A Vehicle Cooperative Localization Method Based on Federated Transfer Learning

Abstract

This application is applicable to the field of intelligent transportation technology and provides a vehicle collaborative positioning method based on federated transfer learning. The vehicle collaborative positioning method includes: the cloud clusters multiple regional states based on a multi-attribute spatio-temporal clustering algorithm to determine a regional state cluster corresponding to the target domain; the edge side aggregates the model weights corresponding to all regional states in the regional state cluster and uses a self-attention mechanism module to obtain the spatio-temporal characteristics of all regional states in the regional state cluster; the edge side filters the aggregated model weights using the spatio-temporal characteristics to obtain the final model weights and constructs a global model using the final model weights; the vehicle side receives the global model sent by the edge side and, based on federated learning, obtains a local model according to the global model; the local model outputs the GPS regional error; the vehicle side uses the GPS regional error to correct its own position information. The vehicle collaborative positioning method of this application can solve the problem of low vehicle positioning accuracy.

Description

Technical Field

[0001] The present application relates to the field of intelligent transportation technology, and in particular to a vehicle collaborative positioning method based on federated transfer learning. Background Art

[0002] The popularity of advanced embedded and communication systems has promoted the penetration of many smart services into our daily lives, such as smart environments, smart cities, universal healthcare, and intelligent transportation systems. However, in order to realize and provide these smart services, contextual information is required in addition to the actual sensor data. Among various contextual information, location information is one of the most critical ones, because many smart services are location-dependent, which is often referred to as location-aware services. Smart navigation is a good example of location-aware services. For many outdoor applications, the Global Positioning System (GPS) is currently the dominant technology. Although it is widely deployed and used, since GPS is a multi-delay-based method, its accuracy is easily affected by the physical environment.

[0003] To address these issues, several research directions have emerged, including infrastructure-based positioning methods, collaborative positioning methods, sensor fusion and cellular system-based positioning methods, and machine learning-based methods. Despite these research efforts, existing methods still have many shortcomings. Most machine learning-based methods, while offering relatively high positioning accuracy, rely heavily on large amounts of data. Furthermore, in most industries, data exists in isolated silos, making data-integrated machine learning unable to guarantee data privacy. Secondly, positioning methods based on sensor fusion and infrastructure, while offering very high positioning accuracy, rely on network bandwidth and very expensive and abundant sensors, making them difficult to widely adopt. Existing collaborative vehicle positioning methods generally utilize federated learning, but the traditional federated learning approach discussed in their research suffers from low vehicle positioning accuracy.

[0004] Application Contents

[0005] This application provides a vehicle collaborative positioning method based on federated transfer learning, which can solve the problem of low vehicle positioning accuracy.

[0006] This application provides a vehicle collaborative localization method based on federated transfer learning, which includes:

[0007] The cloud clusters multiple regional states based on a multi-attribute spatiotemporal clustering algorithm to determine the regional state cluster corresponding to the target domain. Multiple regional states correspond one-to-one to multiple regions, and the target domain is the regional state corresponding to the target region in multiple regions.

[0008] The edge side aggregates the model weights corresponding to all regional states in the regional state cluster, and uses the self-attention mechanism module to obtain the spatio-temporal characteristics of all regional states in the regional state cluster;

[0009] The edge side filters the aggregated model weights using the spatio-temporal characteristics to obtain the final model weights, and constructs a global model using the final model weights;

[0010] The vehicle side within the target area receives the global model sent by the edge side, and based on federated learning, obtains a local model according to the global model; the local model is used to output the GPS area error based on the GPS data of the vehicle side;

[0011] The vehicle side corrects its own position information using the GPS area error.

[0012] Optionally, the f-th regional state S among multiple regional states f is:

[0013] S f =(id sf , e f , n f , t f , w f , num f , a f )

[0014] where f = 1, 2,..., L, L represents the total number of regional states, id sf represents the unique identifier of the f-th regional state, e f represents the longitude of the center of the region corresponding to the f-th regional state, n f represents the latitude of the center of the region corresponding to the f-th regional state, t f represents the time value, w f ∈{0, 1, 2, 3}, w f represents the weather condition of the f-th regional state at time t f , 0 represents non-severe weather, 1 represents heavy rain, 2 represents lightning, 3 represents heavy snow, num f represents the number of vehicles within the region corresponding to the f-th regional state, a f ∈{0, 1}, a f represents the GPS signal status of the f-th regional state, 0 represents unstable GPS signal status, 1 represents stable GPS signal status.

[0015] Optionally, the cloud side clusters multiple regional states based on the multi-attribute spatio-temporal clustering algorithm to determine the regional state cluster corresponding to the target domain, including:

[0016] The cloud calculates the similarity between the target domain and the other regional states among multiple regional states through a multi-attribute spatio-temporal similarity calculation formula;

[0017] The cloud uses the other regional states similar to the target domain, as well as the target domain, as the regional state cluster corresponding to the target domain.

