CN116295409A - Route processing method, route processing device, computer readable medium and electronic equipment - Google Patents

Abstract

The present application can be applied to technical fields such as maps and traffic, and specifically provides a route processing method, device, computer-readable medium, and electronic equipment. The method includes: generating a link expression corresponding to a planned route sample; extracting multiple initial embedding vectors corresponding to each element in the link expression through a pre-training model, and generating an input feature vector corresponding to each element; Multi-layer forward propagation processing of the input feature vector, and model training processing based on the planning route samples, training tasks corresponding to the pre-training model, and intermediate result vectors corresponding to each element, so that the specified route can be corrected based on the converged pre-training module to process. The technical solution of the present application can solve the problem of huge workload caused by requiring a large amount of labeled data for model training, and can improve the accuracy of business processing.

Description

Route processing method, device, computer readable medium and electronic device

technical field

The present application relates to the technical field of computer and communication, and in particular, relates to a route processing method, device, computer readable medium and electronic equipment.

Background technique

Path planning is to find a route from the start point to the end point in a specified area. For example, in the field of map navigation, after the map user selects the start point and end point on the map, he can plan a route from the start point to the end point. route. Another example is in the field of robot control, path planning is to plan a collision-free safe route from the starting point to the end point of the robot. After the planned route is obtained, related business processing such as rationality evaluation can usually be performed on the planned route, so as to improve the path planning strategy based on the evaluation result. In related technologies, developers usually need to model based on industry experience and implement model training based on feature engineering. This solution not only has a huge workload, but also has low accuracy of business processing results.

Contents of the invention

Embodiments of the present application provide a route processing method, device, computer-readable medium, and electronic equipment, which can solve the problem of a huge workload caused by requiring a large amount of labeled data for model training, and can improve the efficiency of business processing. accuracy.

Other features and advantages of the present application will become apparent from the following detailed description, or in part, be learned by practice of the present application.

According to an aspect of an embodiment of the present application, a route processing method is provided, including: generating a link corresponding to the planned route sample according to a plurality of links included in the planned route sample and attribute information of each link expression; a plurality of initial embedding vectors corresponding to each element in the link expression is extracted through a pre-training model, and an input feature vector corresponding to each element is generated according to the plurality of initial embedding feature vectors; according to the The input feature vector corresponding to each element is subjected to multi-layer forward propagation processing, and based on the planned route sample, the training task corresponding to the pre-training model, and the multi-layer forward transfer processing obtained corresponding to each element The intermediate result vector is processed for model training, so as to process the specified route based on the converged pre-training module.

According to an aspect of the embodiments of the present application, there is provided a route processing device, including: a first generating unit configured to generate the The link expression corresponding to the planned route sample; the feature extraction unit is configured to extract a plurality of initial embedding vectors corresponding to each element in the link expression through a pre-training model; the second generation unit is configured to initial embedding feature vectors to generate input feature vectors corresponding to each element; the model training unit is configured to perform multi-layer forward propagation processing according to the input feature vectors corresponding to each element, and based on the planned route samples, the described The training task corresponding to the pre-training model and the intermediate result vector corresponding to each element obtained by the multi-layer forward transmission process are subjected to model training processing, so as to process the specified route based on the converged pre-training module. .

In some embodiments of the present application, based on the foregoing solution, the first generation unit is configured to: generate the user information corresponding to each link according to at least one attribute feature included in the attribute information of each link. According to the order of the plurality of links in the planned route sample, the elements corresponding to the plurality of links are combined to generate the link expression corresponding to the planned route sample .

In some embodiments of the present application, based on the foregoing solution, the first generation unit is configured to: according to the position of the specified link among the multiple links in the planned route sample, in the link expression Add the position flag bit corresponding to the specified link; wherein the position flag bit includes at least one of the following: the first flag bit added before the element corresponding to the starting link in the planned route sample, The second flag bit is added between the elements corresponding to the two specified links, and the third flag bit is added after the element corresponding to the termination link in the planned route sample.

In some embodiments of the present application, based on the foregoing solution, the specified two links include: a link located at a road intersection in the planned route sample, or any two adjacent links among the multiple links adjacent link.

In some embodiments of the present application, based on the foregoing solution, the feature extraction unit is configured to: extract word embedding vectors, position embedding vectors, and route interval embedding vectors corresponding to each element in the link expression through a pre-training model ; Wherein, the position embedding vector is used to represent the specific position of the element in the link expression, and the route interval embedding vector is used to represent the position of the element in the link expression interval.

In some embodiments of the present application, based on the aforementioned solution, the second generation unit is configured to: perform a merge process on the multiple initial embedded feature vectors corresponding to the respective elements based on the set merge processing method, and combine the The result of is used as the input feature vector corresponding to each element.

In some embodiments of the present application, based on the foregoing solution, the model training unit is configured to: perform replacement operations on the links other than the initial link in the planned route samples with set probabilities, Obtain the processed planned route sample; the replacement operation includes at least one of the following: replace with a set flag bit, replace with a random link, and keep it unchanged; take the intermediate result vector corresponding to each element as input, and use the The link represented by the set flag bit is restored to the actual link in the planned route sample as the optimization target corresponding to the first training task, and the model training process is performed.

In some embodiments of the present application, based on the foregoing solution, performing replacement operations on links other than the initial link in the planned route samples according to set probabilities, including: For links other than the initial link, determine whether to perform the replacement operation through the first probability; for the link that needs to be replaced, replace it with the second probability of setting the flag bit, and replace it with a random chain The third probability of the way and the fourth probability that remains unchanged are replaced respectively.

In some embodiments of the present application, based on the aforementioned solution, the model training unit is configured to: replace the second half of the links in the planned route samples with random links by a set probability; wherein, the The planned route sample is divided into the first half link and the second half link at the two specified links; the first flag bit added before the element corresponding to the starting link in the link expression corresponds to The intermediate result vector of is used as input, and the random link is restored to the second half of the links in the planned route sample as the optimization target corresponding to the second training task, and the model training process is performed.

In some embodiments of the present application, based on the aforementioned solution, the route processing device further includes: a processing unit configured to extract the target embedding vector corresponding to each element through the converged pre-training model; The target embedding vector of the target is generated to generate the route embedding vector corresponding to the planned route sample; the route embedding vector and the target input information used to adjust the pre-training model are used as input, and the business is set as the optimization target, Adjusting the model parameters of the pre-trained model.

In some embodiments of the present application, based on the foregoing solution, the processing unit is configured to: calculate the target embedding vector corresponding to the starting link in the planned route sample to obtain the The mean value of the target embedding vector; according to the multiple target embedding vectors respectively corresponding to the multiple links included in the planned route sample, calculate the target embedding vector mean value corresponding to the multiple links; A plurality of target embedding vectors corresponding to the link is calculated to obtain the mean value of the target embedding vector corresponding to the terminating link; according to the mean value of the target embedding vector corresponding to the starting link, the target embedding vector corresponding to the multiple links The mean value and the mean value of the target embedding vector corresponding to the terminated link generate a route embedding vector corresponding to the planned route sample.

In some embodiments of the present application, based on the foregoing solution, the target input information used to adjust the pre-training model includes at least one of the following: route features of the planned route samples obtained through feature engineering; The route features of the planned route samples are the prediction results obtained by performing the prediction processing of the setting business through the machine learning model.

