Data analysis method and system applied to vehicle reaction mechanism

Abstract

The invention provides a data analysis method and system applied to a vehicle reaction mechanism, and the method comprises the steps: obtaining to-be-analyzed vehicle sensing data, first vehicle state data and second vehicle state data; each of the first vehicle state data and the second vehicle state data comprises a state label of at least one potential vehicle state and corresponding reference vehicle sensing data, and the second potential vehicle state is obtained by supplementing the first potential vehicle state; analyzing the to-be-analyzed vehicle sensing data according to the first state tag and the second state tag to obtain at least one third potential vehicle state of the to-be-analyzed vehicle sensing data; and according to the matching degree between the to-be-analyzed vehicle sensing data and the reference vehicle sensing data corresponding to the at least one third potential vehicle state, determining a target vehicle state of the to-be-analyzed vehicle sensing data in the at least one third potential vehicle state. According to the method, the iteration cost of the algorithm can be relieved when the vehicle state category is changed.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a data analysis method and system applied to a vehicle reaction mechanism. Background Art

[0002] With the rapid development of the automobile industry and the increasing perfection of intelligent transportation systems, the safety and intelligence level of vehicles have become important indicators for measuring automobile performance. In complex driving environments, vehicles can accurately identify their own status so that they can respond quickly in emergency situations to ensure driving safety. Traditional vehicle status recognition methods mainly rely on preset rules or simple threshold judgments. These methods can achieve certain results when the vehicle status categories are relatively fixed and limited. However, with the continuous advancement of vehicle technology and the increasing complexity of driving environments, vehicle status categories are also increasing and changing. For example, with the development of autonomous driving technology, vehicles may recognize more driving states, such as following vehicles and changing lanes in autonomous driving mode. Traditional rule- and threshold-based methods are difficult to adapt to such changes and cannot accurately identify newly added vehicle status categories.

[0003] Existing vehicle state recognition methods often lack flexibility and scalability when dealing with newly added vehicle state categories. When the vehicle state category changes, existing methods may reconstruct or redesign the entire model to adapt to the new state category. This not only increases the complexity of the algorithm, but may also introduce new errors and uncertainties. Summary of the invention

[0004] The object of the present invention is to provide a data analysis method and system for vehicle reaction mechanism. The present invention is implemented as follows: In a first aspect, the present invention provides a data analysis method applied to a vehicle reaction mechanism, the method comprising: acquiring vehicle sensor data to be analyzed, first vehicle state data and second vehicle state data, the first vehicle state data comprising a first state label of at least one first potential vehicle state and corresponding reference vehicle sensor data, the second vehicle state data comprising a second state label of at least one second potential vehicle state and corresponding reference vehicle sensor data, the at least one second potential vehicle state being obtained by supplementing the at least one first potential vehicle state; analyzing the vehicle sensor data to be analyzed according to the first state label and the second state label respectively to obtain at least one third potential vehicle state of the vehicle sensor data to be analyzed; determining the target vehicle state of the vehicle sensor data to be analyzed in the at least one third potential vehicle state according to the degree of matching between the vehicle sensor data to be analyzed and the reference vehicle sensor data corresponding to the at least one third potential vehicle state.

[0005] On the other hand, the present invention provides a computer system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.

[0006] The data analysis method applied to the vehicle reaction mechanism provided by the present invention adds a first state label and a first reference vehicle sensor data to the first potential vehicle state. Based on this, when the category of the first potential vehicle state changes, the second state label and the second reference vehicle sensor data of the supplemented second potential vehicle state can be set at the same time. Then, the vehicle sensor data to be analyzed are analyzed according to the first state label and the second state label respectively, and at least one third potential vehicle state of the vehicle sensor data to be analyzed is obtained. Afterwards, according to the matching degree between the vehicle sensor data to be analyzed and the reference vehicle sensor data corresponding to at least one third potential vehicle state, the target vehicle state of the vehicle sensor data to be analyzed is determined in at least one third potential vehicle state. Based on this, when the category of the vehicle state changes, only the state label of the changed vehicle state needs to be modified, and the corresponding reference vehicle sensor data is supplemented at the same time, so that it can adapt to the changed vehicle state in real time, avoid repeated training of the existing algorithm, and be conducive to the effective response of the vehicle emergency mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a flow chart of a data analysis method applied to a vehicle reaction mechanism provided by an embodiment of the present invention.

[0008] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The execution subject of the data analysis method applied to the vehicle reaction mechanism in the embodiment of the present invention is a computer system, including but not limited to a server, a personal computer, a notebook computer, a tablet computer, etc. Figure 1 As shown, the method includes: Step S110: Obtain the vehicle sensor data to be analyzed, the first vehicle status data and the second vehicle status data, wherein the first vehicle status data includes a first status label of at least one first potential vehicle status and the corresponding reference vehicle sensor data, and the second vehicle status data includes a second status label of at least one second potential vehicle status and the corresponding reference vehicle sensor data, and the at least one second potential vehicle status is obtained by supplementing the at least one first potential vehicle status.

[0010] In step S110, the computer system first obtains the sensor data of the vehicle to be analyzed. These data usually come from various sensors installed on the vehicle, such as speed sensors, acceleration sensors, steering angle sensors, etc. These sensors can collect various operating parameters of the vehicle in real time, such as speed, acceleration, steering angle, etc., and transmit these parameters to the computer system in the form of data. For example, for a ski vehicle, its sensors may include wheel speed sensors for measuring vehicle speed, gyroscope sensors for detecting the vehicle's tilt angle, etc. The data collected by these sensors are the sensor data of the vehicle to be analyzed, which can fully reflect the current operating status of the vehicle.

[0011] Next, the computer system obtains the first vehicle state data. The first vehicle state data includes a first state label of no less than one first potential vehicle state and the corresponding reference vehicle sensor data. The first potential vehicle state here refers to various candidate states that the vehicle may be in, and the first state label is a label used to describe these vehicle states. For example, for a ski vehicle, its first potential vehicle state may include normal driving state, sharp turn state, braking state, etc. Each state corresponds to a unique first state label, such as "normal driving", "sharp turn", "braking", etc.

[0012] Corresponding to the first state label is the reference vehicle sensor data. The reference vehicle sensor data refers to the data collected by the sensor when the vehicle is in a first potential vehicle state. These data can be used as a reference for determining whether the vehicle is in this state. For example, when the ski vehicle is in a normal driving state, its speed sensor may collect a relatively stable speed value, the acceleration value collected by the acceleration sensor is close to zero, and the steering angle value collected by the steering angle sensor is also relatively stable. These data constitute the reference vehicle sensor data of the normal driving state.

[0013] In addition to the first vehicle state data, the computer system also obtains second vehicle state data. The second vehicle state data is an update of the first vehicle state data, and it includes no less than a second state tag of a second potential vehicle state and corresponding reference vehicle sensor data. The second potential vehicle state is a supplement to the first potential vehicle state, and they may be a newly emerged vehicle state or a refinement of the original vehicle state. For example, during the operation of a ski vehicle, a special driving state may occur, such as a "sharp turn state on an icy road", which may not be included in the first vehicle state data, but is supplemented in the second vehicle state data.

[0014] Corresponding to the second state label is the second reference vehicle sensor data. These data are also the data collected by the sensors when the vehicle is in a second potential vehicle state. Similar to the first reference vehicle sensor data, they can be used as a reference for determining whether the vehicle is in this state. However, the difference is that since the second potential vehicle state may be a newly emerging or refined state, the second reference vehicle sensor data may be significantly different from the first reference vehicle sensor data.

[0015] Step S120: Analyze the vehicle sensor data to be analyzed according to the first state label and the second state label respectively to obtain at least one third potential vehicle state of the vehicle sensor data to be analyzed.

[0016] In step S120, the computer system processes the sensor data of the vehicle to be analyzed for subsequent classification analysis. This processing process includes steps such as data cleaning, feature extraction and embedding mapping. Data cleaning refers to removing noise and outliers in the sensor data to ensure the accuracy and reliability of the data. Feature extraction is to extract features useful for vehicle status judgment from the sensor data, such as speed, acceleration, steering angle, etc. Embedding mapping is to map high-dimensional sensor data to low-dimensional space for more efficient analysis and calculation.

[0017] Suppose a sensor data set of a ski vehicle contains sensor data of multiple dimensions such as speed, acceleration, steering angle, etc. In order to perform embedding mapping, algorithms such as principal component analysis (PCA) or t-SNE can be used. Next, the sensor data of the vehicle to be analyzed is analyzed according to the first state label and the second state label. This analysis process is actually a classification process, the purpose of which is to determine which vehicle state the sensor data of the vehicle to be analyzed belongs to. To achieve this goal, the computer system uses a classification algorithm in machine learning technology.

[0018] In an embodiment of the present invention, the computer system, for example, uses two embedding mapping algorithms to process the first state label and the second state label respectively. The first embedding mapping algorithm is used to embed and map the state labels under multiple operating modes, and convert the state labels into a low-dimensional state representation array. For example, for the state labels of the ski vehicle such as "normal driving", "sharp turn" and "braking", the first embedding mapping algorithm can convert them into a two-dimensional or three-dimensional state representation array. The second embedding mapping algorithm is used to embed and map the state labels under the operating mode corresponding to the sensor data of the vehicle to be analyzed, which can combine the state labels with the sensor data more closely.

[0019] Assume that there are two state labels: "normal driving" and "sharp turn", and their corresponding state representation arrays are [0.5, 0.8] and [0.2, 0.9] respectively. At the same time, there is also a target sensor representation array [0.45, 0.78] after PCA processing, which represents the main information of the vehicle sensor data to be analyzed.

[0020] Next, the match between the target sensor representation array and each state representation array is calculated. This calculation process can use metrics such as Euclidean distance, cosine similarity, or Manhattan distance.

[0021] The vehicle state may not be accurately determined by relying solely on the first state label, because the vehicle state may be affected by many factors. Therefore, the second state label and the corresponding embedding mapping algorithm are also used for further analysis. Assume that the second state label includes more detailed states such as "normal driving on ice and snow roads" and "sharp turn on dry roads", and their corresponding state representation arrays are [0.4, 0.7] and [0.1, 0.8] respectively. Similarly, the cosine similarity between the target sensor representation array and these state representation arrays can be calculated, and the state to which the vehicle sensor data to be analyzed is closer can be determined based on the size of the similarity. Finally, the analysis results based on the first state label and the second state label are merged to obtain no less than one third potential vehicle state of the vehicle sensor data to be analyzed. These third potential vehicle states may be classification results obtained based on different state labels and embedding mapping algorithms, and they together constitute a candidate set of vehicle states. According to the matching degree calculation in the subsequent steps, the target vehicle state that best matches the vehicle sensor data to be analyzed can be selected from the third potential vehicle states.