[0018] Optionally, the cloud calculates the similarity between the target domain and the other regional states among multiple regional states through a multi-attribute spatio-temporal similarity calculation formula, including:

[0019] The cloud uses the formula:

[0020]

[0021] to calculate the similarity between the target domain S i and the other regional state S l ;

[0022] where (Sp_T_A) il represents the similarity between the target domain S i and the other regional state S l , (Sp_T_A) il =1 indicates that the target domain S i is similar to the other regional state S l , (Sp_T_A) il =0 indicates that the target domain S i is not similar to the other regional state S l , S i ∈(S1, S2,..., S L ), l = 1, 2,..., L, l ≠ i, L represents the total number of regional states, Sp il represents the result of whether the target domain S i is similar to the other regional state S l in the spatial dimension, T il represents the result of whether the target domain S i is similar to the other regional state S l in the temporal dimension, and A represents the result of whether the target domain S i is similar to the other regional state S l in other attributes.

[0023] Optionally, the judgment result Sp i of whether the target domain S l is similar to the other regional state S il in the spatial dimension is:

[0024]

[0025] where Spil = 1 indicates the target domain S i is spatially similar to other regional states S l , Sp il = 0 indicates the target domain S i is not spatially similar to other regional states S l , ΔSp il indicates the spatial distance between the target domain Sx and other regional states S l ; -1 indicates the target domain S i is completely negatively correlated with other regional states S l 0 indicates the target domain S i is spatially independent of other regional states S l 1 indicates the target domain S i is completely correlated with other regional states S l where x i = R * β i , x i represents the projection result of β i , R represents the radius of the Earth, β i represents the longitude of the center of the target area, y i represents the projection result of γ i , γ i represents the latitude of the center of the target area, x l = R * β l , x l represents the projection result of β l , β l represents the longitude of the center of the area corresponding to other regional states S l ; y l represents the projection result of γ l , γ l represents the latitude of the center of the area corresponding to other regional states S l ; spatial_threshold represents the spatial distance threshold;

[0026] The target domain S i and the judgment result T l on whether it is similar to other regional states S il in the time dimension is:

[0027]

[0028] where, T il = 1 indicates that the target domain S i is similar to other regional states S l in time, T il= 0 represents the target domain S i is not similar in time to other region states S l , ΔT il represents the target domain S i is not similar in time to other region states S l The distance in time from, through the formula ΔT il = |t i - t l | is calculated, temporal_threshold represents the time distance threshold, t i represents the target domain S i 's time value, t l represents other region states S l 's time value;

[0029] Target domain S i is whether similar in other attributes to other region states S l The judgment result A il is:

[0030]

[0031] Among them, A il = 1 represents that the target domain S i is similar in other attributes to other region states S l , A il = 0 represents that the target domain S i is not similar in other attributes to other region states S l , ΔA il represents the target domain S i is not similar in other attributes to other region states S l The distance in other attributes from, ΔA il _threshold represents the other attribute similarity threshold.

[0032] Optionally, the edge side aggregates the model weights corresponding to all region states in the region state cluster, including:

[0033] The edge side aggregates the model weights corresponding to the region state cluster to obtain the aggregated model weights

[0034] Among them, represents the total number of parameters in the aggregated model weights , is the j-th parameter of the aggregated model weights:

[0035]

[0036] Among them, W2 is the weight matrix, W2 ∈ RC×C , b2 is the bias term, b2 ∈ R C×C , R C×C represents the dimension;

[0037]

[0038] Among them, W1 is the weight matrix, W1 ∈ R C'×C , b1 is the bias term, b1 ∈ R C‘×C , R C'×C represents the dimension;

[0039]

[0040] Among them, i = 1, 2,..., k, k represents the number of regional states in the regional state cluster, w ij represents the j-th parameter corresponding to the i-th regional state in the regional state cluster, W0 is the convolution kernel, and b0 is the bias term.

[0041] Optionally, the self-attention mechanism module includes multiple self-attention units and a linear layer, and the output end of each self-attention unit is connected to the input end of the linear layer;

[0042] The edge side uses the self-attention mechanism module to obtain the spatio-temporal characteristics of all regional states in the regional state cluster, including:

[0043] The edge side inputs multiple regional states in the regional state cluster into the self-attention mechanism module and obtains the spatio-temporal characteristics of all regional states in the regional state cluster from the linear layer Multiple regional states correspond one-to-one with multiple self-attention units,

[0044] The expression of the self-attention unit is:

[0045]

[0046] Among them, Att TS (Q, K, V) represents the correlation degree between the regional state corresponding to the self-attention unit and other regional states in the regional state cluster, Q, K, and V represent the query input, key input, and value input respectively, d k represents the dimension of K, △T represents the time gap between two regional states, △S represents the space gap between two regional states, Softmax(·) is the normalization exponential function, tanh(·) is the Tanh activation function, Sig(·) is the Sigmoid activation function, represents the unique identifier of the spatio-temporal characteristics, represents the aggregated longitude, represents the aggregated latitude, represents the time value, Indicates the aggregated weather condition, 0 indicates non-severe weather, 1 indicates heavy rain, 2 indicates lightning, and 3 indicates heavy snow; Indicates the aggregated number of vehicles; Indicates the aggregated GPS signal status, 0 indicates unstable GPS signal status, and 1 indicates stable GPS signal status.