In some embodiments of the present application, based on the foregoing solution, the route processing device further includes: a first prediction unit; the first generation unit is further configured to: plan the links included in the route according to the target to be processed, And the attribute information of each link, generate the link expression corresponding to the target planning route; the feature extraction unit is also configured to: extract each link in the target planning route through the converged pre-training model The corresponding target embedding vector; the first predicting unit is configured to: perform forecasting processing of setting services based on the target embedding vector corresponding to each link in the target planned route.

In some embodiments of the present application, based on the foregoing solution, the first prediction unit is configured to: perform at least the following on the target planned route based on the target embedding vectors corresponding to each link in the target planned route One processing: abnormal link detection processing, completion processing of missing links in the target planned route, generation processing of subsequent planned routes, and route correlation processing.

In some embodiments of the present application, based on the foregoing solution, the route processing device further includes: a second prediction unit; the first generation unit is further configured to: plan the links included in the route according to the target to be processed, And the attribute information of each link, generate the link expression corresponding to the target planning route; the second prediction unit is configured to: use the link expression as the input of the converged pre-training model, through the The converged pre-training model outputs the prediction result of the target planned route for the set business; wherein the set business includes at least one of the following: the coverage rate of the driving trajectory and the target planned route, the The travel time of the target planned route, the planning rationality of the target planned route, the navigation completion rate of the target planned route, and the traffic of the target planned route.

According to an aspect of the embodiments of the present application, a computer-readable medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the route processing method as described in the foregoing embodiments is implemented.

According to an aspect of the embodiments of the present application, an electronic device is provided, including: one or more processors; a storage device for storing one or more computer programs, when the one or more computer programs are executed by the When executed by one or more processors, the electronic device implements the route processing method as described in the foregoing embodiments.

According to an aspect of the embodiments of the present application, a computer program product is provided, where the computer program product includes a computer program, and the computer program is stored in a computer-readable storage medium. The processor of the electronic device reads and executes the computer program from the computer-readable storage medium, so that the electronic device executes the route processing methods provided in the various optional embodiments above.

In the technical solution provided by some embodiments of the present application, the link expression corresponding to the planned route sample is generated according to the multiple links contained in the planned route sample and the attribute information of each link, and then the pre- The training model extracts multiple initial embedding vectors corresponding to each element in the link expression, and generates input feature vectors corresponding to each element according to multiple initial embedding feature vectors, and then performs multi-layer forwarding according to the input feature vectors corresponding to each element. Propagation processing, and model training processing based on the planned route samples, training tasks corresponding to the pre-training model, and intermediate result vectors corresponding to each element, to process the specified route based on the converged pre-training module, so that the unlabeled The planned route samples are processed, which solves the problem of a huge workload caused by the need for a large amount of labeled data for model training, and because the detailed features in the planned route samples can be learned through multi-layer forward propagation processing, it can be improved. The accuracy of business processing can also handle different business settings, realizing the effective application of multiple business fields.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the application.

Description of drawings

FIG. 1 shows a schematic diagram of an exemplary system architecture to which the technical solutions of the embodiments of the present application can be applied;

FIG. 2 shows a flowchart of a route processing method according to an embodiment of the present application;

FIG. 3 shows a flowchart of a route processing method according to an embodiment of the present application;

FIG. 4 shows a flow chart of a route processing method according to an embodiment of the present application;

FIG. 5 shows a flow chart of route processing according to an embodiment of the present application;

Fig. 6 shows a schematic diagram of interrupting a route according to an embodiment of the present application;

FIG. 7 shows a schematic diagram of embedding features corresponding to links in a route according to an embodiment of the present application;

Fig. 8 shows a schematic diagram of Transformer forward propagation according to an embodiment of the present application;

Fig. 9 shows a schematic diagram of processing probability for each link according to an embodiment of the present application;

Fig. 10 shows a schematic diagram of model fine-tuning according to an embodiment of the present application;

Fig. 11 shows a block diagram of a route processing device according to an embodiment of the present application;

Fig. 12 shows a schematic structural diagram of a computer system suitable for implementing the electronic device of the embodiment of the present application.

Detailed ways

Example embodiments will now be described in a more complete manner with reference to the accompanying drawings. Example embodiments may, however, be embodied in various forms and should not be construed as limited to these examples; rather, these embodiments are provided so that this application will be thorough and complete, and to fully convey the concepts of example embodiments communicated to those skilled in the art.

Furthermore, the features, structures, or characteristics described herein may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are included so that the embodiments of the present application can be fully understood. However, those skilled in the art should realize that when implementing the technical solutions of the present application, it is not necessary to use all the detailed features in the embodiments, one or more specific details can be omitted, or other methods, elements, and devices can be used , steps, etc.

The block diagrams shown in the drawings are merely functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, or in one or more hardware modules or integrated circuits, or in different networks and/or processor devices and/or microcontroller devices entity.

The flow charts shown in the drawings are only exemplary illustrations, and do not necessarily include all contents and operations/steps, nor must they be performed in the order described. For example, some operations/steps can be decomposed, and some operations/steps can be combined or partly combined, so the actual order of execution may be changed according to the actual situation.

It should be noted that: the "plurality" mentioned in this article refers to two or more than two. "And/or" describes the association relationship of associated objects, indicating that there may be three types of relationships. For example, A and/or B may indicate: A exists alone, A and B exist simultaneously, and B exists independently. The character "/" generally indicates that the contextual objects are an "or" relationship.

It can be understood that in the specific implementation of this application, related data such as planned routes are involved. When the above embodiments of this application are applied to specific products or technologies, it is necessary to obtain user permission or consent, and the collection of relevant data , use and processing need to comply with relevant laws, regulations and standards of relevant countries and regions.

As shown in Figure 1, in the field of map navigation, after the map user selects the starting point and the ending point on the map, he can plan a route from the starting point to the ending point, specifically as shown in Figure 1. The planned routes of Link1, Link2, Link3, and Link4. Among them, Link is the smallest data unit describing a road, and it is a set of structured data, including but not limited to attributes such as the length, width, and road grade of a Link.

After the planned route is obtained, related business processing such as rationality evaluation can usually be performed on the planned route, so as to improve the path planning strategy based on the evaluation result. A machine learning algorithm in artificial intelligence (Artificial Intelligence, AI for short) may be used when evaluating the planned path. in. Artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the nature of intelligence and produce a new kind of intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making. Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes several major directions such as computer vision technology, speech processing technology, natural language processing technology, machine learning/deep learning, automatic driving, and intelligent transportation.

Machine learning (Machine Learning, ML for short) is a multi-field interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. Specializes in the study of how computers simulate or implement human learning behaviors to acquire new knowledge or skills, and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its application pervades all fields of artificial intelligence. Machine learning and deep learning usually include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching and learning.

In related technologies, when processing related businesses such as rationality evaluation on planned routes, developers usually need to model based on industry experience and implement model training based on feature engineering. This solution not only has a workload Huge problem, and less accurate business processing results.

Based on this, a new route processing scheme is proposed in one embodiment of the present application. As shown in FIG. 1, when the vehicle terminal 101 selects the starting point and the ending point on the electronic map, the Generate a planned route including Link1, Link2, Link3 and Link4. In order to perform business processing such as rationality evaluation on the planned route, the server 102 may train a pre-trained model based on the planned route sample, and then evaluate the planned route through the converged pre-trained model.

