model as a basis to build a vehicle state feature extraction network. This network can be regarded as an optimized feature extractor, which can extract the most representative features from the input vehicle mode matching training data to form a training state representation vector, providing strong support for subsequent analysis and processing.

[0178] To summarize, steps S201-S203 debug the second machine learning model by acquiring vehicle pattern matching training data and training representative data, and generate a vehicle state feature extraction network based on the debugging results, which provides important technical support for accurately extracting the characteristics of the vehicle's operating status and performing effective pattern classification, and helps to improve the accuracy and reliability of the emergency handling method of ski resort transport vehicles.

[0179] As an implementation manner, step S201, obtaining vehicle pattern matching training data and training representative data of the vehicle pattern matching training data, includes:

[0180] Step S2011: acquiring vehicle pattern matching training data and third representative data of the vehicle pattern matching training data;

[0181] Step S2012: performing noise enhancement on the third representative data according to the vehicle state characteristics of the third representative data to obtain fourth representative data;

[0182] Step S2013: Generate training representative data according to the third representative data and the fourth representative data.

[0183] In step S2011, vehicle mode matching training data and third representative data related thereto are obtained. The vehicle mode matching training data is a rich data set that covers detailed information of ski resort transport vehicles in various operating modes. These data are derived from various sensors installed on the vehicle, including but not limited to speed sensors, acceleration sensors, gyroscopes, temperature sensors, etc., which record various parameters in real time during the operation of the vehicle.

[0184] For example, the speed sensor will provide real-time feedback on the vehicle's driving speed, the acceleration sensor can monitor the vehicle's acceleration, deceleration, and dynamic acceleration changes during driving, the gyroscope can accurately measure the vehicle's posture angle, and the temperature sensor monitors the temperature of key components such as the engine and brake system. These data show different characteristics in different operating modes. For example, in normal driving mode, the speed is relatively stable, the acceleration fluctuates slightly, the body posture remains stable, and the temperature of key components is within the normal range; in bumpy mode, the acceleration changes frequently, and the body posture angle fluctuates greatly; in slip mode, the speed and wheel speed may have an abnormal relationship, and the lateral acceleration will change significantly.

[0185] The third representative data is a representative subset of data selected from the vehicle pattern matching training data. The purpose of the vehicle-mounted system to obtain the third representative data is to process and analyze subsequent data more efficiently, and to reduce the amount of calculation to a certain extent. For example, suppose the vehicle pattern matching training data contains the operation records of all transport vehicles in different time periods and road conditions in the past month, and the amount of data is very large. The vehicle-mounted system may select a part of these data as the third representative data according to specific rules or random sampling. For example, according to different time periods of the day (morning, noon, and evening) and different road conditions (flat snow roads, undulating snow roads, and curved snow roads), stratification is performed, and then a certain amount of data is randomly selected from each layer to form the third representative data. This ensures that the third representative data can cover the vehicle operation characteristics under different circumstances.

[0186] In step S2012, after obtaining the third representative data, the vehicle system performs noise enhancement operation on the data according to the vehicle state characteristics, thereby obtaining fourth representative data. Noise enhancement is a data enhancement technology, which aims to increase the diversity and complexity of the data by adding a certain amount of noise to the original data, so that the model can learn richer features during the training process and improve the generalization ability of the model.

[0187] Vehicle state characteristics refer to various parameters and characteristic values ​​that can reflect the operating status of the vehicle, such as the speed, acceleration, attitude angle, temperature, etc. mentioned above. The on-board system analyzes the distribution and changes of these vehicle state characteristics in the third representative data, and then adds noise according to specific rules. In this way, each vehicle state characteristic of each record in the third representative data is noised to obtain the fourth representative data. The fourth representative data after noise enhancement increases the variability of the data while retaining the characteristics of the original data, so that the model can be exposed to more different situations during training, which helps to improve the model's adaptability to various actual scenarios.

[0188] In step S2013, after obtaining the third representative data and the fourth representative data enhanced by noise, the vehicle-mounted system generates training representative data based on the two sets of data. The process of generating training representative data is to reasonably combine the third representative data and the fourth representative data to make full use of the characteristics of the original data and the diversity of the data after noise.

[0189] For example, the vehicle system can merge the third representative data and the fourth representative data according to a certain ratio. Assume that the set ratio is 70% of the third representative data and 30% of the fourth representative data. First, randomly select 70% of the data records from the third representative data, and then randomly select 30% of the data records from the fourth representative data, and merge these two parts of data together to form training representative data.