[0047] Optionally, the final model weights are

[0048] where w' j is the j-th parameter of the final model weights, indicates the total number of parameters in the model weights. The edge side obtains the parameter w' through the calculation formula where i = 1, 2,..., k, and k represents the number of regions, j is the j-th parameter of the second-layer linear layer, represents the element-wise product, represents the element-wise addition, and F represents the forget gate, and I i represents the input gate; i represents the input gate;

[0049]

[0050]

[0051] where S i is the target domain, is the spatio-temporal characteristic, Sig(·) is the Sigmoid activation function, and W if and W ii are both weight matrices of the target domain. W if ∈R 7×C and W ii ∈R 7×C and W sf and W si are both weight matrices of the spatio-temporal characteristic. W sf ∈R 7×C and W si ∈R 7×C and b f is the bias term of the forget gate. b f ∈R C and b i is the bias term of the input gate. b i ∈R C and W f is the parameter matrix of the forget gate. W f ∈R C and W iis the parameter matrix of the input gate, W i ∈R C , R 7×C and R C are both dimensions.

[0052] Optionally, when there are multiple target regions in multiple regions, the multiple target regions correspond to the multiple edge sides one by one.

[0053] The above solution of this application has the following beneficial effects:

[0054] In the embodiment of this application, the cloud determines the region state cluster corresponding to the target domain of the target region based on the multi-attribute spatio-temporal clustering algorithm. Then, the edge side aggregates the model weights corresponding to all region states in the region state cluster, uses the self-attention mechanism module to obtain the spatio-temporal characteristics of all region states in the region state cluster, and uses the spatio-temporal characteristics to filter the aggregated model weights to obtain the final model weights. Based on the final model weights, a global model is constructed. Finally, after the vehicle side in the target region receives the global model sent by the edge side, based on federated learning, a local model for outputting the GPS region error based on the GPS data of the vehicle side is obtained according to the global model, and the GPS region error is used to correct its own position information. Among them, since the global model is constructed based on the regions with similar spatio-temporal characteristics to the target region, the accuracy of the final model weights obtained through the region states of these regions and the model weights corresponding to the region states will increase. Therefore, the accuracy of the global model constructed using the final model weights will also increase. Because the region error within the same region and at similar time points is almost the same, the vehicle side can correct the regional error caused by the atmosphere, satellites, and multipath effects by revising its own position information according to the global model, thereby improving the positioning accuracy of the vehicle side.

[0055] Other beneficial effects of this application will be described in detail in the subsequent specific implementation part. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of this application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following drawings are only some embodiments of this application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0057] Figure 1 is the flowchart of the vehicle cooperative positioning method based on federated transfer learning provided by an embodiment of this application;

[0058] Figure 2 is the flowchart of the multi-attribute clustering algorithm provided by an embodiment of this application;

[0059] Figure 3 The flowchart for calculating the weights of the final model provided by an embodiment of the present application;

[0060] Figure 4 The framework diagram of the three-layer federated transfer learning provided by an embodiment of the present application. Detailed implementation manners

[0061] In the following description, specific details such as specific system structures and technologies are presented for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present application. However, those skilled in the art should clearly understand that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of the present application.

[0062] It should be understood that when used in the specification and appended claims of the present application, the term "comprising" indicates the presence of the described features, wholes, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0063] It should also be understood that the term "and / or" used in the specification and appended claims of the present application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0064] As used in the specification and appended claims of the present application, the term "if" can be interpreted as "when", "once", "in response to determining", or "in response to detecting" according to the context. Similarly, the phrase "if determined" or "if [the described condition or event] is detected" can be interpreted as meaning "once determined", "in response to determining", "once [the described condition or event] is detected", or "in response to detecting [the described condition or event]" according to the context.

[0065] In addition, in the description of the specification and appended claims of the present application, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0066] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that a specific feature, structure, or characteristic described in connection with that embodiment is included in one or more embodiments of this application. Thus, statements such as "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments", etc. that appear in different places in this specification do not necessarily all refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized. The terms "comprising", "including", "having" and their variants all mean "including but not limited to", unless otherwise specifically emphasized.

[0067] In view of the problem of low vehicle positioning accuracy, an embodiment of this application provides a vehicle collaborative positioning method based on federated transfer learning. This method determines, by the cloud, a regional state cluster corresponding to the target domain based on a multi-attribute spatio-temporal clustering algorithm, and then the edge side aggregates the model weights corresponding to all regional states in the regional state cluster, uses a self-attention mechanism module to obtain the spatio-temporal characteristics of all regional states in the regional state cluster, and uses the spatio-temporal characteristics to filter the aggregated model weights to obtain the final model weights. A global model is constructed based on the final model weights. Finally, after the vehicle terminal in the target area receives the global model sent by the edge side, based on federated learning, a local model for outputting GPS regional errors based on the GPS data of the vehicle terminal is obtained according to the global model, and the vehicle's own position information is corrected using the GPS regional errors. Among them, since the global model is constructed based on regions with similar spatio-temporal characteristics to the target area, the accuracy of the final model weights obtained through the regional states of these regions and the model weights corresponding to the regional states will increase. Therefore, the accuracy of the global model constructed using the final model weights will also increase. Because the regional errors within the same region and at similar time points are almost the same, the vehicle terminal revises its own position information according to the global model, which can correct the regional errors caused by the atmosphere, satellites, and multipath effects, and improve the positioning accuracy of the vehicle terminal.