In one embodiment of the present application, when the server 102 trains the pre-training model based on the planned route samples, it can generate the corresponding The link expression; then extract multiple initial embedding vectors corresponding to each element in the link expression through the pre-training model, and generate input features corresponding to each element in the link expression according to these multiple initial embedding feature vectors vector; then perform multi-layer forward propagation processing according to the input feature vector corresponding to each element in the link expression, and based on the planned route sample, the training task corresponding to the pre-training model, and the multi-layer forward transmission processing The intermediate result vector corresponding to each element is processed for model training; after the model converges, the specified route can be processed based on the converged pre-training module. For example, the target embedding vector corresponding to each element in the link expression can be extracted through the converged pre-training model, and the prediction processing of setting services can be performed based on the target embedding vector. For example, evaluate the rationality of the planned route samples.

It should be noted that the server 102 can be an independent physical server, or a server cluster or distributed system composed of at least two physical servers, and can also provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage Cloud servers for basic cloud computing services such as network services, cloud communications, middleware services, domain name services, security services, content delivery network (ContentDelivery Network, CDN), and big data and artificial intelligence platforms. The vehicle terminal 101 can specifically refer to a smart phone with vehicle functions, a smart speaker, a speaker with a screen, a smart watch, etc., but is not limited thereto. For example, the vehicle terminal 101 can also be replaced by a mobile terminal such as an aircraft. Each vehicle terminal and server can be directly or indirectly connected through wired or wireless communication, and the number of vehicle terminals and servers can be one or at least two, which is not limited in this application.

The implementation details of the technical solutions of the embodiments of the present application are described in detail below:

FIG. 2 shows a flow chart of a route processing method according to an embodiment of the present application. The route processing method may be executed by a server or by a terminal device (such as the vehicle terminal 101 shown in FIG. 1 ), It can also be completed jointly by the server and the terminal device. Referring to Fig. 2, the route processing method includes at least step S210 to step S230, which are described in detail as follows:

In step S210, a link expression corresponding to the planned route sample is generated according to the multiple links contained in the planned route sample and the attribute information of each link.

In some optional embodiments, a link is a Link, which is the smallest data unit describing a road, and the attribute information of the link includes but not limited to the ID, length, width, road grade, road condition, traffic and other attributes of the link .

In some optional embodiments, according to the multiple links contained in the planned route sample and the attribute information of each link, the process of generating the link expression corresponding to the planned route sample may be: according to each link At least one attribute feature contained in the attribute information of the link, generate the elements corresponding to each link used to represent the attribute feature, and then according to the order of the multiple links in the planned route sample, the elements corresponding to the multiple links are respectively Combined to generate the link expression corresponding to the planned route sample. For example, if the link expression is generated according to the ID of the link, then the planned route sample R can be expressed as: route R={link ₁ , link ₂ , link ₃ ...}; where link ₁ , link ₂ , link ₃ ... are respectively Indicates the ID of the corresponding link, which is the element corresponding to the link in the link expression.

If the link expression is generated according to the road level, road condition and flow of the link, then the planned route sample R is expressed as: route R = {rc_rs_flpw ₁ , rc_rs_flow ₂ , rc_rs_flow ₃ ...}; where rc_rs_flow ₁ represents the link ₁ Road grade, traffic status and flow are the elements corresponding to link 1 in the link expression; rc_rs_flow ₂ indicates the road grade, traffic status and flow of link ₂ , which is corresponding to the link in the link expression 2 elements; rc_rs_flow ₃ indicates the road grade, traffic status and flow of link ₃ , which is the element corresponding to link 3 in the link expression.

In some optional embodiments, when generating the link expression corresponding to the planned route sample, the specified link can also be added to the link expression according to the position of the specified link in the planned route sample among the multiple links The position flag corresponding to the road; wherein, the position flag includes at least one of the following: the first flag added before the element corresponding to the start link in the planned route sample, the element corresponding to the two specified links The second flag bit is added between elements, and the third flag bit is added after the element corresponding to the terminating link in the planned route sample. That is, the link expression is divided into the first half and the second half through the second flag bit. Optionally, the first flag bit may be CLS, and both the second flag bit and the third flag bit may be SEP.

In some optional embodiments, when the link expression is divided into the first half and the second half, it can be divided at the road intersection located in the planned route sample, or at any two adjacent links to divide. That is, the two links specified in the foregoing embodiment include: a link located at a road intersection in the planned route sample, or any two adjacent links among multiple links.

Continuing to refer to FIG. 2, in step S220, multiple initial embedding vectors corresponding to each element in the link expression are extracted through the pre-training model, and input feature vectors corresponding to each element are generated according to the multiple initial embedding feature vectors.

It should be noted that the elements in the link expression include the elements corresponding to each link for representing attribute features, and may also include the elements corresponding to the position flags in the above embodiments.

In some optional embodiments, the multiple initial embedding vectors corresponding to each element may include: a word embedding vector (tokenembedding), a position embedding vector (positionembedding) and a route interval embedding vector (segmentembedding) corresponding to each element; wherein, position The embedding vector is used to indicate the specific position of the element in the link expression, and the route interval embedding vector is used to indicate the interval where the element is located in the link expression (that is, whether it is located in the first half or the second half). Optionally, the multiple initial embedding vectors corresponding to each element may also include: word embedding vectors and position embedding vectors.

In some optional embodiments, the process of generating the input feature vector corresponding to each element according to the multiple initial embedded feature vectors may be: based on the set merging processing method, merging the multiple initial embedded feature vectors corresponding to each element Processing, the result of the merge processing is used as the input feature vector corresponding to each element. For example, multiple initial embedding vectors corresponding to each element may be superimposed, and the superposition result may be used as a corresponding input feature vector.

In step S230, perform multi-layer forward propagation processing according to the input feature vector corresponding to each element, and based on the planned route samples, training tasks corresponding to the pre-training model, and the intermediate results corresponding to each element obtained by multi-layer forward transfer processing The vector performs model training processing to process the specified route based on the converged pre-training module.

In some optional embodiments, the process of performing model training processing based on the planned route samples, the training tasks corresponding to the pre-trained model, and the intermediate result vectors corresponding to each element obtained from the multi-layer forward transmission process may include: The other links in the sample except the initial link are replaced by the set probability to obtain the processed planned route sample; the replacement operation includes at least one of the following: replace with set flag bit, replace with The random link remains unchanged; then the intermediate result vector corresponding to each element is used as input, and the link represented by the set flag bit is restored to the actual link in the planned route sample as the optimization target corresponding to the first training task , to perform model training processing. The technical solution of this embodiment is to perform probabilistic replacement of the links in the original planned route samples except the initial link, and then perform model training by recovering the planned route samples as optimization targets.

Optionally, when performing replacement operations on links other than the initial link in the planned route sample by using set probabilities, the other links in the planned route sample except the initial link can be replaced first. The first probability is used to determine whether to perform the replacement operation. For the link that needs to be replaced, the second probability of setting the flag bit is replaced, the third probability is replaced by a random link, and the unchanged first The four probabilities perform replacement operations respectively. For example, the first probability may be 50%, the second probability may be 50%, the third probability may be 5%, and the fourth probability may be 45%. That is, the sum of the second probability, the third probability and the fourth probability is 1.

In some optional embodiments, the process of performing model training processing based on the planned route samples, the training tasks corresponding to the pre-trained model, and the intermediate result vectors corresponding to each element obtained from the multi-layer forward transmission process may include: The second half of the links in the sample are replaced with random links by the set probability; among them, the planned route sample is divided into the first half of the links and the second half of the links at the two specified links; the link expression In the formula, the intermediate result vector corresponding to the first flag bit added before the element corresponding to the starting link is used as input, and the random link is restored to the second half of the link in the planned route sample as the second training task. The optimization goal of , and perform model training processing. The technical solution of this embodiment is to replace the second half of the original planned route samples with probability, and then perform model training by recovering the planned route samples as optimization targets.