[0190] The training representative data generated in this way not only contains the typical vehicle operation status characteristics in the original data, but also incorporates data with more variability after noise enhancement. For example, in normal driving mode, the normal driving records in the third representative data can provide stable and typical feature samples, while the normal driving records after noise addition in the fourth representative data can provide some samples of minor changes or interference that may occur in actual operation. The same is true for other operating modes (such as bump mode, slip mode, etc.). Through this combination, the training representative data can more comprehensively cover various situations in the vehicle operation process.

[0191] To summarize, steps S2011-S2013 obtain vehicle pattern matching training data and the third representative data, perform noise enhancement on the third representative data to obtain the fourth representative data, and generate training representative data based on these two sets of data, thereby providing a rich, diverse and representative data basis for the subsequent debugging of the second machine learning model, helping to improve the performance and accuracy of the model, and thereby enhancing the effectiveness of the entire ski resort transport vehicle emergency handling method.

[0192] As an implementation mode, step S202, debugging the second machine learning model according to the training representative data of the vehicle pattern matching training data, includes:

[0193] Step S2021: extracting a training state representation vector of the training representative data based on the second state representation component in the second machine learning model;

[0194] Step S2022: Based on the second probability prediction component in the second machine learning model, and according to the training state representation vector of the training representative data, determining the third mode classification probability of the training representative data being classified into each vehicle mode matching training data;

[0195] Step S2023: Debug the parameters in the second state characterization component and the second probability prediction component according to the third mode classification probability until the training representative data representing the same vehicle mode matching training data by the third mode classification probability is classified into the same vehicle mode matching training data.

[0196] In step S2021, the vehicle system uses the second state representation component in the second machine learning model to process the training representative data to extract its training state representation vector. The second state representation component is used to learn and extract a vector representation that can represent the data features from the input data. This vector representation can capture key information in the data for subsequent analysis and processing.

[0197] For example, assume that the vehicle pattern matching training data contains various monitoring data of ski resort transport vehicles in different scenarios, such as speed, acceleration, temperature, pressure, etc. The training representative data is a data set for training obtained after specific processing (such as the processing described in step S201). The second state representation component may use the convolutional neural network (CNN) structure in deep learning to extract features. For one-dimensional time series data (such as the sequence of vehicle speed changes over time), a one-dimensional convolutional layer can be used. Suppose the input training representative data sequence is , where n is the sequence length, the convolution kernel size of the one-dimensional convolution layer is k, and the step size is s, then the convolution operation can be expressed as:

[0198] ;

[0199] Among them, w j is the weight parameter of the convolution kernel, y i is the output of the i-th position after the convolution operation. Through multiple layers of convolution and pooling operations, the second state representation component can gradually extract more advanced features and finally generate a training state representation vector v, which is a low-dimensional vector that can effectively represent the characteristics of the training representative data.

[0200] In step S2022, after obtaining the training state representation vector, the vehicle system determines the third mode classification probability of the training representative data being classified into each vehicle mode matching training data through the second probability prediction component. The main task of the second probability prediction component is to calculate the possibility of the training representative data belonging to each vehicle mode matching training data category based on the training state representation vector.

[0201] For example, suppose there are m different vehicle mode matching training data categories, representing different vehicle operating state modes, such as normal operating mode, bumpy mode, fault mode, etc. For a given training state representation vector v, the second probability prediction component can use a fully connected neural network to calculate the probability that it belongs to each category. Let the weight matrix of the fully connected layer be , where d is the dimension of the training state representation vector and the bias vector is , then the formula for calculating the third mode classification probability p through the softmax function is:

[0202] ;

[0203] Among them, z i is the i-th element of the fully connected layer output, p i Indicates the probability that the training representative data belongs to the i-th vehicle pattern matching training data category. In this way, the vehicle-mounted system can obtain the third mode classification probability distribution of the training representative data for each vehicle pattern matching training data category.

[0204] Step S2023 is an iterative optimization process, in which the vehicle system adjusts the parameters in the second state representation component and the second probability prediction component according to the third mode classification probability to make the prediction result of the model more accurate. Specifically, the goal is to enable the third mode classification probability to correctly classify the training representative data of the same vehicle mode matching training data into the same vehicle mode matching training data category.