[0068] Next, an exemplary description is given of the vehicle collaborative positioning method based on federated transfer learning provided by this application.

[0069] As Figure 1 shown, the vehicle collaborative positioning method based on federated transfer learning provided by an embodiment of this application includes the following steps:

[0070] Step 11, the cloud clusters multiple regional states based on a multi-attribute spatio-temporal clustering algorithm to determine a regional state cluster corresponding to the target domain.

[0071] The above-mentioned multiple regional states correspond one-to-one with multiple regions. The target domain is the regional state corresponding to the target region among the multiple regions. It can be understood that the target region can be one or multiple.

[0072] In some embodiments of the present application, the cloud can cluster multiple regional states through similarity calculation to determine the regional state cluster corresponding to the target domain. It is worth mentioning that by clustering multiple regional states, the amount of data is greatly reduced. In subsequent calculations, only the calculation object needs to be reduced from multiple regional states to the regional state cluster corresponding to the target domain, reducing the number of calculations and accelerating the model convergence speed.

[0073] Step 12, the edge side aggregates the model weights corresponding to all regional states in the regional state cluster and uses the self-attention mechanism module to obtain the spatio-temporal characteristics of all regional states in the regional state cluster.

[0074] In some embodiments of the present application, specifically, dynamic convolution can be used to aggregate the model weights corresponding to all regional states in the regional state cluster obtained by cloud clustering, and aggregate all regional states in the regional state cluster and capture their spatio-temporal characteristics to facilitate the calculations in subsequent steps.

[0075] It is worth mentioning that aggregating the model weights through dynamic convolution can increase the data accuracy of the model weights, and using the self-attention mechanism module to obtain the spatio-temporal characteristics of all regional states in the regional state cluster can increase the accuracy of the obtained spatio-temporal characteristics.

[0076] Step 13, the edge side filters the aggregated model weights using the spatio-temporal characteristics to obtain the final model weights, and constructs a global model using the final model weights.

[0077] In some embodiments of the present application, specifically, the aggregated model weights can be filtered based on the gating mechanism of the forgetting gate and the input gate to filter out the useless features of the model weights and retain the important features, so as to further increase the data accuracy of the model weights, thereby improving the accuracy of the constructed global model.

[0078] Step 14, the vehicle terminal in the target region receives the global model sent by the edge side and obtains a local model based on the global model according to federated learning.

[0079] The above local model is used to output the GPS regional error based on the GPS data of the vehicle terminal. It can be understood that the above global model is also a model for outputting the GPS regional error. Exemplarily, this model can be a GPS error prediction model.

[0080] It should be noted that the position of the satellite, the influence of the ionosphere and the atmosphere on the signal radio wave, and the relevant deviations of the receiver and the antenna are all the main sources of GPS measurement errors. In some embodiments of the present application, according to the nature of the errors, these errors are divided into two types: systematic errors and random errors. Therefore, the GPS error can be expressed as:

[0081] E = E r + E s

[0082] wherein, E s represents the systematic error, and E r represents the random error. The systematic error mainly comes from the position and clock error of the satellite, the ionospheric delay and the influence of the atmosphere on the signal radio wave, as well as the receiver clock and positioning error. The random error mainly includes the error caused by the multipath effect and the receiver noise. Usually, the random error is much smaller than the systematic error. Among the above errors, the errors caused by the satellite position, the ionosphere, the atmosphere and the multipath effect are also called regional errors, and the regional errors have the characteristic that the errors in the same region and at similar time points are almost the same. Therefore, the GPS error can be expressed as:

[0083] E = E r + E p

[0084] wherein, E p is the individual error, and the individual error is mainly the error generated by the receiver and the like.

[0085] For two vehicles V m and V n in the same time and the same region (the diameter of the region does not exceed 2 km), their relative error ||ΔE mn || can be expressed as:

[0086]

[0087] wherein, represents the GPS position of vehicle m, represents the accurate position of vehicle m, and E m represents the GPS positioning error of vehicle m, represents the GPS position of vehicle n, represents the accurate position of vehicle n, and E n represents the GPS positioning error of vehicle n. According to the above formula, it can be obtained that:

[0088]

[0089] wherein, represents the GPS positioning regional error of vehicle m, Represents the individual differences of vehicle m, The GPS positioning area error of vehicle n, The individual differences of vehicle n. As above, the GPS positioning area errors between two vehicles in the same area at the same time are almost equal, that is Therefore, it can be known that the relative error between vehicles m and n in the same area at the same time depends on the individual differences:

[0090]

[0091] Through the above analysis, it can be known that the GPS position information of the vehicle is inaccurate and has a certain error.

[0092] The error caused by the atmosphere, satellites and multipath effects is called the GPS area error, which is expressed as:

[0093]

[0094] Among them, Represents the longitude positioning error of the vehicle, Represents the latitude positioning error of the vehicle.

[0095] It should be noted that according to the spatio-temporal characteristics that the errors of the above area errors in the same area and at similar time points are almost the same, the vehicle cooperative positioning method provided by the present application can predict the GPS area error to cooperate with vehicle positioning, and the GPS area error output by the local model based on the GPS data at the vehicle end is the GPS area error used for prediction.