[0022] Step S130: According to the matching degree between the vehicle sensor data to be analyzed and the reference vehicle sensor data corresponding to the at least one third potential vehicle state, the target vehicle state of the vehicle sensor data to be analyzed is determined in the at least one third potential vehicle state.

[0023] In step S130, the computer system obtains the third potential vehicle state obtained in step S120 and its corresponding reference vehicle sensor data. These third potential vehicle states are candidate vehicle states obtained after classifying and analyzing the sensor data of the vehicle to be analyzed based on the first state label and the second state label, and the reference vehicle sensor data is the sensor data feature corresponding to each candidate vehicle state. For example, for a ski vehicle, its third potential vehicle state may include "normal driving on dry road", "sharp turn on icy road", etc., and the reference vehicle sensor data corresponding to each state may include specific values ​​of features such as speed, acceleration, and steering angle.

[0024] Next, the matching degree of the sensor data of the vehicle to be analyzed and the reference vehicle sensor data corresponding to each third potential vehicle state is calculated. This calculation process is to evaluate the similarity between the sensor data of the vehicle to be analyzed and each candidate vehicle state, so as to determine which state best matches the actual operation of the current vehicle. In the embodiment of the present invention, the computer system uses a target data embedding mapping algorithm to embed and map the reference vehicle sensor data to obtain a first reference representation array corresponding to the third potential vehicle state. The target data embedding mapping algorithm is an algorithm that can map high-dimensional data to a low-dimensional space while retaining the main features of the data. For example, for features such as speed, acceleration, steering angle, etc. in the reference vehicle sensor data, the target data embedding mapping algorithm can convert it into a low-dimensional vector, namely the first reference representation array. This vector can represent the main information of the original reference sensor data and facilitate the subsequent matching degree calculation. After obtaining the first reference representation array, it is compared with the sensor data of the vehicle to be analyzed (the target sensor representation array obtained after appropriate processing) to calculate the matching degree between them. This comparison process can use a variety of measurement methods, such as Euclidean distance, cosine similarity, Manhattan distance, etc.

[0025] The computer system performs the above-mentioned matching degree calculation process for each third potential vehicle state to obtain the matching degree value corresponding to each state. Then, based on the magnitude of these matching degree values, the computer system can determine the target vehicle state for analyzing the vehicle sensor data. Specifically, the computer system will select the third potential vehicle state with the highest matching degree value as the target vehicle state, because this state best matches the actual operation of the current vehicle.

[0026] In order to improve the accuracy of determining the state of the target vehicle, a target joint analysis algorithm can also be used to comprehensively consider multiple factors. The target joint analysis algorithm is a multi-task analysis model that can simultaneously process multiple input data (such as sensor data of the vehicle to be analyzed, the third potential vehicle state, the first reference representation array, etc.) and output a comprehensive judgment result. In this process, the target joint analysis algorithm will use methods such as joint learning or multi-task learning in machine learning technology to combine the learning processes of multiple tasks (such as state recognition, sensor data matching, etc.), thereby improving the overall learning effect and prediction accuracy. For example, assume that the target joint analysis algorithm adopts a model structure based on a neural network. This neural network can accept multiple input data such as sensor data of the vehicle to be analyzed, the third potential vehicle state, the first reference representation array, and output a probability distribution representing the state of the target vehicle through nonlinear transformation of multiple hidden layers. The state of the target vehicle can be determined based on this probability distribution, that is, the state with the highest probability value is selected as the final judgment result.

[0027] In one implementation, the step S120, analyzing the to-be-analyzed vehicle sensor data according to the first state tag and the second state tag respectively to obtain at least one third potential vehicle state of the to-be-analyzed vehicle sensor data, includes: Step S121: embedding and mapping the sensor data of the vehicle to be analyzed to obtain a target sensor representation array; Step S122: embedding and mapping the second state label according to a first embedding and mapping algorithm to obtain a first target state representation array, wherein the first embedding and mapping algorithm is used to embed and map state labels under multiple operating modes; Step S123: embedding and mapping the first state label according to a second embedding and mapping algorithm to obtain a second target state representation array, wherein the second embedding and mapping algorithm is used to embed and map the state label under the operating mode corresponding to the sensor data of the vehicle to be analyzed; Step S124: analyzing the sensor data of the vehicle to be analyzed according to the matching degree between the first target state representation array and the target sensor representation array to obtain a first target analysis result; Step S125: analyzing the sensor data of the vehicle to be analyzed according to the matching degree between the second target state representation array and the target sensor representation array to obtain a second target analysis result; Step S126: Merge the first target analysis result and the second target analysis result to obtain at least one third potential vehicle state of the vehicle sensor data to be analyzed.

[0028] In step S121, the sensor data of the vehicle to be analyzed is processed for subsequent analysis and comparison. Sensor data usually contains information of multiple dimensions, such as speed, acceleration, steering angle, temperature, etc., which may have high dimensions in the original form and are more complicated to process directly. Therefore, the embedding mapping technology is used to map the high-dimensional sensor data to a low-dimensional space while retaining the main features of the data. This process is similar to dimensionality reduction, but the difference is that the embedding mapping not only reduces the dimension of the data, but also retains the relative relationship between the data through a specific mapping function. Embedding mapping can be implemented through a variety of algorithms, such as principal component analysis (PCA), t-SNE, autoencoder, etc.

[0029] In step S122, the computer system processes the second state label so as to match it with the sensor data of the vehicle to be analyzed. The second state label is a label that describes different states that the vehicle may be in, such as "normal driving", "sharp turn", "braking", etc. These labels are discrete, non-directly comparable text or symbols in their original form. In order to match the state label with the sensor data, the computer system uses an embedding mapping algorithm to convert the state label into a low-dimensional vector representation, namely the first target state representation array.

[0030] The first embedding mapping algorithm is an embedding mapping algorithm specially designed for state labels under multiple operating modes. It can map labels under different operating states into a common low-dimensional space, making similar labels closer in space. The implementation of this algorithm can be based on word embedding technology, such as Word2Vec, GloVe, etc., or autoencoder. For vehicle state labels, each label can be regarded as a "vocabulary" and mapped to a low-dimensional vector space through the Word2Vec algorithm. Suppose there are three state labels: "normal driving", "sharp turn", and "brake". After Word2Vec processing, three corresponding vector representations are obtained: [0.5, 0.8], [0.2, 0.9], [-0.3, 0.7]. These vectors are the first target state representation array, which represents the position relationship of different state labels in the low-dimensional space.

[0031] In step S123, the computer system processes the first state label so as to make a more precise match with the sensor data of the vehicle to be analyzed. Unlike the second state label, the first state label is a specific state label for the operating mode corresponding to the sensor data of the vehicle to be analyzed. These labels may be more detailed, such as "normal driving on dry roads", "sharp turns on icy roads", etc. In order to more closely combine these labels with the sensor data, the computer system uses a second embedding mapping algorithm to convert the first state label into a low-dimensional vector representation, namely, the second target state representation array.

[0032] The second embedding mapping algorithm is an embedding mapping algorithm specially designed for analyzing the state labels of the corresponding operating modes of vehicle sensor data. It can map the labels under specific operating states into a low-dimensional space, making the relationship between the labels and the sensor data clearer. The implementation of this algorithm can be based on conditional embedding technology or domain adaptation technology. Taking conditional embedding technology as an example, it is a technology that integrates conditional information (such as operating mode) into the embedding mapping process. For the vehicle state label, the operating mode can be regarded as conditional information, and the first state label is mapped to a low-dimensional space closely related to the sensor data through conditional embedding technology.

[0033] The design goal of the second embedding mapping algorithm is to make the relationship between the state label and the sensor data clearer. The algorithm fully considers the impact of the operating mode on the state label and sensor data, and optimizes the embedding mapping model through training data, so that the representation of the label and sensor data in the low-dimensional space is more consistent. In addition, since the first state label may be more detailed, the second embedding mapping algorithm may require more training data and a more complex model structure to ensure the accuracy of the embedding mapping.

[0034] Through the above three steps, the computer system converts the sensor data of the vehicle to be analyzed, the second state label, and the first state label into low-dimensional vector representations (i.e., the target sensor representation array, the first target state representation array, and the second target state representation array). These vector representations not only retain the main features of the original data, but also enable comparison and matching between different types of data (such as sensor data and state labels). This provides important data support for the subsequent analysis and comparison process and is a key step in achieving accurate judgment of vehicle status.

[0035] In step S124, the computer system calculates the matching degree between the target sensor representation array (i.e., the low-dimensional representation of the sensor data of the vehicle to be analyzed) and the first target state representation array (i.e., the low-dimensional representation of the second state label). The matching degree is an indicator of the similarity between two vectors, which reflects the degree of association between the sensor data of the vehicle to be analyzed and a potential vehicle state. In order to calculate the matching degree, the computer system can use the cosine similarity measurement method.

[0036] For example, suppose there is a sensor data of a ski vehicle to be analyzed, and the target sensor representation array obtained after embedding mapping is [0.5, 0.7]. At the same time, there are two second state labels: "normal driving" and "sharp turn", and the first target state representation arrays obtained after processing by the first embedding mapping algorithm are [0.5, 0.8] and [0.2, 0.9] respectively. By calculating the cosine similarity between the target sensor representation array and the two state representation arrays, the similarity between them can be obtained. The calculation results show that the cosine similarity between the "normal driving" state and the sensor data of the vehicle to be analyzed is 0.98, while the cosine similarity between the "sharp turn" state and the sensor data of the vehicle to be analyzed is 0.85. Since 0.98 is greater than 0.85, the computer system can preliminarily determine that the sensor data of the vehicle to be analyzed is closer to the "normal driving" state. The size of the cosine similarity alone cannot fully determine the target vehicle state of the sensor data of the vehicle to be analyzed, because the vehicle state may be affected by many factors. Therefore, the computer system further analyzes the first target analysis result and combines it with other information (such as the second target analysis result) to make a final judgment.

[0037] Step S125 is similar to step S124, but the object of processing is the first state label. The first state label is a specific state label for the operating mode corresponding to the sensor data of the vehicle to be analyzed. They may be more detailed, such as "normal driving on dry roads", "sharp turns on icy roads", etc. In order to more closely combine these labels with the sensor data, the computer system uses a second embedding mapping algorithm to convert the first state label into a second target state representation array.