Optionally, the specified route in step S230 may be a planned route sample, may also be a route planned during actual application, or may be any selected route.

In some optional embodiments, after the pre-training model converges, the specified route is processed based on the converged pre-training module, specifically, the target embedding vector corresponding to each element may be extracted through the converged pre-training model; According to the target embedding vector corresponding to each element, the route embedding vector corresponding to the planned route sample is generated, and then the route embedding vector and the target input information used to adjust the pre-training model are used as input, and the business is set as the optimization goal. The model parameters of the pre-trained model are adjusted.

In the embodiment of the present application, the target embedding vector corresponding to each element may also include a token embedding vector (token embedding), a position embedding vector (position embedding) and a route segment embedding vector (segment embedding).

The technical solution of this embodiment is to fine-tune the pre-trained model through the target embedding vector and target input information corresponding to each element. Optionally, the setting service in this embodiment includes at least one of the following: evaluating the coverage rate of the driving track and the planned route sample, estimating the driving time of the planned route sample, evaluating the planning rationality of the planned route sample, and estimating the planning The navigation completion rate of the route sample, and the estimated flow of the planned route sample.

In some optional embodiments, according to the target embedding vector corresponding to each element, the process of generating the route embedding vector corresponding to the planned route sample may be: according to the multiple target embedding vectors corresponding to the starting link in the planned route sample , calculate the mean value of the target embedding vector corresponding to the starting link; calculate the mean value of the target embedding vector corresponding to multiple links according to the multiple target embedding vectors corresponding to the multiple links included in the planned route sample; according to the planned route According to the multiple target embedding vectors corresponding to the termination link in the sample, the mean value of the target embedding vector corresponding to the termination link is calculated; then according to the mean value of the target embedding vector corresponding to the starting link, the mean value of the target embedding vector corresponding to multiple links and The mean value of the target embedding vector corresponding to the termination link is used to generate the route embedding vector corresponding to the planned route sample. For example, the mean value of target embedding vectors corresponding to the starting link, the mean value of target embedding vectors corresponding to multiple links, and the mean value of target embedding vectors corresponding to the terminating link may be used as the route embedding vectors corresponding to the planned route samples.

In some optional embodiments, the target input information used to adjust the pre-training model may include at least one of the following: the route features of the planned route samples obtained through feature engineering; according to the route features of the planned route samples, by The machine learning model is the prediction result obtained by performing the prediction processing of the set business. Optionally, the machine learning model may be a decision tree model.

Optionally, the route features may include one or more of mined features, real-time features and static features. The excavated features can include the historical arrival time and historical driving speed of the planned route samples; the real-time features can include the real-time road conditions and real-time driving speed of the planned route samples; the static features can include the road attributes of the planned route samples, such as road width, Road type (such as expressway, provincial road or national road), etc.

FIG. 3 shows a flow chart of a route processing method according to an embodiment of the present application. The route processing method may be executed by a server or by a terminal device (such as the vehicle terminal 101 shown in FIG. 1 ), It can also be completed jointly by the server and the terminal device. Referring to Figure 3, the route processing method includes at least step S310 to step S330, which are described in detail as follows:

In step S310, a link expression corresponding to the planned target route is generated according to the links contained in the planned target route to be processed and the attribute information of each link.

The specific implementation details of this step are similar to those of step S210 and will not be repeated here.

In step S320, the target embedding vector corresponding to each link in the target planned route is extracted through the converged pre-training model.

In the embodiment of the present application, the pre-training model can be trained through the technical solution of the embodiment shown in Figure 2, and the target embedding vector corresponding to each link can include a word embedding vector (token embedding), a position embedding vector (position embedding) and route interval embedding vector (segmentembedding).

In step S330 , based on the target embedding vectors corresponding to each link in the target planned route, the forecast processing of setting traffic is performed.

In some optional embodiments, based on the target embedding vectors corresponding to the links in the target planned route, at least one of the following processing may be performed on the target planned route: abnormal link detection processing, missing links in the target planned route Road completion processing, follow-up planning route generation processing, and route correlation processing.

Specifically, for the abnormal link detection service, a prediction value can be generated for each link in the target planning route through the target embedding vector, and the prediction value indicates the probability that each link belongs to the target planning route, if If the predicted value corresponding to a certain link is small, it may indicate that there may be an anomaly in this link. The completion of the missing link can be when a certain link is missing in the target planning route, and the target embedding vector is used to predict which link the missing link should be supplemented with. The generation process of the subsequent planned route may be to predict the missing link or the link that should be supplemented between the starting location and the ending location according to the target embedding vector after the starting location and the ending location are known. The correlation processing of the routes may be to calculate the distance between the routes according to the target embedding vectors corresponding to the routes, and then determine the correlation according to the distance value.

FIG. 4 shows a flow chart of a route processing method according to an embodiment of the present application. The route processing method may be executed by a server or by a terminal device (such as the vehicle terminal 101 shown in FIG. 1 ), It can also be completed jointly by the server and the terminal device. Referring to Figure 4, the route processing method includes at least step S410 to step S420, which are described in detail as follows:

In step S410, a link expression corresponding to the planned target route is generated according to the links contained in the planned target route to be processed and the attribute information of each link.

The specific implementation details of this step are similar to those of step S210 and will not be repeated here.

In step S420, the link expression is used as an input of the converged pre-training model, and the predicted result of the target planned route for the set service is output through the converged pre-training model.

In some optional embodiments, the setting service includes at least one of the following: the coverage rate of the driving trajectory and the target planned route, the driving time of the target planned route, the planning rationality of the target planned route, and the navigation completion rate of the target planned route , The traffic of the target planning route.

The technical solutions of the above-mentioned embodiments of the present application can solve the problem of huge workload caused by the need for a large amount of labeled data for model training by processing unlabeled planned route samples, and because it can be processed through multi-layer forward propagation The detailed features in the planned route samples are learned, so the accuracy of business processing can be improved, and different set businesses can be processed at the same time, realizing the effective application in multiple business fields.

The implementation details of the technical solution of the embodiment of the present application are described in detail below in conjunction with FIGS. 5 to 10 , taking the rationality of the estimated route as an example:

In related technologies, there are mainly the following schemes for estimating the rationality of routes: Scheme 1, by manually modeling the route, using a machine learning model to estimate the rationality of a characteristic-fixed-length route sample, machine learning The model may be, for example, a logistic regression model, a GBDT (Gradient Boosting Decision Tree, gradient descent tree) model, a LambdaRank (a sorting algorithm) model, and the like. Solution 2, through the method of deep learning combined with manual modeling, such as manual feature extraction, fixed-length feature modeling of route samples, combined with the route rationality estimation scheme of the deep model structure. Solution 3, through the embedding method, deep learning, and artificial modeling methods, such as through the two-tower model structure, artificial feature extraction is performed on the route and users, and embedding is performed on the data units and users of the route respectively, and then through the deep network The structure makes a reasonable estimate of the route.

However, the manual modeling in Scheme 1 and Scheme 2 in related technologies strongly depends on the industry experience and business understanding of developers, and it is difficult to achieve the completeness of business modeling, so it is difficult to achieve the optimal effect under business goals. The route embedding scheme in Scheme 3 is carried out based on the link topology of the route. The information of the embedding result mainly includes the topological attributes of the road network formed by the route, which is difficult to match with the expected business goals, and the applicable scenarios are small.