[0096] In some embodiments of the present application, the vehicle end can obtain the local model based on the traditional federated learning according to the global model. Specifically, after the edge side obtains the final model weights, it uses the final model weights as the initial weights of the global model, and then the global model is updated and the model weights are sent to the vehicle end. The vehicle end uses the model weights to construct the local model and trains the local model with local data (such as GPS data). After training, the corresponding model weights are uploaded to the edge side. The edge side aggregates and filters the model weights, and outputs the final model weights again and updates the global model until the local model converges.

[0097] It should be noted that each time the global model is updated, all parameters on the edge side will be updated.

[0098] Step 15, the vehicle end corrects its own position information using the GPS area error.

[0099] In the related art, the vehicle terminal can be positioned based on traditional vehicle positioning methods. However, the regional errors generated by traditional positioning methods are high, resulting in inaccurate positioning. In some embodiments of the present application, the vehicle is positioned by the collaborative positioning method provided in the present application, which can effectively correct the GPS regional errors caused by the atmosphere, satellites, and multipath effects, thereby improving the vehicle positioning accuracy.

[0100] It is worth mentioning that since the global model is constructed based on regions with similar spatio-temporal characteristics to the target region, the accuracy of the final model weights obtained through the regional states of these regions and the model weights corresponding to the regional states will increase. Therefore, the accuracy of the global model constructed using the final model weights will also be improved. Because the regional errors within the same region and at similar time points are almost the same, the vehicle terminal can revise its own position information according to the global model, which can correct the regional errors caused by the atmosphere, satellites, and multipath effects and improve the positioning accuracy of the vehicle terminal.

[0101] The following specifically describes the specific steps of step 11 in an exemplary manner in combination with specific embodiments.

[0102] In some embodiments of the present application, the specific implementation process of the above step 11 includes the following steps:

[0103] Step 11.1, define the regional state. A regional state is defined for each region to represent the dynamic changes of each region. The f-th regional state S among multiple regional states f is:

[0104] S f =(id sf ,e f ,n f ,t f ,w f ,num f ,a f )

[0105] where f = 1, 2,..., L, and L represents the total number of regional states. id sf represents the unique identifier of the f-th regional state, e f represents the longitude of the center of the region corresponding to the f-th regional state, n f represents the latitude of the center of the region corresponding to the f-th regional state, t f represents the time value, w f ∈{0, 1, 2, 3} represents the weather condition of the f-th regional state at time t f , 0 represents non-severe weather, 1 represents heavy rain, 2 represents lightning, 3 represents heavy snow, num f represents the number of vehicles within the region corresponding to the f-th regional state, af ∈ {0, 1}, a f represents the GPS signal status of the f-th area status, 0 indicates that the GPS signal status is unstable, and 1 indicates that the GPS signal status is stable.

[0106] In some embodiments of the present application, the model weights corresponding to the above area status are recorded as W = {W1,..., W p ,..., W L}, p = 1, 2,..., L, where each W i is a list of length , that is w ij is the j-th parameter in the i-th list, is the total number of parameters of the list. The above area status and the model weights corresponding to the area status are stored in the cloud in real time.

[0107] Step 11.2, the cloud calculates the similarity between the target domain and other area statuses among multiple area statuses through the multi-attribute spatio-temporal similarity calculation formula.

[0108] The multi-attribute spatio-temporal clustering algorithm provided by the present application is a density-based supervised learning algorithm that clusters the source domains with similar spatio-temporal and other attributes centered on the target domain. There are 4 clustering parameters to be set: spatial distance threshold spatial_threshold, temporal distance threshold temporal_threshold, other attribute similarity threshold ΔA_threshold, and spatio-temporal object quantity threshold MinPts. The first 3 parameters are used to determine the multi-attribute proximity domain, and the last parameter is used to determine the number of objects in the spatio-temporal proximity domain. First, set the thresholds of each distance, and use the above multiple area statuses as the clustering objects.

[0109] In some embodiments of the present application, the target domain is represented as S i , and other area statuses are represented as S l . Select the target domain S i from all area statuses, and judge whether the target domain S i is a core object. If not, reselect the object. If so, perform the following similarity calculation:

[0110] The cloud passes the formula:

[0111]

[0112] to calculate the similarity between the target domain S i and other area statuses S l ;

[0113] Among them, (Sp_T_A)il Denote the target domain as S i The similarity between the target domain S and other regional states S l , (Sp_T_A) il =1 indicates that the target domain S i is similar to other regional states S l , (Sp_T_A) il =0 indicates that the target domain S i is not similar to other regional states S l , S i ∈(S1, S2,..., S L ), l = 1, 2,..., L, l ≠ i, where L represents the total number of regional states, and Sp il denotes the result of whether the target domain S i is similar to other regional states S l in the spatial dimension, T il denotes the result of whether the target domain S i is similar to other regional states S l in the temporal dimension, and A represents the result of whether the target domain S i is similar to other regional states S l in other attributes.