[0038] Next, the computer system also uses cosine similarity to calculate the matching degree between the target sensor representation array and the second target state representation array. This calculation process is similar to step S124, but the processed data and labels are different. Assume that there are two first state labels: "normal driving on dry roads" and "sharp turn on ice and snow roads", and the second target state representation arrays obtained after being processed by the second embedding mapping algorithm are [0.6, 0.7] and [0.1, 0.8] respectively. By calculating the cosine similarity between the target sensor representation array and the two state representation arrays, the similarity between them can be obtained. The calculation results show that the cosine similarity between the "normal driving on dry roads" state and the sensor data of the vehicle to be analyzed is 0.95, while the cosine similarity between the "sharp turn on ice and snow roads" state and the sensor data of the vehicle to be analyzed is 0.70. Since 0.95 is greater than 0.70, the computer system can preliminarily judge that the sensor data of the vehicle to be analyzed is closer to the "normal driving on dry roads" state. Similar to step S124, the computer system further analyzes the second target analysis results and makes a final judgment in combination with the first target analysis results. During this process, the computer system may consider multiple factors, such as the priority of different state labels, the confidence of sensor data, etc.

[0039] In step S126, the computer system combines the first target analysis result and the second target analysis result to obtain at least one third potential vehicle state of the vehicle sensor data to be analyzed. These third potential vehicle states are classification results obtained based on different state labels and embedding mapping algorithms, and they together constitute a candidate set of vehicle states.

[0040] The process of merging the first target analysis result and the second target analysis result may involve a variety of strategies, such as weighted average, voting mechanism, etc. In the embodiment of the present invention, it can be assumed that the computer system adopts a simple weighted average strategy. Specifically, for each candidate vehicle state, the computer system calculates the weighted average of its corresponding first target analysis result (i.e., cosine similarity) and second target analysis result (also cosine similarity). The selection of weights may be based on a variety of factors, such as the priority of the state label, the confidence of the sensor data, the statistical information of the historical data, etc.

[0041] Assume that there are two candidate vehicle states: "normal driving" (based on the second state label) and "normal driving on dry roads" (based on the first state label). For the "normal driving" state, its first target analysis result is 0.98 (cosine similarity with the sensor data of the vehicle to be analyzed), and the second target analysis result is 0 (because there is no corresponding first state label); for the "normal driving on dry roads" state, its first target analysis result is 0 (because there is no corresponding second state label), and the second target analysis result is 0.95. Assume that the weights assigned to the first target analysis result and the second target analysis result are 0.6 and 0.4 respectively (this is just an assumed value, and the actual weights may be adjusted according to specific circumstances). By calculating the weighted average, the final score of each candidate vehicle state can be obtained: the score of the "normal driving" state is 0.98*0.6+0*0.4=0.588, and the score of the "normal driving on dry roads" state is 0*0.6+0.95*0.4=0.38. Since 0.588 is greater than 0.38, the computer system ultimately determines that the state of the target vehicle whose vehicle sensor data is to be analyzed is "normal driving".

[0042] In one implementation, the step S130, according to the matching degree between the sensor data of the vehicle to be analyzed and the reference vehicle sensor data corresponding to the at least one third potential vehicle state, determines the target vehicle state of the sensor data of the vehicle to be analyzed in the at least one third potential vehicle state, including: Step S131: embedding and mapping the reference vehicle sensor data corresponding to the at least one third potential vehicle state according to the target data embedding and mapping algorithm to obtain a first reference representation array corresponding to the third potential vehicle state; Step S132: passing the to-be-analyzed vehicle sensor data, the at least one third potential vehicle state, and the first reference representation array into a target joint analysis algorithm, and determining a first state recognition support corresponding to the first reference representation array, wherein the first state recognition support is used to indicate a matching degree between the to-be-analyzed vehicle sensor data and the first reference representation array; Step S133: Determine the target vehicle state from the at least one third potential vehicle state according to the first state recognition support.

[0043] In step S131, the computer system processes the reference vehicle sensor data corresponding to the third potential vehicle state for subsequent analysis and comparison. The reference vehicle sensor data is the data collected by the sensor describing the vehicle in different states, and they constitute an important basis for judging the vehicle state. However, the original form of the reference vehicle sensor data may have a high dimension and is more complicated to process directly. Therefore, the computer system uses a target data embedding mapping algorithm to map the high-dimensional reference vehicle sensor data to a low-dimensional space while retaining the main features of the data.

[0044] The target data embedding mapping algorithm maps high-dimensional data to a low-dimensional space while retaining the relative relationship between the data. It is optimized for specific tasks (such as vehicle state recognition) and can better capture key information in the data. For example, for a ski vehicle, its reference vehicle sensor data may include data in multiple dimensions such as speed, acceleration, steering angle, temperature, etc. The target data embedding mapping algorithm can map high-dimensional sensor data to a low-dimensional vector space, namely the first reference representation array, by learning the relationship between these data and the vehicle state.

[0045] Assume that there are two third potential vehicle states: "normal driving" and "sharp turn", and their corresponding reference vehicle sensor data are [v1, a1, θ1, t1] and [v2, a2, θ2, t2], where v represents speed, a represents acceleration, θ represents steering angle, and t represents temperature. After being processed by the target data embedding mapping algorithm, two corresponding first reference representation arrays are obtained: [v1', a1', θ1'] and [v2', a2', θ2']. These vectors not only retain the main features of the original reference sensor data, but also make the data in different states more separable in low-dimensional space.

[0046] In step S132, the sensor data of the vehicle to be analyzed, the third potential vehicle state and the first reference representation array are used as inputs and passed into the target joint analysis algorithm for analysis. The target joint analysis algorithm is a multi-task analysis model that can process multiple input data at the same time and output a comprehensive judgment result. This process is similar to multi-task learning or joint learning in machine learning, which aims to improve the overall learning effect and prediction accuracy by learning multiple related tasks at the same time.

[0047] The target joint analysis algorithm can be implemented based on a variety of machine learning models such as neural networks, support vector machines, and decision trees. In the embodiment of the present invention, it is assumed that the target joint analysis algorithm adopts a model structure based on a neural network. This neural network can accept multiple input data (such as sensor data of the vehicle to be analyzed, the third potential vehicle state, the first reference representation array, etc.), and output a probability distribution representing the target vehicle state through nonlinear transformation of multiple hidden layers.

[0048] The hypothetical neural network model includes an input layer, two hidden layers, and an output layer. The input layer accepts the sensor data of the vehicle to be analyzed, the third potential vehicle state, and the first reference representation array as input; the hidden layer transforms the data and extracts features through nonlinear activation functions (such as ReLU, Sigmoid, etc.); the output layer outputs a probability distribution, indicating the probability that the sensor data of the vehicle to be analyzed belongs to each third potential vehicle state. For example, for the two states of "normal driving" and "sharp turn", the output layer may output a probability distribution [0.8, 0.2], indicating that the sensor data of the vehicle to be analyzed has an 80% probability of being in the "normal driving" state and a 20% probability of being in the "sharp turn" state. This probability distribution is the first state recognition support, which indicates the matching degree of the sensor data of the vehicle to be analyzed with the reference vehicle sensor data corresponding to each third potential vehicle state. By comparing the probability values ​​of different states, it can be determined which state best matches the actual operation of the current vehicle.

[0049] In step S133, the computer system determines the target vehicle state of the vehicle sensor data to be analyzed according to the first state recognition support. Specifically, the computer system selects the third potential vehicle state with the highest probability value as the target vehicle state because this state best matches the actual operation of the current vehicle.

[0050] Assume that there are two third potential vehicle states: "normal driving" and "sharp turn", and their corresponding first state recognition supports are 0.8 and 0.2 respectively. Since 0.8 is greater than 0.2, the computer system can determine that the target vehicle state of the vehicle sensor data to be analyzed is "normal driving". In practical applications, there may be situations where the probability values ​​of multiple states are similar. In this case, a more complex decision strategy may be used to determine the target vehicle state. For example, factors such as the priority between different states, the statistical information of historical data, and the confidence of sensor data can be considered to comprehensively determine which state best fits the current situation.

[0051] In one implementation, the method provided by the present application further includes: Step S140: obtaining the confirmation result of the target vehicle status; Step S150: When the confirmation result indicates that the to-be-analyzed vehicle sensor data matches the target vehicle state, the to-be-analyzed vehicle sensor data is added to the corresponding reference vehicle sensor data of the at least one second potential vehicle state according to the target vehicle state.

[0052] In step S140, the computer system obtains the confirmation result of the target vehicle state. The confirmation result is a verification of the accuracy of the target vehicle state judgment, which may come from a variety of channels, such as manual inspection, auxiliary verification of other sensors, comparative analysis of historical data, etc. The purpose of obtaining the confirmation result is to ensure the accuracy of the target vehicle state and avoid misjudgment or missed judgment that causes the vehicle emergency mechanism to be triggered or not triggered. In order to achieve this goal, a feasible method is to use multi-sensor fusion technology. Multi-sensor fusion is a method of fusing data from different sensors to improve data accuracy and reliability. For example, in a ski car scene, in addition to sensor data such as speed, acceleration, and steering angle, other sensor data such as GPS data and gyroscope data can also be introduced. Another method is to use a rule-based reasoning system. A rule-based reasoning system is a method of reasoning and judging input data according to preset rules. The computer system can predefine a series of rules, such as "if the speed is greater than X and the acceleration is greater than Y, it is judged to be a sharp turn state." When the target vehicle state judgment result meets these rules, it can be considered that the judgment is accurate; if it does not meet them, the judgment result may be further verified or adjusted. Another method is to use the verification mechanism of the machine learning model. If the judgment of the target vehicle state is based on a machine learning model (such as a neural network, support vector machine, etc.), then the accuracy of the model can be verified by methods such as cross-validation and holdout. For example, the historical data can be divided into a training set and a validation set, the training set is used to train the model, and the validation set is used to evaluate the accuracy of the model. If the model performs well on the validation set, then it can be considered that the judgment of the target vehicle state is accurate.

[0053] In step S150, when the confirmation result indicates that the sensor data of the vehicle to be analyzed matches the target vehicle state, the computer system adds the sensor data of the vehicle to be analyzed to the reference vehicle sensor data of the corresponding second potential vehicle state. This process is actually a process of data updating and model optimization, which helps to improve the accuracy and adaptability of the system's judgment of future vehicle states.

[0054] Specifically, when the confirmation result indicates that the target vehicle state judgment is accurate, the sensor data of the vehicle to be analyzed and its corresponding target vehicle state information are added to the reference vehicle sensor data of the second potential vehicle state. The advantage of this is that as the data continues to accumulate, the reference vehicle sensor data will become richer and more accurate, providing a more reliable basis for subsequent vehicle state judgments.

[0055] For example, suppose there is a second potential vehicle state of a ski vehicle, "sharp turn on icy road", and the corresponding reference vehicle sensor data originally only contains a small amount of historical data. As time goes by, when the computer system continuously receives new sensor data of the vehicle to be analyzed and verifies that these data match the target vehicle state (such as "sharp turn on icy road") through confirmation results, these new data will be added to the reference vehicle sensor data of the "sharp turn on icy road" state. As the data continues to accumulate, the reference vehicle sensor data will become more complete and accurate, which will help improve the system's ability to recognize the "sharp turn on icy road" state.