[0205] For example, in each iteration, the vehicle system first calculates the loss function to measure the difference between the model prediction result and the true label. A commonly used loss function can be the cross entropy loss function. Let the true label be y, where Indicates that the training representative data belongs to the i-th category, Indicates that it does not belong. Then the cross entropy loss function L can be expressed as:

[0206] ;

[0207] Then, the vehicle system uses an optimization algorithm (such as the stochastic gradient descent algorithm SGD) to update the parameter variables in the second state representation component and the second probability prediction component. (including weights and biases), the update formula can be expressed as:

[0208] ;

[0209] in, is the parameter value of the current iteration, is the learning rate, is the gradient of the loss function with respect to the parameter. By continuously calculating the loss function and updating the parameter, the vehicle system gradually optimizes the third mode classification probability until the stopping condition is met, such as the value of the loss function is less than a preset threshold, or the number of iterations reaches the maximum limit. At this point, the model can accurately classify the training representative data of the same vehicle mode matching training data into the same vehicle mode matching training data category, completing the debugging process of the second machine learning model.

[0210] During the entire debugging process, the on-board system continuously adjusts the model's parameters based on the third-mode classification probability to improve the model's classification accuracy for the training representative data, so that the model can better learn the characteristics and rules of the vehicle pattern matching training data, providing a reliable model foundation for subsequent vehicle status feature extraction and analysis.

[0211] As an implementation manner, step S2022, determining the third mode classification probability of the training representative data being classified into each vehicle mode matching training data according to the training state representation vector of the training representative data, includes:

[0212] Step S20221: determining a third commonality metric coefficient between a training state representation vector of the training representative data and a state representation vector of each vehicle mode matching training data; the state representation vector of the vehicle mode matching training data is pre-set in the second probability prediction component; the state representation vector of the vehicle mode matching training data is optimized in the debugging link of the neural network;

[0213] Step S20222: Determine the third mode classification probability of the training representative data being classified into each vehicle mode matching training data according to the third commonality metric coefficient.

[0214] In step S20221, the third commonality metric coefficient between the training state representation vector of the training representative data and the state representation vector of each vehicle mode matching training data is calculated. The third commonality metric coefficient is used to measure the degree of similarity between the two vectors, which reflects the proximity between the training representative data and each vehicle mode matching training data in the feature space.

[0215] For example, assuming that the vehicle system has obtained the training state representation vector of the training representative data through the second state representation component , whose dimension is d, that is At the same time, for each vehicle mode matching training data, there is also a corresponding state representation vector , where i represents different vehicle pattern matching training data categories, .

[0216] One method to calculate the third commonality measure coefficient is cosine similarity. Cosine similarity measures the similarity of two vectors by calculating the cosine value of the angle between them. The value range is between [-1, 1]. The closer the value is to 1, the more similar the two vectors are, and the closer it is to -1, the less similar they are. The calculation formula of cosine similarity is:

[0217] ;

[0218] in, Representation vector The dot product of Represents vectors Through this formula, the vehicle system can calculate the cosine similarity between the training state representation vector of the training representative data and the state representation vector of each vehicle pattern matching training data, that is, the third commonality metric coefficient.

[0219] For example, suppose there is a training state representation vector v representing the training data t =[1, 2, 3], and the state representation vectors v1=[2, 4, 6] and v2=[-1, -2, -3] of the two vehicle pattern matching training data. First, calculate v t Cosine similarity with v1:

[0220] ;

[0221] ;

[0222] ;

[0223] ;

[0224] Then calculate v t Cosine similarity with v2:

[0225] ;

[0226] ;

[0227] ;

[0228] From the calculation results, we can see that v t The cosine similarity with v1 is 1, indicating that they are very similar; tThe cosine similarity with v2 is -1, indicating that they are completely opposite.

[0229] In step S20222, after calculating the third commonality metric coefficients, the vehicle system determines the third mode classification probability of the training representative data being classified into each vehicle mode matching training data according to these coefficients. The third mode classification probability indicates the probability of the training representative data belonging to each vehicle mode matching training data category.

[0230] One feasible method is to obtain the third mode classification probability by normalizing the third commonality metric coefficient. Assume that the third commonality metric coefficients of the training representative data and the m vehicle mode matching training data have been calculated and are These commonality measurement coefficients can be converted into probability values ​​using the softmax function, and the calculation formula of the softmax function is:

[0231] ;

[0232] Among them, p i It represents the probability of the third mode classification of the training representative data being classified into the i-th vehicle mode matching training data. Through the softmax function, the vehicle system can convert the third commonality metric coefficient into a probability value in the range of [0, 1], and the sum of all probability values ​​is 1, which meets the definition of probability.