[0114] Among them, for the judgment result Sp of whether the target domain S i is similar to other regional states S l in the spatial dimension: il is:

[0115]

[0116] Among them, Sp il =1 indicates that the target domain S i is similar to other regional states S l in space, and Sp il =0 indicates that the target domain S i is not similar to other regional states S l in space. ΔSp il denotes the distance between the target domain S i and other regional states S l in space. -1 indicates that the target domain S i is completely negatively correlated with other regional states S l in space, 0 indicates that the target domain S i is independent of other regional states S l in space, and 1 indicates that the target domain S i is completely correlated with other regional states S l in space, where x i =R*βi , x i represents the projection result of β i . R represents the radius of the Earth, and β i represents the longitude of the center of the target area, y i represents the projection result of γ i , and γ i represents the latitude of the center of the target area. x l = R * β l , x l represents the projection result of β l , and β l represents the longitude of the center of the area corresponding to other area status S l , y l represents the projection result of γ l , and γ l represents the latitude of the center of the area corresponding to other area status S l . spatial_threshold represents the spatial distance threshold.

[0117] Search for all other area statuses S i adjacent to the target domain S l through the above formula. If S l does not belong to any existing area status cluster, then calculate whether other attributes between S i and S l are similar through the following algorithm.

[0118] The judgment result T i on whether the target domain S l is similar to other area status S il in the time dimension is as follows:

[0119]

[0120] Among them, T il = 1 indicates that the target domain S i is similar to other area status S l in time. T il = 0 indicates that the target domain S i is not similar to other area status S l in time. ΔT il represents the time distance between the target domain S i and other area status S l . It is calculated through the formula ΔT il = |t i - t l |. temporal_threshold represents the time distance threshold, and t iRepresents the target domain S i The time value, t l Represents the state of other areas S l The time value.

[0121] Target domain S i And the state of other areas S l The judgment result A on whether they are similar in other attributes il Is:[[]]

[0122]

[0123] Among them, other attributes include: weather conditions, number of vehicles, GPS signal status, A il A = 1 indicates that the target domain S i Is similar to the state of other areas S l In other attributes, A il A = 0 indicates that the target domain S i Is not similar to the state of other areas S l In other attributes, ΔA il Represents the distance between the target domain S o And the state of other areas S l In other attributes, ΔA il _threshold represents the similarity threshold of other attributes, and the Dice similarity coefficient is used to calculate ΔA il , and the calculation formula is:[[]]

[0124]

[0125] Among them, A i =(w i , num i , a o ), A l =(w l , num l , a l ), w i ∈{0, 1, 2, 3}, w i Represents the weather condition of the i-th area state at time t i Time, t i Represents the time value, 0 means non-severe weather, 1 means heavy rain, 2 means lightning, 3 means heavy snow, num i Represents the number of vehicles in the area corresponding to the i-th area state, a i ∈{0, 1}, a i Represents the GPS signal status of the i-th area state, where 0 means the GPS signal status is unstable, 1 means the GPS signal status is stable; w l ∈{0, 1, 2, 3}, w lIndicates the weather condition of the l-th area state at time t l where t l represents the time value, num l represents the number of vehicles in the area corresponding to the l-th area state, a l ∈ {0, 1}, a l represents the GPS signal status of the l-th area state, where 0 indicates unstable GPS signal status and 1 indicates stable GPS signal status.

[0126] The formula for calculating the spatio-temporal object quantity threshold is

[0127] MinPts = ln(|D s |)

[0128] where D s is the set of area states, and |D s | is the total number of area states.

[0129] Step 11.3: The cloud will take other area states similar to the target domain, as well as the target domain, as the area state cluster corresponding to the target domain.

[0130] Specifically, after the above calculation, if S l is similar to S i , then put S l into the newly created cluster, and select the next area state for similarity calculation until all spatially adjacent and similar area states of the target domain S i have been put into the newly created cluster, and take the newly created cluster as the area state cluster S' = {S′1,..., S′ i ,..., S′ k} corresponding to the target domain, and the model weights W' = {W′1,..., W′ i ,..., W′ k} corresponding to the area state cluster are sent to the edge side together.

[0131] The above steps will be exemplarily described below with a specific embodiment.

[0132] The flowchart of the above multi-attribute spatio-temporal clustering algorithm is as Figure 2 [[ID=5'5]]shown. After starting, establish the area state set D S , then set four thresholds: spatial distance threshold, time distance threshold, other attribute similarity threshold, spatio-temporal object quantity threshold, and then select the target domain S S from D i . Next, determine whether S i belongs to an existing cluster. If so, return to the previous step to re-select S i ; if not, then determine the object point (with the same meaning as the above area state) S iWhether it is a core object. If not, return to re-select S i ; If so, start searching for S i 's spatially adjacent point Q i (i.e., the above-mentioned other area state S l ), and then judge whether Q i belongs to the existing cluster. If so, return to the previous step to re-search for Q i ; If not, then continue to judge S i and Q i 's time attributes and other attributes are similar. If not, re-search for Q i ; If so, put Q i into the new cluster, and then judge whether the objects in each cluster are core objects. If not, re-search for Q i ; If so, take all the area states in the same cluster as the target domain S i as the output, and the algorithm ends.

[0133] It is worth mentioning that the cloud clusters multiple area states, reducing the calculation objects from multiple area states to the area state clusters corresponding to the target domain, reducing the data volume, being able to reduce the number of calculations, and accelerating the positioning process.