[0056] In addition, adding the sensor data of the vehicle to be analyzed to the sensor data of the reference vehicle has another important role, which is to optimize the machine learning model. If the judgment of the state of the target vehicle is based on the machine learning model, then as new data is added, the model can continuously optimize its parameters and structure through retraining or online learning, thereby improving the accuracy and robustness of the judgment of the future vehicle state. For example, suppose a neural network-based model is used to judge the vehicle state. When the new sensor data of the vehicle to be analyzed is added to the sensor data of the reference vehicle, these data can be input into the neural network as new training samples for training. The weights and biases of the neural network are adjusted through the back propagation algorithm, so that the output of the model is closer to the true label (that is, the state of the target vehicle). In this way, as time goes by and data continues to accumulate, the performance of the neural network will continue to improve, and the ability to judge the vehicle state will continue to increase.

[0057] Through the above two steps, not only the target vehicle status is confirmed and verified, but also the system's adaptability and accuracy are improved through data update and model optimization. This process constitutes a closed-loop feedback mechanism, which helps to ensure the effective response of the vehicle emergency mechanism and the safe operation of the vehicle.

[0058] In one implementation, the process of determining the second vehicle status data includes: Step S101: obtaining vehicle state data to be supplemented, first example sensor data, and a first example vehicle state of the first example sensor data, wherein the vehicle state data to be supplemented includes at least one original state label of a target supplemented vehicle state and corresponding original reference sensor data, and the target supplemented vehicle state is obtained by supplementing the first potential vehicle state; Step S102: analyzing the first example sensor data according to the first state label and the original state label respectively to obtain at least one example potential vehicle state of the first example sensor data; Step S103: determining an inferred vehicle state of the first example sensor data from the at least one example potential vehicle state according to a matching degree between the first example sensor data and the original reference sensor data corresponding to the at least one example potential vehicle state; Step S104: When the inferred vehicle state matches the first example vehicle state, the vehicle state data to be supplemented is determined as the second vehicle state data.

[0059] In step S101, the computer system obtains the vehicle state data to be supplemented, the first example sensor data, and the first example vehicle state of the first example sensor data. The vehicle state data to be supplemented is the data supplemented by the computer system to improve the vehicle state library, which contains no less than one original state label of the target supplement vehicle state and the corresponding original reference sensor data. These target supplement vehicle states are obtained by supplementing the first potential vehicle state, and they may represent the newly appeared state of the vehicle, the refined state, or the state under different operating environments.

[0060] For example, suppose there is a first potential vehicle state library for a ski vehicle, which includes basic states such as "normal driving" and "sharp turn". However, in actual applications, the states of the ski vehicle under different road conditions may also be considered, such as "normal driving on icy and snowy roads" and "sharp turn on dry roads". These states are the target supplementary vehicle states, which supplement the first potential vehicle state. In order to obtain data on these states, the computer system collects the vehicle state data to be supplemented, including state labels (such as "normal driving on icy and snowy roads") and corresponding reference sensor data (such as speed, acceleration, steering angle, etc. when driving normally on icy and snowy roads).

[0061] At the same time, the computer system also obtains the first example sensor data and the first example vehicle state of the first example sensor data. The first example sensor data is data used to train and verify the machine learning model, which contains the sensor data of the vehicle in different states. The first example vehicle state is the vehicle state label corresponding to these sensor data. For example, a series of sensor data of skis under different road conditions can be collected and associated with corresponding vehicle state labels (such as "normal driving", "sharp turn", etc.) to form the first example sensor data and the first example vehicle state.

[0062] In step S102, the computer system analyzes the first example sensor data to obtain the corresponding example potential vehicle state. This process is similar to the classification analysis of the vehicle sensor data to be analyzed, but the difference is that the object of analysis here is the first example sensor data, and the classification is based on the first state label and the original state label.

[0063] Specifically, the computer system preprocesses the first example sensor data, such as data cleaning, feature extraction, and embedding mapping. Then, the first state label and the original state label are used to analyze the first example sensor data. This process can be implemented based on classification algorithms in machine learning, such as support vector machines, decision trees, neural networks, etc. For example, a neural network-based classification model can be used, which accepts the first example sensor data as input, and outputs a probability distribution representing the vehicle state through nonlinear transformation of multiple hidden layers. For each first example sensor data, the first state label and the original state label can be used as inputs of the model to obtain two probability distributions. Then, the state with the highest probability value can be selected as the example potential vehicle state.

[0064] Assume a first example sensor data, including sensor data of a ski vehicle under certain road conditions. The sensor data is analyzed using the first state label (such as "normal driving") and the original state label (such as "normal driving on icy and snowy roads"). The results show that in the probability distribution obtained by analyzing using the first state label, the probability value of the "normal driving" state is the highest; while in the probability distribution obtained by analyzing using the original state label, the probability value of the "normal driving on icy and snowy roads" state is the highest. Therefore, it can be considered that the example potential vehicle state corresponding to this first example sensor data is "normal driving on icy and snowy roads".

[0065] In step S103, the inferred vehicle state of the first example sensor data is determined according to the matching degree of the first example sensor data and the original reference sensor data corresponding to the example potential vehicle state. This process is similar to the matching degree calculation and analysis process in step S130, but the difference is that the objects analyzed here are the first example sensor data and the example potential vehicle state, and the matching degree calculation is based on the original reference sensor data.

[0066] Specifically, the computer system performs embedding mapping processing on the original reference sensor data corresponding to each example potential vehicle state to obtain a first reference representation array. Then, it compares the first example sensor data (the target sensor representation array obtained after appropriate preprocessing and embedding mapping) with each first reference representation array to calculate the matching degree between them. This comparison process can use a variety of measurement methods, such as Euclidean distance, cosine similarity, Manhattan distance, etc. In the embodiment of the present invention, cosine similarity is used as the matching degree calculation indicator.

[0067] By calculating the cosine similarity between the target sensor characterization array and each first reference characterization array, the degree of similarity between them can be obtained. Assume that the calculation results show that the cosine similarity between the "normal driving on icy and snowy roads" state and the first example sensor data is 0.95, while the cosine similarity between the "normal driving" state and the first example sensor data is 0.80. Since 0.95 is greater than 0.80, the computer system can determine that the inferred vehicle state of the first example sensor data is "normal driving on icy and snowy roads". When determining the inferred vehicle state, multiple factors may be considered, such as the priority between different states, the confidence of the sensor data, the statistical information of historical data, etc. These factors can be comprehensively considered by setting different weights or adopting decision models such as decision trees and Bayesian networks.

[0068] In step S104, it is determined whether the vehicle state data to be supplemented can be used as the second vehicle state data according to the matching of the inferred vehicle state and the first example vehicle state. If the inferred vehicle state matches the first example vehicle state, it means that the vehicle state data to be supplemented is accurate and reliable, and can be added to the second vehicle state data; if it does not match, the vehicle state data to be supplemented may be further corrected or the data may be collected again.

[0069] The determination of the matching degree can be implemented based on a variety of methods, such as threshold determination, probability comparison, etc. In the embodiment of the present invention, a matching degree threshold (such as 0.9) can be set, and only when the matching degree between the inferred vehicle state and the first example vehicle state is greater than this threshold, they are considered to be matched.

[0070] Assume that there is a first example sensing data, and its first example vehicle state is "normal driving on icy and snowy roads". After the analysis of step S103, it is obtained that the inferred vehicle state is also "normal driving on icy and snowy roads". At this time, compare the matching degree of the inferred vehicle state with that of the first example vehicle state. If the matching degree is greater than a preset threshold (such as 0.9), the computer system can deem that the vehicle state data to be supplemented is accurate and reliable, and can determine it as part of the second vehicle state data. In this way, a new vehicle state (such as "normal driving on icy and snowy roads") is supplemented to the second vehicle state data, providing more comprehensive and accurate data support for subsequent vehicle state judgments.

[0071] In one implementation, the process of determining the second vehicle status data further includes: Step S105: when the inferred vehicle state does not match the first example vehicle state, obtaining correction data for the at least one target supplementary vehicle state, the correction data including an iterative state label and / or iterative sensor data; Step S106: supplementing the original state label corresponding to the at least one target supplementary vehicle state according to the iterative state label; Step S107: supplementing the original reference sensor data corresponding to the at least one target supplement vehicle state according to the iterative sensor data.

[0072] In step S105, the computer system identifies the mismatch between the inferred vehicle state and the first example vehicle state, and obtains corresponding correction data to correct the vehicle state data to be supplemented. The mismatch between the inferred vehicle state and the first example vehicle state may be caused by a variety of reasons, such as data noise, model error, unclear state definition, etc. In order to solve this problem, the computer system obtains correction data, including iterative state labels and / or iterative sensor data.

[0073] Iterative state labels are further refinements or corrections to the target supplementary vehicle state, which may be derived from new vehicle state definitions, more accurate sensor data, or more advanced model algorithms. For example, suppose there was originally a target supplementary vehicle state "normal driving on icy and snowy roads", but in actual applications, it is found that the definition of this state is too broad and cannot accurately reflect the driving state of the vehicle on icy and snowy roads. At this time, more refined iterative state labels can be introduced, such as "normal driving on light icy and snowy roads", "normal driving on heavy icy and snowy roads", etc., to more accurately describe the driving state of the vehicle under different icy and snowy road conditions. Iterative sensor data are supplements or corrections to the original reference sensor data, which may be derived from more accurate sensor measurements, more comprehensive data collection, or more advanced data processing techniques. For example, suppose there was originally an original reference sensor data [v1, a1, θ1], which represents the speed, acceleration, and steering angle of the vehicle under certain icy and snowy road conditions. However, in actual applications, it is found that these data may be affected by noise or outliers, resulting in inaccurate inference of vehicle states. At this point, more accurate iterative sensor data [v2, a2, θ2] can be introduced to more accurately reflect the driving state of the vehicle on icy and snowy roads. In order to obtain correction data, a variety of methods may be used, such as manual collection, sensor calibration, model prediction, etc. For example, for the acquisition of iterative state labels, it can be defined based on expert knowledge or domain rules; for the acquisition of iterative sensor data, the computer system can use more accurate sensors or more advanced data processing technology for collection and processing.

[0074] In step S106, the computer system supplements the original state label according to the iterative state label to improve the definition of the target supplementary vehicle state. Specifically, the computer system matches and integrates the iterative state label with the original state label to form a more comprehensive and accurate set of vehicle state labels. For example, suppose there was originally a target supplementary vehicle state "normal driving on icy and snowy roads", and its corresponding original state label was "normal driving". However, in step S105, the iterative state labels "normal driving on light icy and snowy roads" and "normal driving on heavy icy and snowy roads" were obtained. In order to improve the definition of the vehicle state, the computer system matches and integrates these two iterative state labels with the original state label. Specifically, it can subdivide the "normal driving" state into two sub-states: "normal driving on light icy and snowy roads" and "normal driving on heavy icy and snowy roads", and assign corresponding iterative state labels to each sub-state.