Based on this, in the technical solution provided by the present application, as shown in FIG. 5 , the data used include high-quality navigation routes, which means that routes composed of link strings are reasonable and acceptable. The quantification of rationality usually includes: the coverage rate of the historical driving trajectory and the navigation route, the navigation completion rate under the navigation route, the rationality probability of the navigation route, etc.

It should be noted that the navigation route is composed of a series of continuous links in the topological structure of the road network. The link data unit includes: 1. Geographic location information of the road, such as latitude and longitude coordinates, length, road grade, speed limit, administrative subordinate information, etc. ;2. Road topology information, such as upstream and downstream linkids, surrounding facilities, etc.; 3. Real-time information, such as vehicle speed at a certain moment, road conditions, vehicle speed at historical moments, etc.

After the high-quality navigation route is obtained, it can be processed to obtain the pre-training route data.

Specifically, the following designs can be used as input samples for the pre-training model:

First, use the id string of the link as the expression of the route sample. For example, if the link expression is generated according to the ID of the link, then the route sample R can be expressed as: route R={link ₁ , ink ₂ , ink ₃ ...}; where link ₁ , ink ₂ , ink ₃ ... respectively represent ID of the corresponding link.

It is also possible to use the combination of link attributes as the expression of route samples for different business scenarios. For example, route sample R={road grade_traffic status_flow...}, specifically expressed as: route R={rc_rs_flow ₁ ,c_rs_flow ₂ ,c_rs_flow ₃ …}; Among them, rc__flow ₁ indicates the road level, traffic status and flow of link ₁ ; rc__flow ₂ indicates the road level, traffic status and flow of link ₂ ; rc__flow ₃ indicates the road level, traffic status and flow of link ₃ .

It should be noted that for routes R with different expressions, pre-training will produce different embedding results.

Secondly, as shown in Figure 6, the route R is interrupted at any intersection of the route R, or the route R can be interrupted between any two links to form the first half and the second half of the route. Then add the flag bit CLS before the start link of the route R, and add the flag bit SEP after the break and end link. Then get an input sample R _sample of a pre-trained model:

R _sample = {CLS, link ₁ , link ₂ , link ₃ , SEP, link ₄ , link ₅ , link ₆ , SEP}

After obtaining the input samples of the pre-training model, step S501 shown in FIG. 5 can be executed to perform the pre-training process of the model.

In one embodiment of the present application, a BERT (Bidirectional Encoder Representations from Transformers, Transformer-based bidirectional encoder representation) model may be used for pre-training. The input training samples of the BERT model are the first half and the second half of a sentence, which may correspond to the first half and the second half of the route in the embodiment of this application, and a link in the route may correspond to a word in a sentence. Specifically, as shown in FIG. 7 , a link in a route may have three types of embeddings: TokenEmbedding, SegmentEmbedding, and PositionEmbedding. Among them, Segment Embeddings identifies whether a link is located in the first half or the second half. If it is located in the first half, it can be expressed as E _A ; if it is located in the second half, it can be expressed as E _B . Position Embeddings are used to represent the positional characteristics of each link in the route.

The forward propagation of the BERT model in the embodiment of this application is divided into two stages: the forward propagation of the Transformer structure and the forward propagation of two subtasks, which will be introduced separately below.

The forward propagation process of the Transformer structure is as follows: in the Embedding stage, for each link and flag in the route sample, search for the corresponding Token Embedding in the Embedding data; according to the part of the route where each link and flag are located, find the corresponding Token Embedding Segment Embedding; Find the corresponding Position Embedding according to the route position of each link and marker. Then the TokenEmbedding, Segment Embedding, and Position Embedding vectors found by each link and flag bit are accumulated (in other embodiments of the application, it can also be based on Token Embedding, Segment Embedding, and Position Embedding) to obtain this The embedding _input in this route sample is a link or flag, that is, E ₁ , E ₂ , . . . , E _N shown in FIG. 8 .

Continue to refer to shown in Figure 8, use Embedding _input as input to carry out multi-layer Transformer forward propagation (shown in Figure 8 the example of two layers, can be more layers in other embodiments), when the last layer of Transformer output , to obtain the intermediate result vector Transfermor _out , that is, T ₁ , T ₂ , ..., T _N shown in Figure 8, depending on the attention mechanism of the Transformer structure, the intermediate result Transfermor _out of the corresponding position of the Embedding _input is fused with the route sample Link information for other locations.

The forward propagation process of the two subtasks is specifically: because the BERT model designs two subtasks, namely MLM (Mask Language Model, mask language model) and NSP (Next Sentence Prediction, next sentence prediction).

For the MLM task, the input sample R _sample is first processed, specifically: the first link of the sample is not operated; for each of the remaining links, as shown in Figure 9, a link is operated with the probability of P _op , and the operated link is The probability of P _MASK is set as the flag MASK, the probability of P _RAND is replaced with a random link, and the probability of P _KEEP is kept unchanged. Optionally, P _op =50%, P _MASK =50%, P _RAND =5%, P _KEEP =45%.

The processed samples include flag MASK and random link _RAND , and the intermediate result vector Transfermor _out of each link and flag in the MLM task input route is calculated for each MASK flag through the DNN (Deep Neural Network, deep neural network) model softmax, get the probability distribution of all links on the MASK flag, and then compare the true value linkid set as the MASK flag, calculate the gradient and backpropagate. Optionally, in other embodiments of the present application, the DNN model may also be replaced with other deep learning network models.

For the NSP task, the input sample R _sample is firstly processed, specifically, the second half of the route identified by the SEP flag is replaced with a random second half of the route with the probability of P _op , optionally, P _op =50%. The NSP task inputs the intermediate result vector Transfermor _{out of} the CLS flag at the starting position of the route, calculates the softmax through the DNN model, and obtains the probability that the second half of the route belongs to the R _sample , compares whether it is replaced, calculates the gradient and backpropagates.

Through pre-training, it is possible to optimize the converged Token Embedding, SegmentEmbedding, Position Embedding results, and the Transformer structure of each layer in a large number of route samples.

The fine-tuning process of step S502 shown in FIG. 5 specifically deals with the route business as follows: First, the embedding expression of the route level is composed of the embedding of the start link, the average embedding of the route and the embedding of the end link in the route. Specifically as follows:

Embedding _start ＝∑Token Embedding _start ,Segment Embedding _start ,PositionEmbedding _start

Embedding _end ＝ΣToken Embedding _end ,Segment Embedding _end ,PositionEmbedding _end

Among them, TokenEmbedding _start , SegmentEmbedding _start and PositionEmbedding _start are the embeddings of the starting link; TokenEmbedding _end , SegmentEmbedding _end and PositionEmbedding _end are the embeddings of the ending link; TokenEmbedding _i , SegmentEmbedding _i and PositionEmbedding _i are the embeddings of the i-th link; n indicates the route The number of links in . In other embodiments of the present application, the route-level embedding may also be based on other calculation and processing results of Token Embedding, Segment Embedding, and Position Embedding vectors.

Secondly, the route features designed by artificial feature engineering and the estimated results output by the machine learning method of the decision tree model can be obtained, and these two parts can be used as the input information of the fine-tuning model.

Optionally, the route features may include one or more of mined features, real-time features and static features. The excavated features can include the historical arrival time and historical driving speed of the planned route samples; the real-time features can include the real-time road conditions and real-time driving speed of the planned route samples; the static features can include the road attributes of the planned route samples, such as road width, Road type (such as expressway, provincial road or national road), etc.