[0233] Through the above steps, the on-board system can determine the third mode classification probability of the training representative data being classified into each vehicle mode matching training data based on the third commonality measurement coefficient between the training state representation vector of the training representative data and the state representation vector of each vehicle mode matching training data, providing an important basis for subsequent model debugging and classification tasks.

[0234] As an implementation mode, step S2023, debugging the parameter variables in the second state characterization component and the second probability prediction component according to the third mode classification probability until the training representative data representing the same vehicle mode matching training data by the third mode classification probability is classified into the same vehicle mode matching training data, includes:

[0235] Step S20231: normalizing the third mode classification probability;

[0236] Step S20232: determining matching vehicle pattern matching data corresponding to the training representative data in each vehicle pattern matching training data based on the third pattern classification probability and the third critical pattern classification probability after the normalization process;

[0237] Step S20233: Debugging the parameters in the second state characterization component and the second probability prediction component according to the matching vehicle pattern matching data corresponding to the training representative data, until the training representative data of the same vehicle pattern matching training data are classified into the same vehicle pattern matching training data.

[0238] In step S20231, the third mode classification probability is standardized, the purpose of which is to adjust the probability value to an appropriate range so that different probability values ​​are comparable and conform to the characteristics of probability distribution. Standardization helps to make accurate decisions and adjust model parameters based on these probability values. Please refer to the above description for the standardization method, which will not be repeated here.

[0239] In step S20232, after completing the normalization processing of the third mode classification probability, the vehicle-mounted system compares it with the third critical mode classification probability to determine the matching vehicle mode matching data corresponding to the training representative data. The third critical mode classification probability is a pre-set threshold value used to determine whether the normalized probability is high enough to determine the category to which the training representative data belongs.

[0240] For example, suppose the third critical mode classification probability is , the standardized third mode classification probability is The vehicle system will traverse all probability values, and when a When , the i-th vehicle pattern matching training data is determined as the matching vehicle pattern matching data corresponding to the training representative data.

[0241] For example, assuming m=5, the standardized third mode classification probabilities are , the third critical mode classification probability = 0.3. In this case, only , so the vehicle system will determine the first vehicle pattern matching training data as the matching vehicle pattern matching data corresponding to the training representative data.

[0242] If all normalized probabilities are less than the third critical mode classification probability, the on-board system may adopt some additional strategies, such as marking the training representative data as the “uncertain” category, or further adjusting the model parameters to improve the classification accuracy.

[0243] In step S20233, the vehicle system adjusts the parameters in the second state representation component and the second probability prediction component according to the matching vehicle pattern matching data corresponding to the training representative data, so as to enable the model to more accurately classify the training representative data of the same vehicle pattern matching training data into the same category. This is an iterative process until a satisfactory classification effect is achieved.

[0244] For example, the vehicle system can use the back-propagation algorithm to adjust the parameters. The back-propagation algorithm is based on the principle of gradient descent, which calculates the gradient of the loss function with respect to the parameters, and then updates the parameters in the opposite direction of the gradient to reduce the value of the loss function.

[0245] Assume that the loss function L is used to measure the difference between the model prediction result and the true label, for example, the cross entropy loss function can be used.

[0246] The vehicle system first calculates the gradient of the loss function with respect to the parameters (such as weights and biases) in the second probability prediction component, as well as the gradient of the parameters (such as convolution kernel weights, etc.) in the second state representation component. Then, the parameters are updated according to the gradient descent algorithm. The vehicle system will repeat this process continuously, and each iteration will calculate the gradient and update the parameters based on the new training representative data and the corresponding matching vehicle pattern matching data. As the iterations proceed, the classification accuracy of the model will gradually improve, and eventually the training representative data of the same vehicle pattern matching training data can be accurately classified into the same vehicle pattern matching training data category.

[0247] During the implementation of step S2023, the vehicle system normalizes the third mode classification probability, determines the matching vehicle mode matching data based on the normalized probability and the critical probability, and then debugs the model parameters according to the matching results, continuously optimizing the performance of the model so that it can better adapt to the characteristics of the vehicle mode matching training data, and provide strong support for accurate vehicle state classification and subsequent emergency handling. This iterative optimization process ensures that the model can continuously learn and improve when processing complex vehicle operation mode data, thereby improving the accuracy and reliability of classification.

[0248] Figure 2 A schematic diagram of a hardware entity of a vehicle-mounted system provided in an embodiment of the present application is shown as follows: Figure 2 As shown, the hardware entity of the vehicle-mounted system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and when the processor 1001 executes the program, the steps in the method of any of the above embodiments are implemented.