[0075] In the implementation process, the computer system may use rule-based methods or machine learning methods to supplement the status labels. The rule-based method refers to matching and integrating the status labels according to preset rules or logic; the machine learning method refers to automatically dividing and integrating the status labels using machine learning algorithms (such as clustering algorithms, classification algorithms, etc.). For example, a clustering algorithm can be used to perform cluster analysis on iterative status labels and original status labels, classify similar status labels into one category, and assign a representative status label to each category.

[0076] In step S107, the computer system supplements the original reference sensor data according to the iterative sensor data to improve the accuracy and reliability of the sensor data. Specifically, the computer system matches and integrates the iterative sensor data with the original reference sensor data to form a more complete and accurate sensor data set.

[0077] For example, suppose there is a target supplementary vehicle state "normal driving on icy and snowy roads", and its corresponding original reference sensor data is [v1, a1, θ1]. However, in step S105, iterative sensor data [v2, a2, θ2] are obtained, which more accurately reflect the driving state of the vehicle on icy and snowy roads. In order to improve the accuracy of the sensor data, the computer system matches and integrates the two sensor data sets. Specifically, it can combine [v1, a1, θ1] and [v2, a2, θ2] by weighted averaging or selecting the optimal value to obtain a more accurate reference sensor data set.

[0078] In the implementation process, a variety of methods can be used to supplement and integrate sensor data. One method is the weighted average method, which assigns different weights to different sensor data according to the reliability and accuracy of the data, and then performs weighted average calculation. For example, if the accuracy of the iterative sensor data is higher, it can be assigned a larger weight; conversely, if the accuracy of the original reference sensor data is higher, it can be assigned a smaller weight. Another method is the optimal value selection method, which selects the optimal sensor data as the final result according to certain criteria (such as minimum value, maximum value, median, etc.).

[0079] In one implementation, the step S102 of analyzing the first example sensor data according to the first state label and the original state label to obtain at least one example potential vehicle state of the first example sensor data includes: Step S1021: performing embedding mapping on the first example sensor data to obtain a first example sensor representation array; Step S1022: embedding and mapping the original state label according to a first embedding and mapping algorithm to obtain a first example state representation array, wherein the first embedding and mapping algorithm is used to embed and map the state labels under multiple operation modes; Step S1023: embedding and mapping the first state label according to a second embedding and mapping algorithm to obtain a second example state representation array, wherein the second embedding and mapping algorithm is used to embed and map the state label under the operating mode corresponding to the sensor data of the vehicle to be analyzed; Step S1024: analyzing the first example sensor data according to the matching degree between the first example state representation array and the first example sensor representation array to obtain a first example analysis result; Step S1025: analyzing the first example sensor data according to the matching degree between the second example state representation array and the first example sensor representation array to obtain a second example analysis result; Step S1026: Merge the first example analysis result and the second example analysis result to obtain no less than one example potential vehicle state of the first example sensor data.

[0080] In step S1021, the first example sensor data is processed for subsequent analysis and comparison. The first example sensor data describes the data collected by the sensor of the vehicle in different states, which contains important basis for judging the vehicle state. The original sensor data may have a high dimension, and there may be redundancy or correlation between data of different dimensions, which is more complicated to process directly. Therefore, the embedding mapping technology is used to map the high-dimensional sensor data to a low-dimensional space, while retaining the main features of the data, so as to facilitate subsequent analysis and calculation.

[0081] In step S1022, the original state label is processed to match with the first example sensor data. The original state label is a label describing different states that the vehicle may be in, such as "normal driving", "sharp turn", etc. These labels are discrete, non-directly comparable text or symbols in their original form. In order to match the state label with the sensor data, the computer system uses a first embedding mapping algorithm to convert the state label into a low-dimensional vector representation, namely the first example state representation array.

[0082] In step S1023, the computer system processes the first state label so as to make a more precise match with the first example sensor data. Different from the original state label, the first state label is a specific state label for the operating mode corresponding to the sensor data of the vehicle to be analyzed. These labels may be more detailed, such as "normal driving on dry roads", "sharp turns on icy roads", etc. In order to more closely combine these labels with the sensor data, the computer system uses a second embedding mapping algorithm to convert the first state label into a low-dimensional vector representation, namely the second example state representation array.

[0083] The second embedding mapping algorithm is an embedding mapping algorithm specially designed for analyzing the state labels of the corresponding operating modes of vehicle sensor data. It can map the labels under specific operating states into a low-dimensional space, making the relationship between the labels and the sensor data clearer. The implementation of this algorithm can be based on conditional embedding technology or domain adaptation technology. Taking conditional embedding technology as an example, it is a technology that integrates conditional information (such as operating mode) into the embedding mapping process. For the vehicle state label, the operating mode can be regarded as conditional information, and the first state label is mapped to a low-dimensional space closely related to the sensor data through conditional embedding technology.

[0084] Compared with the first embedding mapping algorithm, the second embedding mapping algorithm pays more attention to the relationship between the state label and the sensor data under a specific operating state. It introduces conditional information (such as operating mode) to make the embedding mapping result of the state label closer to the actual sensor data, thereby improving the accuracy of classification and matching.

[0085] Through the above three steps, the computer system converts the first example sensor data, the original state label and the first state label into low-dimensional vector representations (i.e., the first example sensor representation array, the first example state representation array and the second example state representation array). These vector representations not only reduce the dimension of the data, but also retain the relative relationship between the data, so that different types of data (such as sensor data and state labels) can be compared and matched. This provides important data support for subsequent classification and matching operations, and is a key step in achieving accurate judgment of vehicle status.

[0086] In step S1024, the matching degree between the first example state representation array (i.e., the low-dimensional representation of the original state label) and the first example sensor representation array (i.e., the low-dimensional representation of the first example sensor data) is calculated, and the first example sensor data is analyzed based on the matching degree. The matching degree is an indicator of the similarity between two vectors, which reflects the degree of association between the state label and the sensor data. In order to calculate the matching degree, the computer system can use a cosine similarity measurement method.

[0087] Step S1025 is similar to step S1024, but the object of processing is the first state label instead of the original state label. The first state label is a specific state label for the operating mode corresponding to the sensor data of the vehicle to be analyzed, which may be more detailed, such as "normal driving on dry roads", "sharp turn on icy and snowy roads", etc. In order to combine these labels with the sensor data more closely, the computer system adopts a second embedding mapping algorithm to convert the first state label into a second example state representation array. Next, the cosine similarity is also used to calculate the matching degree between the first example sensor representation array and the second example state representation array, and the first example sensor data is analyzed based on this matching degree. Assume that there are two first state labels: "normal driving on dry roads" and "sharp turn on icy and snowy roads", and the second example state representation arrays obtained after they are processed by the second embedding mapping algorithm are [0.6, 0.7] and [0.1, 0.8] respectively. By calculating the cosine similarity between the first example sensor representation array and the two state representation arrays, the similarity between them can be obtained. The calculation results show that the cosine similarity between the state of "driving normally on a dry road" and the first example sensor data is 0.95, while the cosine similarity between the state of "sharp turn on icy and snowy road" and the first example sensor data is 0.70. Since 0.95 is greater than 0.70, the computer system can preliminarily determine that the first example sensor data is closer to the state of "driving normally on a dry road".

[0088] Similar to step S1024, the second example analysis result is also a probability distribution, score or other form of representation, which is used to indicate the degree of match between each first state label and the first example sensor data. This analysis result will provide important data support for subsequent steps, helping the computer system to more accurately determine the example potential vehicle state of the first example sensor data.

[0089] In step S1026, the computer system combines the first example analysis result and the second example analysis result to obtain at least one example potential vehicle state of the first example sensor data. These example potential vehicle states are classification results obtained based on different state labels and embedding mapping algorithms, and together constitute a set of vehicle states that the first example sensor data may be in.

[0090] The process of merging the first example analysis result and the second example analysis result may involve a variety of strategies, such as weighted average, voting mechanism, probability fusion, etc. In an embodiment of the present invention, it can be assumed that the computer system adopts a simple weighted average strategy. Specifically, for each candidate vehicle state (i.e., the classification result corresponding to each state label), the computer system calculates the weighted average of the corresponding first example analysis result (such as cosine similarity or probability value) and the second example analysis result (also cosine similarity or probability value). The selection of weights may be based on a variety of factors, such as the priority of the state label, the confidence of the sensor data, the statistical information of the historical data, etc.

[0091] In one implementation, the step S103, according to the matching degree between the first example sensor data and the original reference sensor data corresponding to the at least one example potential vehicle state, determines the inferred vehicle state of the first example sensor data in the at least one example potential vehicle state, including: Step S1031: embedding and mapping the original reference sensor data corresponding to the at least one example potential vehicle state according to the target data embedding and mapping algorithm to obtain a second reference representation array; Step S1032: passing the first example sensor data, the at least one example potential vehicle state, and the second reference representation array into a target joint analysis algorithm, and determining a second state recognition support corresponding to the second reference representation array, where the second state recognition support is used to indicate a matching degree between the first example sensor data and the second reference representation array; Step S1033: Determine the inferred vehicle state from the at least one example potential vehicle state according to the second state recognition support.

[0092] In step S1031, the computer system processes the original reference sensor data corresponding to the example potential vehicle state for subsequent analysis and comparison. The original reference sensor data is the data collected by the sensor describing the vehicle in different states, which constitutes an important basis for judging the vehicle state. However, the original form of the reference sensor data may have a higher dimension, and there may be redundancy or correlation between the data of different dimensions, which is more complicated to process directly. Therefore, the computer system adopts the target data embedding mapping algorithm to map the high-dimensional reference sensor data to a low-dimensional space, while retaining the main features of the data, for subsequent analysis and calculation.

[0093] The target data embedding mapping algorithm is an algorithm that can map high-dimensional data to a low-dimensional space while preserving the relative relationship between the data. It is similar to dimensionality reduction algorithms such as principal component analysis (PCA) or t-SNE, but the difference is that the target data embedding mapping algorithm is usually optimized for specific tasks (such as vehicle state recognition) and can better capture key information in the data. In an embodiment of the present invention, the role of the target data embedding mapping algorithm is to map the original reference sensor data corresponding to each example potential vehicle state into a low-dimensional vector space, that is, the second reference representation array.

[0094] In step S1032, the first example sensor data, the example potential vehicle state and the second reference representation array are used as inputs and passed into the target joint analysis algorithm for analysis. The target joint analysis algorithm is a multi-task analysis model that can process multiple input data at the same time and output a comprehensive judgment result. This process is similar to multi-task learning or joint learning in machine learning, which aims to improve the overall learning effect and prediction accuracy by learning multiple related tasks at the same time.