[0134] The following will exemplarily illustrate the specific steps of the above step 12 in combination with specific embodiments.

[0135] In some embodiments of the present application, the specific implementation process of the above step 12 includes the following steps:

[0136] Step 12.1, the edge side inputs the model weights W’ = {W′1,..., W′ i ,..., W′ k} corresponding to the area state cluster into the dynamic convolution module for aggregation, and performs dimensional conversion through the linear layer to obtain the aggregated model weights.

[0137] The edge side aggregates the model weights corresponding to the area state cluster to obtain the aggregated model weights

[0138] Among them, represents the total number of parameters in the aggregated model weights , is the j-th parameter of the aggregated model weights:

[0139]

[0140] Among them, W2 is the weight matrix, W2 ∈ R C×C , b2 is the bias term, b2 ∈ R C×C , R C×CRepresents the dimension;

[0141]

[0142] Among them, W1 is the weight matrix, W1 ∈ R C'×C , b1 is the bias term, b1 ∈ R C‘×C , R C'×C Represents the dimension;

[0143]

[0144] Among them, i = 1, 2,..., k, k represents the number of regional states in the regional state cluster, w ij represents the j-th parameter corresponding to the i-th regional state in the regional state cluster, W0 is the convolution kernel, and b0 is the bias term.

[0145] For the above model weights W’ = {W'1,…,W' i ,…,W' k}, the aggregation function is a linear combination, that is, in the above steps, the w ij of each group is successively subjected to convolution operations, and the dimension of the model weights after convolution operations is converted by using two linear layers to obtain the aggregated model weights. Among them, the above two linear layers include the first linear layer and the second linear layer. The calculation formula of the first linear layer is: The calculation formula of the second linear layer is:

[0146] It should be noted that when the cloud selects regional states using the multi-attribute spatio-temporal clustering algorithm, the number of regional states selected each time is uncertain, resulting in an uncertain number of model weights corresponding to the regional states received by the edge side. However, the number of convolution kernels must be determined. To solve this problem, in some embodiments of the present application, max_filter convolution kernels are initialized in advance. When the number of model weights is insufficient, they are all replaced with zero matrices to dynamically adapt to the number of model weights. The calculation of max_filter is max_filter = |D s |, where D s represents the total number of cloud regional states.

[0147] It is worth mentioning that by aggregating the model weights through the above steps to obtain the aggregated model weights, it is for the convenience of subsequent step calculations, and aggregating the model weights using dynamic convolution can improve the data accuracy of the model weights.

[0148] Step 12.2, the edge side uses the self-attention mechanism module to obtain the spatio-temporal characteristics of all regional states in the regional state cluster.

[0149] In some embodiments of the present application, the self-attention mechanism module includes a plurality of self-attention units and a linear layer, and the output end of each self-attention unit is connected to the input end of the linear layer.

[0150] The edge side inputs multiple regional states in the regional state cluster into the self-attention mechanism module, calculates the correlation degree between each regional state and all other regional states, and the multiple regional states correspond to the multiple self-attention units one by one;

[0151] The expression of the self-attention unit is:

[0152]

[0153] where Att TS (Q, K, V) represents the correlation degree between the regional state corresponding to the self-attention unit and other regional states in the regional state cluster, Q, K, and V respectively represent the query input, key input, and value input, and d k represents the dimension of K, △T represents the time gap between two regional states, △S represents the space gap between two regional states, Softmax(·) is the normalized exponential function, tanh(·) is the Tanh activation function, Sig(·) is the Sigmoid activation function, and the output value is [0, 1]. represents the unique identifier of the spatio-temporal characteristics represents the longitude after aggregating the longitudes of the centers of the regions corresponding to all regional states in the regional state cluster, represents the latitude after aggregating the latitudes of the centers of the regions corresponding to all regional states in the regional state cluster, represents the time value, represents the weather condition after aggregating the weather conditions of the regions corresponding to all regional states in the regional state cluster, 0 represents non-severe weather, 1 represents heavy rain, 2 represents lightning, and 3 represents heavy snow; represents the number of vehicles after aggregating the number of vehicles in the regions corresponding to all regional states in the regional state cluster; represents the GPS signal status after aggregating the GPS signal statuses of the regions corresponding to all regional states in the regional state cluster, 0 represents unstable GPS signal status, and 1 represents stable GPS signal status.

[0154] The above query input Q = W Q S j The key input K = W K S j The value input V = W V S j W Q W K WK are the weight matrices of Q, K, and V respectively.

[0155] The calculated correlation is input into the linear layer, and the dimension is transformed by the linear layer to obtain the spatiotemporal characteristics of all regional states.

[0156] The above steps are exemplarily described below with reference to a specific embodiment.

[0157] The corresponding region state S of a self-attention unit in the self-attention mechanism module r , regional status S r With regional status S r+1 The time difference between r,r+1 for:

[0158] △T r,r+1 =|t r -t r+1 |

[0159] Where r = 1, 2, ..., R, R is the total number of regional states in the regional state cluster, and regional state S r With regional status S r+1 Each is a regional state in the regional state cluster.