Finally, the coverage rate T of the historical driving trajectory and the navigation route is used as the quantitative standard of the route rationality to fine-tune the model as the optimization target of the model. A more reasonable route has a higher coverage rate, where T∈(0,1) . The fine-tuning process is shown in Figure 10. Route-level embedding vectors (ie, Embedding _start , Embedding _avg , and Embedding _end ) are transformed into D1 vectors through the multi-layer perceptron MLP ₁ . Link the Embedding _start , Embeddingavg and Embeddingend vectors, and transform them into D2 vectors, D3 vectors and D4 vectors through multi-layer perceptrons MLP2, MLP ₃ and MLP ₄ . Then, the artificially extracted features and the prediction results of the tree model are used as input, and the D5 vector is constructed by linking D4. The D6 vector, the D7 vector and the D8 vector are obtained by transforming the multi-layer perceptrons MLP ₅ , MLP ₆ and MLP ₇ . Optionally, D1=128, D2=1536, D3=384, D4=8, D5=8+N(features)+1, D6=64, D=16, D8=1. The D8 vector is a 1-dimensional vector, which is the output of route rationality estimation results. The number of layers of the multi-layer perceptron shown in FIG. 10 is only an example, and there may be more layers in other embodiments of the present application.

In the above technical solution of this application, the Token Embedding, SegmentEmbedding, and Position Embedding results obtained through pre-training can fully consider whether the composition of the entire route is reasonable in the process of route rationality estimation. In machine learning schemes that are different from general manual feature extraction, due to the design of the global feature of the route, the model will lose the detailed information of the route during the prediction process. In the above technical solution of the present application, based on the attention mechanism in the Transformer structure, routes with higher rationality have better quality route parts. In addition, the embedding results obtained through pre-training in the above-mentioned embodiments of the present application can also be used for abnormal route detection, route completion, automatic route generation, road correlation analysis, urban road planning, etc. The function of the route. In addition, the technical solutions of the above-mentioned embodiments of the present application can not only perform the route rationality prediction task, but also can optimize the target by modifying the pre-training subtask, and can also be used for route time estimation, route rationality quantification, route probability estimation, Route traffic forecasting and other services.

The technical solutions of the above-mentioned embodiments of the present application are an unsupervised learning solution in the pre-training stage. In the navigation business scenario, route data is a kind of data that is difficult to collect user feedback. For the labeled data, the navigation service-oriented information in the unlabeled data can be extracted by using the pre-training method. The information obtained through pre-training can be used to use smaller-scale labeling data in the fine-tuning model stage to obtain a more effective route rationality prediction model, which can effectively solve problems such as under-labeling of data in the business field, data noise, and Position-Bias. . In addition, the information obtained in the pre-training stage is rich and diverse, and the information tends to depend on the structural design of the pre-training scheme. Therefore, for different route learning objectives, the pre-training model structure can be designed in a targeted manner. The technical solution of the embodiment of the application has strong Versatility and scalability, applicable to business problems with different goals. For example: route time estimation, route rationality quantification, route probability estimation, route traffic estimation, etc.

The following introduces device embodiments of the present application, which can be used to implement the route processing method in the foregoing embodiments of the present application. For the details not disclosed in the device embodiment of the present application, please refer to the embodiment of the above-mentioned route processing method in the present application.

Fig. 11 shows a block diagram of a route processing device according to an embodiment of the present application.

Referring to FIG. 11 , a route processing device 1100 according to an embodiment of the present application includes: a first generation unit 1102 , a feature extraction unit 1104 , a second generation unit 1106 and a model training unit 1108 .

Wherein, the first generation unit 1102 is configured to generate a link expression corresponding to the planned route sample according to the multiple links contained in the planned route sample and the attribute information of each link; the feature extraction unit 1104 is configured to pass The pre-training model extracts multiple initial embedding vectors corresponding to each element in the link expression; the second generating unit 1106 is configured to generate an input feature vector corresponding to each element according to the multiple initial embedding feature vectors; the model The training unit 1108 is configured to perform multi-layer forward propagation processing according to the input feature vector corresponding to each element, and based on the planned route sample, the training task corresponding to the pre-training model, and the multi-layer forward propagation processing The obtained intermediate result vector corresponding to each element is subjected to model training processing, so as to process the specified route based on the converged pre-training module.

In some embodiments of the present application, based on the foregoing solution, the first generation unit 1102 is configured to: generate the corresponding link information of each link according to at least one attribute feature included in the attribute information of each link. An element used to represent an attribute feature; according to the order of the multiple links in the planned route sample, combine elements corresponding to the multiple links respectively to generate a link expression corresponding to the planned route sample Mode.

In some embodiments of the present application, based on the foregoing solution, the first generating unit 1102 is configured to: according to the position of the specified link among the multiple links in the planned route sample, express in the link The location flag corresponding to the specified link is added in the formula; wherein, the location flag includes at least one of the following: the first flag added before the element corresponding to the starting link in the planned route sample , the second flag bit added between the elements corresponding to the two specified links, and the third flag bit added after the element corresponding to the termination link in the planned route sample.

In some embodiments of the present application, based on the foregoing solution, the specified two links include: a link located at a road intersection in the planned route sample, or any two adjacent links among the multiple links adjacent links.

In some embodiments of the present application, based on the foregoing solution, the feature extraction unit 1104 is configured to: extract the word embedding vector, position embedding vector and route interval embedding corresponding to each element in the link expression through a pre-training model vector; wherein, the position embedding vector is used to indicate the specific position of the element in the link expression, and the route interval embedding vector is used to indicate the location of the element in the link expression interval.

In some embodiments of the present application, based on the foregoing solution, the second generation unit 1106 is configured to: perform a merge process on the multiple initial embedded feature vectors corresponding to each element based on the set merge processing method, and combine The processing result is used as the input feature vector corresponding to each element.

In some embodiments of the present application, based on the foregoing solution, the model training unit 1108 is configured to: perform replacement operations on the links other than the initial link in the planned route samples with set probabilities , to obtain the processed planned route sample; the replacement operation includes at least one of the following: replacing with a set flag bit, replacing with a random link, and keeping it unchanged; taking the intermediate result vector corresponding to each element as input, and The link represented by the set flag bit is restored to the actual link in the planned route sample as the optimization target corresponding to the first training task, and the model training process is performed.

In some embodiments of the present application, based on the foregoing solution, performing replacement operations on links other than the initial link in the planned route samples according to set probabilities, including: For links other than the initial link, determine whether to perform the replacement operation through the first probability; for the link that needs to be replaced, replace it with the second probability of setting the flag bit, and replace it with a random chain The third probability of the way and the fourth probability that remains unchanged are replaced respectively.

In some embodiments of the present application, based on the aforementioned solutions, the model training unit 1108 is configured to: replace the second half of the links in the planned route samples with random links with a set probability; wherein, The planned route sample is divided into the first half link and the second half link at the two specified links; the first flag bit added before the element corresponding to the starting link in the link expression The corresponding intermediate result vector is used as an input, and the random link is restored to the second half of the links in the planned route sample as the optimization target corresponding to the second training task to perform model training processing.

In some embodiments of the present application, based on the foregoing solution, the route processing device 1100 further includes: a processing unit 1110 configured to extract the target embedding vector corresponding to each element through the converged pre-training model; The target embedding vector corresponding to the element is used to generate the route embedding vector corresponding to the planned route sample; the route embedding vector and the target input information used to adjust the pre-training model are used as input, and the setting business is used as an optimization The target is to adjust the model parameters of the pre-trained model.