[0095] The implementation of the target joint analysis algorithm can be based on a variety of machine learning models such as neural networks, support vector machines, and decision trees. In the embodiment of the present invention, it is assumed that the target joint analysis algorithm adopts a model structure based on a neural network. This neural network can accept multiple input data (such as first example sensor data, example potential vehicle state, second reference representation array, etc.), and through nonlinear transformation of multiple hidden layers, output a probability distribution representing the degree of matching of each example potential vehicle state with the first example sensor data, that is, the second state recognition support.

[0096] Specifically, the assumed neural network model includes an input layer, two hidden layers, and an output layer. The input layer accepts the first example sensor data (the first example sensor representation array obtained after embedding mapping), the example potential vehicle state (represented in one-hot encoding or other forms), and the second reference representation array as input; the hidden layer transforms the data and extracts features through nonlinear activation functions (such as ReLU, Sigmoid, etc.); the output layer outputs a probability distribution, indicating the degree of match between each example potential vehicle state and the first example sensor data. For example, for the two states of "normal driving" and "sharp turn", the output layer may output a probability distribution [0.8, 0.2], indicating that the first example sensor data matches the "normal driving" state by 80% and the "sharp turn" state by 20%.

[0097] This probability distribution is the second state recognition support, which indicates the matching degree of the first example sensor data with the original reference sensor data corresponding to each example potential vehicle state. By comparing the probability values ​​of different states, the computer system can determine which state best matches the actual operation of the current first example sensor data.

[0098] In step S1033, the computer system determines the inferred vehicle state of the first example sensor data according to the second state recognition support. Specifically, the computer system selects the example potential vehicle state with the highest probability value as the inferred vehicle state, because this state best matches the actual operation of the current first example sensor data.

[0099] Assume that there are two potential vehicle states: "normal driving" and "sharp turn", and their corresponding second state recognition support is 0.8 and 0.2 respectively. Since 0.8 is greater than 0.2, the computer system can determine that the inferred vehicle state of the first example sensor data is "normal driving".

[0100] In actual applications, there may be situations where the probability values ​​of multiple states are similar. In this case, a more complex decision-making strategy can be used to determine the inferred vehicle state. For example, factors such as the priority between different states, the statistical information of historical data, and the confidence of sensor data can be considered to comprehensively determine which state best fits the current situation.

[0101] Through the above three steps, the first example sensor data and the original reference sensor data corresponding to the example potential vehicle state are matched and analyzed, and the inference vehicle state is determined based on the recognition support of the second state. This process not only involves complex embedding mapping technology and multi-task analysis models, but also involves decision-making strategies, integrated learning methods, and uncertainty estimation mechanisms. These technical details work together to achieve the ultimate goal and provide strong support for the accurate judgment of vehicle status.

[0102] In one implementation, the process of determining the second embedding mapping algorithm includes: Step S201: obtaining second example sensor data and at least one category status label in the operating mode corresponding to the sensor data of the vehicle to be analyzed; Step S202: performing embedding mapping on the second example sensor data to obtain a second example sensor representation array; Step S203: embedding and mapping the at least one category state label according to a preset embedding and mapping algorithm to obtain a third example state representation array; Step S204: embedding and mapping the at least one category state label according to the first embedding and mapping algorithm to obtain a fourth example state representation array; Step S205: According to the matching error between the second example sensor characterization array and the third example state characterization array, the matching error between the second example sensor characterization array and the fourth example state characterization array, and the inference error between the third example state characterization array and the fourth example state characterization array, the algorithm weights of the preset embedding mapping algorithm are optimized to obtain the second embedding mapping algorithm.

[0103] In step S201, the computer system collects the second example sensor data and the corresponding category status label in the operating mode corresponding to the sensor data of the vehicle to be analyzed. The second example sensor data is the sensor collection data describing the vehicle in different operating states, which contains an important basis for judging the vehicle state. These data may come from various sensors on the vehicle, such as speed sensors, acceleration sensors, steering angle sensors, etc., which can reflect the operating state of the vehicle in real time.

[0104] Class state labels are labels used to describe the different states that the vehicle may be in, such as "normal driving", "sharp turn", "braking", etc. These labels are discrete, not directly comparable text or symbols, but they are closely related to the sensor data. In order to match the state labels with the sensor data, they are mapped into a common low-dimensional space, which is the task of the embedding mapping algorithm.

[0105] In step S202, the computer system performs embedding mapping processing on the second example sensor data to convert it into a low-dimensional vector representation, namely, the second example sensor representation array. Embedding mapping is a data dimensionality reduction and feature extraction technology that can map high-dimensional data to a low-dimensional space while retaining the relative relationship between the data. In an embodiment of the present invention, algorithms such as principal component analysis (PCA), t-SNE, and autoencoder can be used to implement embedding mapping.

[0106] In step S203, the computer system performs embedding mapping processing on the category state label and converts it into a low-dimensional vector representation, namely, the third example state representation array. Unlike the second example sensor data, the category state label is a discrete text or symbol and cannot be directly numerically calculated. Therefore, a special embedding mapping algorithm, such as Word2Vec, GloVe, etc., is used to map the state label into a low-dimensional vector space.

[0107] In step S204, embedding mapping is performed on the category state label according to the first embedding mapping algorithm to obtain a fourth example state representation array. The first embedding mapping algorithm is an embedding mapping algorithm specially designed for state labels under multiple operating modes, which can map labels under different operating states into a common low-dimensional space, so that similar labels are closer in space.

[0108] Different from the preset embedding mapping algorithm, the first embedding mapping algorithm may use a more complex model structure and optimization method to better adapt to the state label of the operating mode corresponding to the vehicle sensor data to be analyzed. For example, a conditional embedding model based on a neural network can be used as the first embedding mapping algorithm. This model first accepts the operating mode (such as "dry road" or "ice and snow road") as input, and then embeds and maps the state label according to this condition.

[0109] In step S205, the computer system optimizes the algorithm weights of the preset embedding mapping algorithm to obtain a second embedding mapping algorithm. The optimization goal is to minimize the matching error between the second example sensor representation array and the third example state representation array, the matching error between the second example sensor representation array and the fourth example state representation array, and the reasoning error between the third example state representation array and the fourth example state representation array.

[0110] The matching error is an indicator to measure the similarity between the sensor data representation array and the state representation array. It can be obtained by calculating the Euclidean distance, cosine similarity or other measurement methods between two vectors. In the embodiment of the present invention, cosine similarity can be used as a measurement method for matching error.

[0111] The reasoning error is an indicator to measure the consistency between two state representation arrays. It can be obtained by calculating a certain distance or difference between two vectors. In the embodiment of the present invention, the Euclidean distance can be used as a measurement method of the reasoning error.

[0112] In order to optimize the algorithm weights of the preset embedding mapping algorithm, optimization algorithms such as gradient descent, stochastic gradient descent, and Adam can be used. These optimization algorithms iteratively adjust the algorithm weights to minimize the sum of the matching error and the inference error. Specifically, for each training sample (i.e., a pair of second example sensor data and category state labels), the matching error and the inference error are calculated, and then the algorithm weights of the preset embedding mapping algorithm are adjusted according to the gradient of the error. This process is repeated until a predetermined convergence condition is reached (such as the error is less than a certain threshold or the number of iterations reaches an upper limit).

[0113] Assume that there is a second example sensor representation array [v', a'] and a category state label "normal driving". Use the preset embedding mapping algorithm and the first embedding mapping algorithm to embed and map them respectively, and obtain the third example state representation array [0.5, 0.8] and the fourth example state representation array [0.6, 0.7]. Then, calculate the matching error (using cosine similarity) between these two state representation arrays and the second example sensor representation array and the reasoning error (using Euclidean distance) between them. Next, use the gradient descent algorithm to optimize the algorithm weights of the preset embedding mapping algorithm to minimize the sum of the two errors. After multiple iterations, the optimized algorithm weights are obtained, that is, the algorithm weights of the second embedding mapping algorithm.

[0114] In one implementation, the step S205 optimizes the algorithm weight of the preset embedding mapping algorithm according to the matching error between the second example sensor characterization array and the third example state characterization array, the matching error between the second example sensor characterization array and the fourth example state characterization array, and the reasoning error between the third example state characterization array and the fourth example state characterization array to obtain the second embedding mapping algorithm, including: Step S2051: Acquire a second example vehicle state corresponding to the second example sensor data; Step S2052: determining a first matching error according to the second example sensor characterization array, the third example state characterization array and the second example vehicle state; Step S2053: determining a second matching error according to the second example sensor characterization array, the fourth example state characterization array and the second example vehicle state; Step S2054: determining a reasoning error according to the third example state representation array and the fourth example state representation array; Step S2055: Optimizing the algorithm weight of the preset embedding mapping algorithm according to the first matching error, the second matching error and the reasoning error to obtain the second embedding mapping algorithm.

[0115] In step S2051, the computer system obtains the second example vehicle state corresponding to the second example sensor data. The second example sensor data is the sensor collection data describing the vehicle in different operating states, and the second example vehicle state is the actual operating state of the vehicle corresponding to these data. The purpose of obtaining the second example vehicle state is to provide a true label for the subsequent matching error and reasoning error calculation to ensure the accuracy and effectiveness of the optimization process.

[0116] In practical applications, the second example vehicle state may be obtained through manual labeling, simulation experiments, or historical data records. For example, for the data analysis of a ski car, a series of sensor data under different road conditions can be collected, and the vehicle state corresponding to each piece of data can be determined through experimental records or video analysis, such as "normal driving", "sharp turn", "braking", etc. These state labels are the second example vehicle states, which will be used for subsequent comparison with the output of the embedding mapping algorithm.

[0117] In step S2052, the computer system calculates the matching error between the second example sensor representation array and the third example state representation array, and compares the error with the second example vehicle state to determine a first matching error. The matching error is an indicator of the similarity between two vectors, which reflects the degree of association between the sensor data representation array and the state representation array. In the embodiment of the present invention, cosine similarity can be used as a measurement method for matching error.

[0118] Step S2053 is similar to step S2052, but the object of processing is the fourth example state representation array instead of the third example state representation array. The fourth example state representation array is a low-dimensional representation of the category state label after being processed by the first embedding mapping algorithm, which may more accurately reflect the state information of the operating mode corresponding to the sensor data of the vehicle to be analyzed.

[0119] Likewise, the cosine similarity is used to calculate the matching error between the second example sensor representation array and the fourth example state representation array, and this error is compared with the second example vehicle state to determine a second matching error.