[0249] The memory 1002 stores computer programs that can be run on the processor. The memory 1002 is configured to store instructions and applications executable by the processor 1001. It can also cache data to be processed or processed by the processor 1001 and various modules in the vehicle-mounted system 1000 (for example, image data, audio data, voice communication data, and video communication data). This can be achieved through flash memory (FLASH) or random access memory (Random Access Memory, RAM).

[0250] When the processor 1001 executes the program, the steps of any of the above-mentioned methods for emergency handling of ski resort transport vehicles based on deep learning are implemented. The processor 1001 generally controls the overall operation of the vehicle-mounted system 1000.

[0251] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A deep learning-based emergency handling method for ski resort transport vehicles, characterized in that: include: Acquire a real-time monitoring data sequence of a ski resort transport vehicle, and determine a target vehicle operation mode corresponding to the real-time monitoring data sequence in each first vehicle operation mode prepared in advance based on a pattern matching neural network and a state characterization vector of the real-time monitoring data sequence; Determine the first vehicle mode matching data belonging to the target vehicle operation mode according to the attribution between the target vehicle operation mode and the first vehicle mode matching data prepared in advance; wherein the vehicle state characteristics of the first vehicle mode matching data belonging to the same first vehicle operation mode are similar; Determining target vehicle pattern matching data corresponding to the real-time monitoring data sequence based on the first vehicle pattern matching data belonging to the target vehicle operation mode; Based on a preset emergency management strategy mapping relationship, the emergency management strategy corresponding to the target vehicle pattern matching data is used as a real-time emergency management strategy.

2. The method according to claim 1, characterized in that The pattern matching neural network based method determines the target vehicle operation mode corresponding to the real-time monitoring data sequence in each first vehicle operation mode prepared in advance according to the state characterization vector of the real-time monitoring data sequence, including: Extracting a state representation vector of the real-time monitoring data sequence based on a first state representation component in the pattern matching neural network; Based on a first probability prediction component in the pattern matching neural network, determining a first mode classification probability of the real-time monitoring data sequence being classified into each of the first vehicle operation modes according to a state characterization vector of the real-time monitoring data sequence; According to the first mode classification probability, a target vehicle operating mode corresponding to the real-time monitoring data sequence is determined in each of the first vehicle operating modes.

3. The method according to claim 2, characterized in that The determining, based on the state characterization vector of the real-time monitoring data sequence, a first mode classification probability of the real-time monitoring data sequence being classified into each of the first vehicle operation modes comprises: Determining a first commonality metric coefficient between a state characterization vector of the real-time monitoring data sequence and a state characterization vector of each of the first vehicle operation modes; the state characterization vector of the first vehicle operation mode is set in the first probability prediction component; determining a first mode classification probability of the real-time monitoring data sequence being classified into each of the first vehicle operation modes according to the first commonality metric coefficient; Determining the target vehicle operation mode corresponding to the real-time monitoring data sequence in each of the first vehicle operation modes according to the first mode classification probability includes: performing normalization processing on the first mode classification probability; Based on the first mode classification probability and the first critical mode classification probability after normalization, a target vehicle operation mode corresponding to the real-time monitoring data sequence is determined in each of the first vehicle operation modes.

4. The method according to claim 1, characterized in that Also includes: Acquire first representative data of the first vehicle pattern matching data; The data capacity of the first representative data is smaller than the first reference data capacity; Determining the attribution between the first vehicle operating mode and the first representative data according to the attribution between the first vehicle operating mode and the first vehicle mode matching data; The first machine learning model is debugged according to the attribution between the first representative data, the first vehicle operation mode and the first representative data, and the first machine learning model obtained by debugging is the pattern matching neural network.

5. The method according to claim 4, characterized in that The debugging of the first machine learning model according to the attribution between the first representative data, the first vehicle operation mode and the first representative data includes: Extracting a state representation vector of the first representative data based on a first state representation component in the first machine learning model; Determining, based on a first probability prediction component in the first machine learning model and according to a state representation vector of the first representative data, a second mode classification probability of the first representative data being classified into each of the first vehicle operation modes; The parameters in the first state characterization component and the first probability prediction component are debugged according to the second mode classification probability until the matching vehicle operation mode corresponding to the first representative data represented by the second mode classification probability conforms to the attribution between the first vehicle operation mode and the first representative data.