In some embodiments of the present application, based on the foregoing solution, the processing unit 1110 is configured to: calculate and obtain the corresponding target link of the starting link according to the multiple target embedding vectors corresponding to the starting link in the planned route sample. The mean value of the target embedding vectors; according to the multiple target embedding vectors corresponding to the multiple links contained in the planned route sample, calculate the target embedding vector mean value corresponding to the multiple links; according to the planned route sample A plurality of target embedding vectors corresponding to the termination link is calculated to obtain the target embedding vector mean value corresponding to the termination link; according to the target embedding vector mean value corresponding to the start link, the target embedding vector value corresponding to the multiple links The vector mean value and the target embedding vector mean value corresponding to the terminated link generate a route embedding vector corresponding to the planned route sample.

In some embodiments of the present application, based on the foregoing solution, the target input information used to adjust the pre-training model includes at least one of the following: route features of the planned route samples obtained through feature engineering; The route features of the planned route samples are the prediction results obtained by performing the prediction processing of the setting business through the machine learning model.

In some embodiments of the present application, based on the foregoing solutions, the route processing device 1100 further includes: a first prediction unit; the first generating unit 1102 is further configured to: plan the chains included in the route according to the target to be processed Road, and the attribute information of each link, generate the link expression corresponding to the target planning route; the feature extraction unit 1104 is also configured to: extract each of the target planning routes through the converged pre-training model target embedding vectors corresponding to each link; the first predicting unit is configured to: perform forecasting processing of setting services based on the target embedding vectors corresponding to each link in the target planned route.

In some embodiments of the present application, based on the foregoing solution, the first prediction unit is configured to: perform at least the following on the target planned route based on the target embedding vectors corresponding to each link in the target planned route One processing: abnormal link detection processing, completion processing of missing links in the target planned route, generation processing of subsequent planned routes, and route correlation processing.

In some embodiments of the present application, based on the foregoing solution, the route processing device 1100 further includes: a second prediction unit; the first generation unit 1102 is further configured to: plan the chains included in the route according to the target to be processed Road, and the attribute information of each link, generate the link expression corresponding to the target planned route; the second prediction unit is configured to: use the link expression as the input of the converged pre-training model, Outputting the prediction result of the target planned route for the set business through the converged pre-training model; wherein the set business includes at least one of the following: the coverage rate of the driving trajectory and the target planned route, The travel time of the target planned route, the planning rationality of the target planned route, the navigation completion rate of the target planned route, and the traffic of the target planned route.

Fig. 12 shows a schematic structural diagram of a computer system suitable for implementing the electronic device of the embodiment of the present application.

It should be noted that the computer system 1200 of the electronic device shown in FIG. 12 is only an example, and should not limit the functions and scope of use of the embodiments of the present application.

As shown in FIG. 12 , a computer system 1200 includes a central processing unit (Central Processing Unit, CPU) 1201, which can be stored in a program in a read-only memory (Read-Only Memory, ROM) 1202 or loaded from a storage part 1208 to a random It accesses the programs in the memory (Random Access Memory, RAM) 1203 to execute various appropriate actions and processes, for example, executes the methods described in the above-mentioned embodiments. In RAM 1203, various programs and data necessary for system operation are also stored. The CPU 1201 , ROM 1202 , and RAM 1203 are connected to each other via a bus 1204 . An input/output (Input/Output, I/O) interface 1205 is also connected to the bus 1204 .

The following components are connected to the I/O interface 1205: an input section 1206 including a keyboard, a mouse, etc.; an output section 1207 including a cathode ray tube (Cathode Ray Tube, CRT), a liquid crystal display (Liquid Crystal Display, LCD), etc., and a speaker ; a storage section 1208 including a hard disk or the like; and a communication section 1209 including a network interface card such as a LAN (Local Area Network) card, a modem, or the like. The communication section 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to the I/O interface 1205 as needed. A removable medium 1211, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is mounted on the drive 1210 as necessary so that a computer program read therefrom is installed into the storage section 1208 as necessary.

In particular, according to the embodiments of the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments of the present application include a computer program product, which includes a computer program carried on a computer-readable medium, where the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program may be downloaded and installed from a network via communication portion 1209 and/or installed from removable media 1211 . When this computer program is executed by a central processing unit (CPU) 1201, various functions defined in the system of the present application are performed.

It should be noted that the computer-readable medium shown in the embodiment of the present application may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable Programmable Read-Only Memory (Erasable Programmable Read Only Memory, EPROM), flash memory, optical fiber, portable compact disk read-only memory (Compact Disc Read-Only Memory, CD-ROM), optical storage device, magnetic storage device, or any suitable The combination. In the present application, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, however, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, in which a computer-readable computer program is carried. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. . A computer program embodied on a computer readable medium can be transmitted using any appropriate medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. Wherein, each block in the flowchart or block diagram may represent a module, a program segment, or a part of the code, and the above-mentioned module, program segment, or part of the code includes one or more executable instruction. It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or they may sometimes be executed in the reverse order, depending upon the functionality involved. It should also be noted that each block in the block diagrams or flowchart illustrations, and combinations of blocks in the block diagrams or flowchart illustrations, can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a It is realized by a combination of special hardware and computer programs.

The units described in the embodiments of the present application may be implemented by software or by hardware, and the described units may also be set in a processor. Wherein, the names of these units do not constitute a limitation of the unit itself under certain circumstances.

As another aspect, the present application also provides a computer-readable medium. The computer-readable medium may be included in the electronic device described in the above embodiments; it may also exist independently without being assembled into the electronic device. middle. The above-mentioned computer-readable medium carries one or more computer programs, and when the above-mentioned one or more computer programs are executed by an electronic device, the electronic device is made to implement the methods described in the above-mentioned embodiments.

It should be noted that although several modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory. Actually, according to the embodiment of the present application, the features and functions of two or more modules or units described above may be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided to be embodied by a plurality of modules or units.

Through the description of the above implementations, those skilled in the art can easily understand that the example implementations described here can be implemented by software, or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of the present application can be embodied in the form of software products, which can be stored in a non-volatile storage medium (which can be CD-ROM, U disk, mobile hard disk, etc.) or on the network , including several instructions to make a computing device (which may be a personal computer, server, touch terminal, or network device, etc.) execute the method according to the embodiment of the present application.

Other embodiments of the present application will be readily apparent to those skilled in the art from consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any modification, use or adaptation of the application, these modifications, uses or adaptations follow the general principles of the application and include common knowledge or conventional technical means in the technical field not disclosed in the application .

It should be understood that the present application is not limited to the precise constructions which have been described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the application is limited only by the appended claims.