[0120] In step S2054, the computer system calculates the reasoning error between the third example state representation array and the fourth example state representation array. The reasoning error is an indicator to measure the consistency between two state representation arrays, which reflects the difference between the processing results of different embedding mapping algorithms for the same category state label. In the embodiment of the present invention, the Euclidean distance can be used as a measurement method for the reasoning error.

[0121] In step S2055, the computer system optimizes the algorithm weights of the preset embedding mapping algorithm according to the first matching error, the second matching error and the inference error to obtain a second embedding mapping algorithm. The optimization goal is to minimize the sum of the three errors so that the second embedding mapping algorithm can more accurately map the category state label to the low-dimensional space and effectively match it with the sensor data.

[0122] To achieve this goal, the computer system can use optimization algorithms such as gradient descent, stochastic gradient descent, and Adam. These optimization algorithms iteratively adjust the algorithm weights to minimize the sum of errors. Specifically, for each training sample (i.e., a pair of second example sensor data and category state labels), the first matching error, the second matching error, and the reasoning error are calculated, and then the algorithm weight of the preset embedding mapping algorithm is adjusted according to the gradient of the error. During the optimization process, the gradient of the error is continuously calculated, and the algorithm weight is updated according to the gradient direction. The calculation of the gradient can be implemented by the back propagation algorithm. The back propagation algorithm is an algorithm for training a neural network, which calculates the partial derivative of the error with respect to the weight (i.e., the gradient), and then updates the weight in the opposite direction of the gradient to minimize the error. In an embodiment of the present invention, the preset embedding mapping algorithm can be regarded as a neural network model, and the partial derivative of the error with respect to the algorithm weight is calculated by the back propagation algorithm, and the weight is updated. By continuously iteratively updating the weights, the sum of the first matching error, the second matching error, and the reasoning error can be gradually reduced, thereby obtaining the optimized algorithm weight, i.e., the algorithm weight of the second embedding mapping algorithm. This weight will enable the second embedding mapping algorithm to more accurately map the category state labels into a low-dimensional space and effectively match them with the sensor data.

[0123] In one implementation, the determination process of the target joint analysis algorithm includes: Step S301: Acquire sample reference sensor data; Step S302: splitting the example reference sensor data into data items to obtain a first data item splitting result; Step S303: shielding the target data item in the first data item splitting result to obtain the second data item splitting result; Step S304: embedding and mapping the first data item splitting result according to the target data embedding and mapping algorithm to obtain a third reference representation array; Step S305: passing the second data item splitting result and the third reference representation array into a preset joint analysis algorithm to obtain the shielded reasoning result; Step S306: According to the inference result and the first data item splitting result, the algorithm weight of the preset joint analysis algorithm is optimized to obtain the target joint analysis algorithm.

[0124] In step S301, the computer system collects sample reference sensor data. Sample reference sensor data is data collected by sensors describing the vehicle in different states, which constitutes an important basis for judging the vehicle state. These data may come from the vehicle's historical records, simulation experiments or actual tests, and contain rich vehicle state information.

[0125] In order to obtain sample reference sensor data, the computer system may collect data from multiple channels and perform data cleaning and preprocessing. Data cleaning refers to removing noise, outliers, or missing values ​​in sensor data to ensure the accuracy and completeness of the data. Preprocessing may include operations such as data normalization, standardization, or feature extraction to facilitate subsequent analysis and calculation.

[0126] For example, for the data analysis of ski vehicles, a series of sensor data under different road conditions can be collected, including data in multiple dimensions such as speed, acceleration, steering angle, temperature, etc. These data will be used as example reference sensor data to train and optimize the target joint analysis algorithm.

[0127] In step S302, the computer system performs data item splitting on the example reference sensor data to obtain a first data item splitting result. Data item splitting refers to extracting each dimension (or data item) in the sensor data separately to form a feature vector or array. This process facilitates subsequent in-depth analysis and comparison of the sensor data.

[0128] Assume that there is an example reference sensor data, which contains data of four dimensions: velocity v, acceleration a, steering angle θ, and temperature T. In order to split the data items, the computer system can extract the data of these four dimensions separately to form a four-dimensional feature vector [v, a, θ, T]. This vector is the result of the first data item splitting, which will be used for subsequent analysis and calculation.

[0129] In step S303, the computer system masks the target data items in the first data item splitting result to obtain the second data item splitting result. The target data items refer to the data items for inference or prediction. By masking these data items, it is possible to simulate whether the algorithm can accurately perform inference or prediction in the absence of such information.

[0130] For example, suppose you want to test whether the target joint analysis algorithm can accurately determine the vehicle state in the absence of steering angle information. Then, in step S303, the computer system will mask the steering angle data item in the first data item splitting result to obtain a three-dimensional second data item splitting result [v, a, T]. This result will be used for subsequent analysis and calculation to evaluate the performance of the algorithm in the absence of specific information.

[0131] In step S304, the computer system performs embedding mapping processing on the split result of the first data item, and converts it into a low-dimensional vector representation, that is, the third reference representation array. Embedding mapping is a technology for data dimensionality reduction and feature extraction, which can map high-dimensional data to a low-dimensional space while retaining the relative relationship between the data. In the embodiment of the present invention, the target data embedding mapping algorithm is used to achieve this conversion.

[0132] The target data embedding mapping algorithm is an algorithm specifically used to map sensor data into a low-dimensional space. It may be implemented based on algorithms such as principal component analysis (PCA), t-SNE, and autoencoder. These algorithms project the data into a low-dimensional space by finding the principal components or potential features of the data, thereby simplifying subsequent analysis and calculations.

[0133] Assume that the PCA algorithm is used as the target data embedding mapping algorithm. For the first data item split result [v, a,θ, T], the PCA algorithm calculates its covariance matrix and finds the eigenvalues ​​and eigenvectors of the covariance matrix. Then, it selects several eigenvectors with the largest eigenvalues ​​as the principal component directions, projects the data onto these directions, and obtains a low-dimensional third reference representation array. For example, if the first two principal component directions are selected, the obtained third reference representation array may be a two-dimensional vector [v', a'], which represents the main information of the original sensor data.

[0134] In step S305, the computer system uses the splitting result of the second data item and the third reference representation array as input and passes them into the preset joint analysis algorithm for analysis. The preset joint analysis algorithm is a multi-task analysis model that can process multiple input data at the same time and output a comprehensive judgment result. In the embodiment of the present invention, it is assumed that the preset joint analysis algorithm adopts a model structure based on a neural network.

[0135] The neural network model can accept multiple input data (such as the split result of the second data item and the third reference representation array), and output a probability distribution representing the vehicle state or reasoning result through nonlinear transformation of multiple hidden layers. In step S305, the split result of the second data item with the target data item shielded and the third reference representation array are used as input and passed into the neural network model for reasoning. The model will calculate a probability distribution based on the input data, indicating the possibility of different vehicle states. This probability distribution is the shielded reasoning result, which reflects the algorithm's reasoning ability when specific information is missing.

[0136] For example, suppose there is a preset joint analysis algorithm that uses a three-layer neural network model. This model accepts the second data item split result [v, a, T] (shielded steering angle θ) and the third reference representation array [v', a'] as input, and outputs a four-dimensional probability distribution [p1, p2, p3, p4] through the nonlinear transformation of the hidden layer, which respectively represents the possibility of "normal driving", "sharp turn", "braking" and "other states". This probability distribution is the shielded reasoning result.

[0137] In step S306, the computer system optimizes the algorithm weight of the preset joint analysis algorithm according to the inference result and the first data item splitting result to obtain the target joint analysis algorithm. The optimization goal is to minimize the difference between the inference result and the true label (i.e., the vehicle state corresponding to the first data item splitting result) so that the algorithm can more accurately judge the vehicle state.

[0138] To achieve this goal, the computer system can use optimization algorithms such as gradient descent, stochastic gradient descent, and Adam. These optimization algorithms iteratively adjust the algorithm weights to minimize the inference error (i.e., the difference between the inference result and the true label). During the optimization process, the computer system continuously calculates the gradient of the inference error and updates the algorithm weights according to the gradient direction.

[0139] By continuously iterating and updating the weights, the computer system can gradually reduce the reasoning error, thereby obtaining the optimized algorithm weight, that is, the algorithm weight of the target joint analysis algorithm. This weight will enable the algorithm to more accurately judge the vehicle status and give reasonable reasoning results even in the absence of specific information.

[0140] In one implementation, the determination process of the target joint analysis algorithm further includes: Step S307: Acquire third example sensor data and corresponding third example vehicle state; Step S308: obtaining a third example sensor representation array of the third example sensor data; Step S309: using the third example sensor representation array as a reference sensor representation array for the third example vehicle state; Step S3010: passing the third example sensor data, the third example vehicle state and the third example sensor representation array into a preset joint analysis algorithm to obtain a third state recognition support of the third example vehicle state; Step S3011: Optimizing the algorithm weight of the preset joint analysis algorithm according to the third state recognition support to obtain the target joint analysis algorithm.

[0141] In step S307, the computer system collects third example sensor data and corresponding third example vehicle states. The third example sensor data is data collected by sensors describing the vehicle in different states, and these data are different from the previous example reference sensor data, for example, from different vehicles, different operating environments, or different time periods. The third example vehicle state is the actual operating state of the vehicle corresponding to these data, which provides real labels for subsequent algorithm optimization.

[0142] In step S308, the computer system performs embedding mapping processing on the third example sensor data to convert it into a low-dimensional vector representation, namely, the third example sensor representation array. This step is similar to step S304, but the object of processing is the third example sensor data instead of the example reference sensor data.

[0143] In step S309, the computer system uses the third example sensor representation array as a reference sensor representation array for the third example vehicle state. The reference sensor representation array is a low-dimensional vector representation used to describe the vehicle state, which can be used as a benchmark for subsequent algorithm optimization. By associating the third example sensor representation array with the third example vehicle state, a reference sensor representation array can be constructed for each vehicle state for subsequent algorithm training and testing.

[0144] In step S3010, the computer system uses the third example sensor data, the third example vehicle state and the third example sensor representation array as inputs and passes them into a preset joint analysis algorithm for analysis. The preset joint analysis algorithm is a multi-task analysis model that can process multiple input data at the same time and output a comprehensive judgment result. In the embodiment of the present invention, it is assumed that the preset joint analysis algorithm adopts a model structure based on a neural network.

[0145] The neural network model can accept multiple input data (such as the third example sensor data, the third example vehicle state and the third example sensor representation array), and output a probability distribution representing the vehicle state or the inference result through nonlinear transformation of multiple hidden layers. In step S3010, the third example sensor data, the third example vehicle state and the third example sensor representation array are used as input and passed into the neural network model for inference. The model will calculate a probability distribution based on the input data, indicating the possibility of different vehicle states. This probability distribution is the third state recognition support, which reflects the algorithm's recognition ability of the third example vehicle state.