6. The method according to claim 5, characterized in that The determining, based on the state characterization vector of the first representative data, a second mode classification probability of the first representative data being classified into each of the first vehicle operation modes comprises: Determining a second commonality metric coefficient between the state characterization vector of the first representative data and the state characterization vector of each of the first vehicle operation modes; the state characterization vector of the first vehicle operation mode is pre-set in the first probability prediction component; the state characterization vector of the first vehicle operation mode is optimized in the debugging link of the neural network; determining a second mode classification probability of the first representative data being classified into each of the first vehicle operation modes according to the second commonality measurement coefficient; The step of debugging the parameter variables in the first state representation component and the first probability prediction component according to the second mode classification probability until the matching vehicle operation mode corresponding to the first representative data represented by the second mode classification probability conforms to the attribution between the first vehicle operation mode and the first representative data includes: performing normalization processing on the second mode classification probability; Determining a matching vehicle operating mode corresponding to the first representative data in each of the first vehicle operating modes based on the second mode classification probability and the second critical mode classification probability after the normalization process; The parameters in the first state characterization component and the first probability prediction component are debugged according to the matching vehicle operation mode corresponding to the first representative data until the matching vehicle operation mode corresponding to the first representative data conforms to the attribution between the first vehicle operation mode and the first representative data.

7. The method according to claim 1, characterized in that Also includes: acquiring second representative data of the first vehicle pattern matching data; The data capacity of the second representative data is greater than the data capacity of the second reference data; Extracting a state representation vector of the second representative data based on a vehicle state feature extraction network; Performing cluster analysis on the first vehicle mode matching data according to the state characterization vector of the second representative data to obtain the attribution between the first vehicle operation mode and the first vehicle mode matching data; Acquire vehicle pattern matching training data and training representative data of the vehicle pattern matching training data; The data capacity of the training representative data is greater than the third reference data capacity; Debugging a second machine learning model according to the training representative data of the vehicle pattern matching training data, wherein the debugged second machine learning model is used to extract a training state representation vector of the training representative data and classify the training representative data of the same vehicle pattern matching training data into the same vehicle pattern matching training data according to the training state representation vector; The vehicle state feature extraction network is generated based on the second machine learning model obtained through debugging.

8. The method according to claim 7, characterized in that The acquiring of vehicle pattern matching training data and training representative data of the vehicle pattern matching training data comprises: acquiring the vehicle pattern matching training data and third representative data of the vehicle pattern matching training data; performing noise enhancement on the third representative data according to the vehicle state characteristics of the third representative data to obtain fourth representative data; generating the training representative data according to the third representative data and the fourth representative data; The debugging of the second machine learning model according to the training representative data of the vehicle pattern matching training data includes: Extracting a training state representation vector of the training representative data based on a second state representation component in the second machine learning model; Based on the second probability prediction component in the second machine learning model, determining a third mode classification probability of the training representative data being classified into each of the vehicle mode matching training data according to the training state representation vector of the training representative data; The parameters in the second state characterization component and the second probability prediction component are debugged according to the third pattern classification probability until the training representative data representing the same vehicle pattern matching training data as the third pattern classification probability is classified into the same vehicle pattern matching training data.

9. The method according to claim 8, characterized in that The determining, based on the training state representation vector of the training representative data, a third mode classification probability of the training representative data being classified into each of the vehicle mode matching training data comprises: Determining a third commonality metric coefficient between the training state representation vector of the training representative data and the state representation vector of each of the vehicle mode matching training data; the state representation vector of the vehicle mode matching training data is pre-set in the second probability prediction component; the state representation vector of the vehicle mode matching training data is optimized in the debugging link of the neural network; Determining a third mode classification probability of the training representative data being classified into each of the vehicle mode matching training data according to the third commonality metric coefficient; The step of debugging the parameter variables in the second state characterization component and the second probability prediction component according to the third pattern classification probability until the training representative data representing the same vehicle pattern matching training data is classified into the same vehicle pattern matching training data by the third pattern classification probability comprises: performing normalization processing on the third mode classification probability; Based on the standardized third mode classification probability and the third critical mode classification probability, determining the matching vehicle mode matching data corresponding to the training representative data in each of the vehicle mode matching training data; The parameters in the second state characterization component and the second probability prediction component are debugged according to the matching vehicle pattern matching data corresponding to the training representative data until the training representative data of the same vehicle pattern matching training data are classified into the same vehicle pattern matching training data.

10. An in-vehicle system, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the program, the steps in the method according to any one of claims 1 to 9 are implemented.

Images