Claims (19)

1. A route processing method, characterized in that, comprising: Generate a link expression corresponding to the planned route sample according to the multiple links contained in the planned route sample and the attribute information of each link; Extracting a plurality of initial embedding vectors corresponding to each element in the link expression through a pre-training model, and generating an input feature vector corresponding to each element according to the plurality of initial embedding feature vectors; Perform multi-layer forward propagation processing according to the input feature vector corresponding to each element, and based on the planned route sample, the training task corresponding to the pre-training model, and the various multi-layer forward propagation processing obtained The intermediate result vector corresponding to the element is subjected to model training processing, so as to process the specified route based on the converged pre-training module. 2. The route processing method according to claim 1, wherein the link expression corresponding to the planned route sample is generated according to the plurality of links contained in the planned route sample and the attribute information of each link ,include: According to at least one attribute characteristic included in the attribute information of each link, generate an element for representing an attribute characteristic corresponding to each link; According to the order of the multiple links in the planned route sample, combine elements corresponding to the multiple links respectively to generate a link expression corresponding to the planned route sample. 3. The route processing method according to claim 2, characterized in that, according to the plurality of links contained in the planned route sample and the attribute information of each link, the link expression corresponding to the planned route sample is generated ,Also includes: According to the position of the specified link among the multiple links in the planned route sample, add the position flag bit corresponding to the specified link to the link expression; Wherein, the location identification bit includes at least one of the following: a first flag bit added before the element corresponding to the start link in the planned route sample, and a first flag bit between the elements corresponding to the specified two links The second flag bit added, and the third flag bit added after the element corresponding to the terminating link in the planned route sample. 4. The route processing method according to claim 3, wherein the specified two links comprise: a link located at a road intersection in the planned route sample, or a link among the plurality of links Any two adjacent links. 5. The route processing method according to claim 1, wherein extracting a plurality of initial embedding vectors corresponding to each element in the link expression through a pre-training model comprises: extracting the link expression through a pre-training model The word embedding vector, position embedding vector and route interval embedding vector corresponding to each element in the road expression; Wherein, the position embedding vector is used to represent the specific position of the element in the link expression, and the route interval embedding vector is used to represent the interval of the element in the link expression . 6. The route processing method according to claim 1, wherein generating an input feature vector corresponding to each element according to the plurality of initial embedded feature vectors comprises: Based on the set merging processing manner, the multiple initial embedded feature vectors corresponding to each element are combined, and the result of the merging process is used as an input feature vector corresponding to each element. 7. The route processing method according to claim 1, characterized in that, based on the planned route samples, the training tasks corresponding to the pre-training model, and the corresponding elements obtained by the multi-layer forward transmission process The intermediate result vector of is used for model training processing, including: For the other links in the planned route samples except the initial link, the replacement operation is respectively performed through the set probability to obtain the processed planned route sample; the replacement operation includes at least one of the following: replacing with the set Set the flag, replace it with a random link, and keep it unchanged; The intermediate result vector corresponding to each element is used as input, and the link represented by the set flag bit is restored to the actual link in the planned route sample as the optimization target corresponding to the first training task, and the model training process. 8. The route processing method according to claim 7, characterized in that, for the other links in the planned route sample except the initial link, the replacement operation is respectively performed according to the set probability, including: For links other than the initial link in the planned route sample, respectively determine whether to perform a replacement operation through the first probability; For the link that needs to be replaced, the replacement operation is performed by replacing with the second probability of setting the flag bit, the third probability of replacing with a random link, and the unchanged fourth probability. 9. The route processing method according to claim 1, characterized in that, based on the planned route samples, the training tasks corresponding to the pre-training model, and the corresponding elements obtained by the multi-layer forward transmission process The intermediate result vector of is used for model training processing, including: For the second half of the links in the planned route samples, the set probability is used to replace them with random links; wherein, the planned route samples are divided into the first half of the links and the second half of the links at the specified two links link; Taking the intermediate result vector corresponding to the first flag bit added before the element corresponding to the starting link in the link expression as input, restoring the random link to the second half of the planned route sample Part of the links are used as optimization targets corresponding to the second training task to perform model training processing. 10. The route processing method according to claim 1, wherein the specified route is processed based on the converged pre-training module, including: Extracting target embedding vectors corresponding to each element through the converged pre-training model; Generate a route embedding vector corresponding to the planned route sample according to the target embedding vector corresponding to each element; The route embedding vector and the target input information for adjusting the pre-training model are used as input, and the set service is used as an optimization target to adjust the model parameters of the pre-training model. 11. The route processing method according to claim 10, wherein generating the route embedding vector corresponding to the planned route sample according to the target embedding vector corresponding to each element comprises: According to the plurality of target embedding vectors corresponding to the starting link in the planned route sample, calculate and obtain the mean value of the target embedding vector corresponding to the starting link; According to the multiple target embedding vectors corresponding to the multiple links included in the planned route sample, calculate and obtain the mean value of the target embedding vectors corresponding to the multiple links; According to the plurality of target embedding vectors corresponding to the terminating link in the planned route sample, calculate and obtain the mean value of the target embedding vector corresponding to the terminating link; Generate a route embedding corresponding to the planned route sample according to the mean value of target embedding vectors corresponding to the starting link, the mean value of target embedding vectors corresponding to the plurality of links, and the mean value of target embedding vectors corresponding to the termination link vector. 12. The route processing method according to claim 10, wherein the target input information for adjusting the pre-training model includes at least one of the following: The route feature of the planned route sample obtained through feature engineering; According to the route characteristics of the planned route sample, the prediction result obtained by performing the prediction processing of the set service through a machine learning model. 13. The route processing method according to any one of claims 1 to 12, wherein the specified route is processed based on the converged pre-training module, including: Generate a link expression corresponding to the target planned route according to the links contained in the target planned route to be processed and the attribute information of each link; Extracting target embedding vectors corresponding to each link in the target planning route through the converged pre-training model; Based on the target embedding vectors corresponding to the links in the target planned route, the forecast processing of setting services is performed. 14. The route processing method according to claim 13, characterized in that, based on the target embedding vectors corresponding to each link in the target planned route, the predictive processing of setting services is performed, comprising: Based on the target embedding vectors corresponding to each link in the target planned route, perform at least one of the following processing on the target planned route: abnormal link detection processing, completion of missing links in the target planned route processing, generation processing of follow-up planning routes, and route correlation processing. 15. The route processing method according to any one of claims 1 to 12, wherein the specified route is processed based on the converged pre-training module, including: Generate a link expression corresponding to the target planned route according to the links contained in the target planned route to be processed and the attribute information of each link; Using the link expression as an input of the converged pre-training model, outputting the prediction result of the target planning route for the set service through the converged pre-training model; Wherein, the setting service includes at least one of the following: the coverage rate of the driving track and the target planned route, the driving time of the target planned route, the planning rationality of the target planned route, the Navigation completion rate, traffic of the target planned route. 16. A route processing device, comprising: The first generation unit is configured to generate a link expression corresponding to the planned route sample according to the multiple links contained in the planned route sample and the attribute information of each link; The feature extraction unit is configured to extract a plurality of initial embedding vectors corresponding to each element in the link expression through a pre-training model; The second generation unit is configured to generate an input feature vector corresponding to each element according to the plurality of initial embedded feature vectors; The model training unit is configured to perform multi-layer forward propagation processing according to the input feature vector corresponding to each element, and based on the planned route sample, the training task corresponding to the pre-training model, and the multi-layer forward propagation The obtained intermediate result vector corresponding to each element is subjected to model training processing, so as to process the specified route based on the converged pre-training module. 17. A computer-readable medium, on which a computer program is stored, wherein, when the computer program is executed by a processor, the route processing method according to any one of claims 1 to 15 is implemented. 18. An electronic device, characterized in that it comprises: one or more processors; Memory, used to store one or more computer programs, when the one or more computer programs are executed by the one or more processors, the electronic device realizes any one of claims 1 to 15 The route processing method described above. 19. A computer program product, characterized in that the computer program product includes a computer program, the computer program is stored in a computer-readable storage medium, and a processor of an electronic device reads and executes the program from the computer-readable storage medium. Executing the computer program causes the electronic device to execute the route processing method according to any one of claims 1 to 15.

Images