[0146] In step S3011, the computer system optimizes the algorithm weight of the preset joint analysis algorithm according to the third state recognition support to obtain the target joint analysis algorithm. The optimization goal is to minimize the difference between the third state recognition support and the true label (i.e., the third example vehicle state) so that the algorithm can more accurately judge the vehicle state. In order to achieve this goal, optimization algorithms such as gradient descent, random gradient descent, and Adam can be used. These optimization algorithms iteratively adjust the algorithm weights to minimize the inference error (i.e., the difference between the third state recognition support and the true label). During the optimization process, the gradient of the inference error is continuously calculated, and the algorithm weight is updated according to the gradient direction. During the optimization process, the inference error is calculated according to the difference between the third state recognition support and the true label, and the partial derivative (i.e., gradient) of the error to the weight is calculated by the back propagation algorithm. Then, the weight is updated according to the Adam algorithm to minimize the inference error. By continuously iteratively updating the weights, the accuracy of the algorithm can be gradually improved, thereby obtaining the optimized target joint analysis algorithm. During the optimization process, the computer system may adjust the hyperparameters of the algorithm (such as learning rate, momentum coefficient, batch size, etc.) to obtain better optimization effects.

[0147] An embodiment of the present invention provides a computer system, such as Figure 2 As shown, the computer system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiments of the present invention.

[0148] An embodiment of the present invention provides a computer system, and the computer system in the embodiment of the present invention includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above-mentioned data analysis method applied to the vehicle reaction mechanism is implemented.

Claims

1. A data analysis method applied to vehicle reaction mechanisms, characterized in that: The method comprises: Acquire sensor data of a vehicle to be analyzed, first vehicle state data, and second vehicle state data, wherein the first vehicle state data includes a first state tag of at least one first potential vehicle state and corresponding reference vehicle sensor data, and the second vehicle state data includes a second state tag of at least one second potential vehicle state and corresponding reference vehicle sensor data, wherein the at least one second potential vehicle state is obtained by supplementing the at least one first potential vehicle state; Analyze the to-be-analyzed vehicle sensor data according to the first state tag and the second state tag respectively, and obtain at least one third potential vehicle state of the to-be-analyzed vehicle sensor data; According to the matching degree between the vehicle sensor data to be analyzed and the reference vehicle sensor data corresponding to the at least one third potential vehicle state, the target vehicle state of the vehicle sensor data to be analyzed is determined in the at least one third potential vehicle state.

2. The data analysis method for vehicle reaction mechanism according to claim 1, characterized in that: The analyzing the to-be-analyzed vehicle sensor data according to the first state tag and the second state tag respectively to obtain at least one third potential vehicle state of the to-be-analyzed vehicle sensor data includes: Embedding and mapping the sensor data of the vehicle to be analyzed to obtain a target sensor representation array; Embedding and mapping the second state label according to a first embedding and mapping algorithm to obtain a first target state representation array, wherein the first embedding and mapping algorithm is used to embed and map state labels under multiple operating modes; Embedding and mapping the first state label according to a second embedding and mapping algorithm to obtain a second target state representation array, wherein the second embedding and mapping algorithm is used to embed and map the state label corresponding to the operating mode of the sensor data of the vehicle to be analyzed; Analyzing the sensor data of the vehicle to be analyzed according to the matching degree between the first target state representation array and the target sensor representation array to obtain a first target analysis result; Analyzing the sensor data of the vehicle to be analyzed according to the matching degree between the second target state representation array and the target sensor representation array to obtain a second target analysis result; Combining the first target analysis result and the second target analysis result to obtain at least one third potential vehicle state of the vehicle sensor data to be analyzed; The step of determining a target vehicle state for obtaining the vehicle sensor data to be analyzed from the at least one third potential vehicle state based on a degree of matching between the vehicle sensor data to be analyzed and the reference vehicle sensor data corresponding to the at least one third potential vehicle state comprises: Embed and map the reference vehicle sensor data corresponding to the at least one third potential vehicle state according to a target data embedding and mapping algorithm to obtain a first reference representation array corresponding to the third potential vehicle state; The sensor data of the vehicle to be analyzed, the at least one third potential vehicle state and the first reference representation array are passed into the target joint analysis algorithm to determine the first state recognition support corresponding to the first reference representation array, where the first state recognition support is used to indicate the matching degree between the sensor data of the vehicle to be analyzed and the first reference representation array; The target vehicle state is determined from the at least one third potential vehicle state according to the first state recognition support.

3. The data analysis method applied to vehicle reaction mechanism according to claim 1, characterized in that: The method further comprises: Obtaining a confirmation result of the target vehicle status; When the confirmation result indicates that the to-be-analyzed vehicle sensor data matches the target vehicle state, the to-be-analyzed vehicle sensor data is added to the corresponding reference vehicle sensor data of the at least one second potential vehicle state according to the target vehicle state.

4. The data analysis method for vehicle reaction mechanism according to claim 1, characterized in that: The process of determining the second vehicle status data includes: Acquire vehicle state data to be supplemented, first example sensor data, and a first example vehicle state of the first example sensor data, wherein the vehicle state data to be supplemented includes at least one original state label of a target supplemented vehicle state and corresponding original reference sensor data, and the target supplemented vehicle state is obtained by supplementing the first potential vehicle state; Analyze the first example sensor data according to the first state label and the original state label respectively to obtain at least one example potential vehicle state of the first example sensor data; determining, based on a degree of matching between the first example sensor data and the original reference sensor data corresponding to the at least one example potential vehicle state, an inferred vehicle state of the first example sensor data in the at least one example potential vehicle state; When the inferred vehicle state matches the first example vehicle state, the vehicle state data to be supplemented is determined as the second vehicle state data.

5. The data analysis method applied to vehicle reaction mechanism according to claim 4, characterized in that: The process of determining the second vehicle status data further includes: When the inferred vehicle state does not match the first example vehicle state, obtaining correction data for the at least one target supplementary vehicle state, the correction data comprising an iterative state label and / or iterative sensor data; Supplementing the original state label corresponding to the at least one target supplementary vehicle state according to the iterative state label; The original reference sensor data corresponding to the at least one target supplementary vehicle state is supplemented according to the iterative sensor data.

6. The data analysis method for vehicle reaction mechanism according to claim 4, characterized in that: The step of analyzing the first example sensor data according to the first state label and the original state label to obtain at least one example potential vehicle state of the first example sensor data includes: Performing embedding mapping on the first example sensor data to obtain a first example sensor representation array; Embedding and mapping the original state label according to a first embedding and mapping algorithm to obtain a first example state representation array, wherein the first embedding and mapping algorithm is used to embed and map the state labels under multiple operating modes; Embedding and mapping the first state label according to a second embedding and mapping algorithm to obtain a second example state representation array, wherein the second embedding and mapping algorithm is used to embed and map the state label corresponding to the operating mode of the sensor data of the vehicle to be analyzed; Analyzing the first example sensor data according to the matching degree between the first example state representation array and the first example sensor representation array to obtain a first example analysis result; Analyze the first example sensor data according to the matching degree between the second example state characterization array and the first example sensor characterization array to obtain a second example analysis result; Combining the first example analysis result with the second example analysis result to obtain at least one example potential vehicle state of the first example sensor data; The determining, based on the matching degree between the first example sensor data and the original reference sensor data corresponding to the at least one example potential vehicle state, the inferred vehicle state of the first example sensor data in the at least one example potential vehicle state comprises: Embedding and mapping the original reference sensor data corresponding to the at least one example potential vehicle state according to a target data embedding and mapping algorithm to obtain a second reference representation array; Passing the first example sensor data, the at least one example potential vehicle state, and the second reference representation array into a target joint analysis algorithm, determining a second state recognition support corresponding to the second reference representation array, wherein the second state recognition support is used to indicate a matching degree between the first example sensor data and the second reference representation array; The inferred vehicle state is determined from the at least one example potential vehicle state according to the second state recognition support.

7. The data analysis method applied to vehicle reaction mechanism according to claim 2 or 6, characterized in that: The determination process of the second embedding mapping algorithm includes: Acquire second example sensor data and at least one category status label in the operating mode corresponding to the sensor data of the vehicle to be analyzed; Performing embedding mapping on the second example sensor data to obtain a second example sensor representation array; Embedding and mapping the at least one category state label according to a preset embedding and mapping algorithm to obtain a third example state representation array; Perform embedding mapping on the at least one category state label according to the first embedding mapping algorithm to obtain a fourth example state representation array; According to the matching error between the second example sensor characterization array and the third example state characterization array, the matching error between the second example sensor characterization array and the fourth example state characterization array, and the inference error between the third example state characterization array and the fourth example state characterization array, the algorithm weights of the preset embedding mapping algorithm are optimized to obtain the second embedding mapping algorithm.

8. The data analysis method applied to vehicle reaction mechanism according to claim 7, characterized in that: The method of optimizing the algorithm weight of the preset embedding mapping algorithm according to the matching error between the second example sensor characterization array and the third example state characterization array, the matching error between the second example sensor characterization array and the fourth example state characterization array, and the reasoning error between the third example state characterization array and the fourth example state characterization array to obtain the second embedding mapping algorithm includes: Acquire a second example vehicle state corresponding to the second example sensor data; determining a first matching error based on the second example sensor representation array, the third example state representation array, and the second example vehicle state; determining a second matching error based on the second example sensor representation array, the fourth example state representation array, and the second example vehicle state; Determining a reasoning error based on the third example state representation array and the fourth example state representation array; According to the first matching error, the second matching error and the reasoning error, the algorithm weight of the preset embedding mapping algorithm is optimized to obtain the second embedding mapping algorithm.

9. The data analysis method applied to vehicle reaction mechanism according to claim 2 or 6, characterized in that: The determination process of the target joint analysis algorithm includes: Obtaining example reference sensor data; Performing data item splitting on the example reference sensor data to obtain a first data item splitting result; Shielding the target data item in the first data item splitting result to obtain a second data item splitting result; Perform embedding mapping on the first data item splitting result according to the target data embedding mapping algorithm to obtain a third reference representation array; Passing the splitting result of the second data item and the third reference representation array into a preset joint analysis algorithm to obtain the shielded reasoning result; According to the inference result and the first data item splitting result, the algorithm weight of the preset joint analysis algorithm is optimized to obtain the target joint analysis algorithm; Acquire third example sensor data and corresponding third example vehicle state; acquiring a third example sensory representation array of the third example sensory data; using the third example sensor representation array as a reference sensor representation array for the third example vehicle state; Passing the third example sensor data, the third example vehicle state and the third example sensor representation array into a preset joint analysis algorithm to obtain a third state recognition support of the third example vehicle state; The algorithm weight of the preset joint analysis algorithm is optimized according to the third state recognition support to obtain the target joint analysis algorithm.

10. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 9 is implemented.

Images