[0160] Regional status S r With regional status S r+1 The spatial gap between r,r+1 for:

[0161]

[0162] Among them, e r Indicates the regional status S r The longitude of the center of the corresponding area, e r+1 Indicates the regional status S r+1 The longitude of the center of the corresponding area, n r Indicates the regional status S r The latitude of the center of the corresponding area, n r+1 Indicates the regional status S r+1 The latitude of the center of the corresponding region.

[0163] The regional state S is calculated by the above formula r With regional status S r+1 The gap in time and space, and the regional state S is obtained through the self-attention unit r With regional status S r+1 Similarly, calculate the regional state S rThe degree of association between each regional state and all other regional states in the regional state cluster is calculated by each self-attention unit corresponding to the regional state, and the degree of association between each regional state and all other regional states in the regional state cluster is obtained, and these degrees of association are output to the linear layer, and the spatio-temporal characteristics of all regional states are obtained through dimensional conversion by the linear layer.

[0164] It is worth mentioning that the self-attention mechanism is a calculation with the same input dimension and output dimension, and the output data is not conducive to subsequent calculations. Therefore, dimensional conversion is performed through the linear layer to obtain spatio-temporal characteristics that are convenient for subsequent calculations, solving the problem that the data is not conducive to subsequent calculations. Using the above self-attention mechanism module to obtain the spatio-temporal characteristics of all regional states in the regional state cluster can make the obtained spatio-temporal characteristics more accurate.

[0165] The following specifically illustrates the specific steps of the above step 13 in combination with specific embodiments.

[0166] In some embodiments of the present application, the specific implementation process of the above step 13 includes the following steps:

[0167] In the first step, the edge side filters the aggregated model weights using spatio-temporal characteristics to obtain the final model weights.

[0168] In some embodiments of the present application, a gating unit is set on the edge side, and the gating unit filters the parameters in the aggregated model weights in remembers the important features of the parameters and forgets the redundant features to obtain the final model weights

[0169] The edge side obtains the parameter w' through the calculation formula , where w' j is the j-th parameter of the final model weight, j represents the number of parameters in the model weight, i = 1, 2,..., k, k represents the number of regions, is the j-th parameter of for , represents the element-wise product, represents the element-wise addition, F i represents the forget gate, and I i represents the input gate:

[0170]

[0171]

[0172] Among them, Si is the target domain, is the spatio-temporal characteristic, Sig(·) is the Sigmoid activation function, W if and W ii are both weight matrices of the target domain, W if ∈R 7×C , W ii ∈R 7×C , W sf and W si are both weight matrices of the spatio-temporal characteristic, W sf ∈R 7×C , W si ∈R 7×C , b f is the bias term of the forget gate, b f ∈R C , b i is the bias term of the input gate, b i ∈R C , W f is the parameter matrix of the forget gate, W f ∈R C , W i is the parameter matrix of the input gate, W i ∈R C , R 7×C and R C are both dimensions.

[0173] It should be noted that the above forget gate F i is used to control which features in i need to be forgotten and which need to be remembered; the input gate I i is used to selectively remember the target domain S i and record some important features in the target domain S

[0174] Next, a specific example is used to exemplarily illustrate the calculation process of the weights of the above final model.

[0175] As Figure 3 shown, the calculation process of the weights of the final model is as follows: the weights of the model are input into the dynamic convolution module and undergo dimensional transformation through the first linear layer and the second linear layer to obtain the aggregated weights of the model. At the same time, the regional state cluster is input into the self-attention mechanism module to obtain the spatio-temporal characteristics. The aggregated weights of the model and the spatio-temporal characteristics are input into the gated unit together. After being filtered by the gated unit, the final weights of the model are output.

[0176] In the second step, the edge side constructs a global model using the final weights of the model obtained in the above step.

[0177] It is worth mentioning that the model weights are filtered to filter out the useless features of the model weights and retain the important features, so as to further improve the data accuracy of the model weights, thereby enhancing the accuracy of the constructed global model.

[0178] The following takes a specific example to exemplarily illustrate the above vehicle collaborative positioning method based on federated transfer learning.

[0179] As Figure 4 shown, the three-layer federated transfer learning framework specifically includes the cloud, the edge side, and the local side. The cloud is connected to multiple edge sides, and from edge side 1 to edge side A are all connected to the cloud, where A represents the number of the last edge side; each edge side is connected to multiple local sides, and each edge side is connected to vehicle side 1 to vehicle side m, where m represents the number of the last vehicle side. Among them, when there are multiple target areas in multiple regions, the multiple target areas correspond to the multiple edge sides one by one. That is, one edge side corresponds to one target area.

[0180] Such a three-layer federated transfer learning framework can implement the vehicle collaborative positioning method provided in this application, correct the regional errors caused by the atmosphere, satellites, and multipath effects, and improve the positioning accuracy of the vehicle side.

[0181] The above is the preferred implementation manner of this application. It should be noted that for those of ordinary skill in the art, without departing from the principle described in this application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of this application.

Claims

1. A vehicle cooperative positioning method based on federated transfer learning, characterized in that Including: The cloud clusters the states of multiple regions based on a multi-attribute spatio-temporal clustering algorithm to determine the region state cluster corresponding to the target domain; The multiple region states correspond to multiple regions one by one, and the target domain is the region state corresponding to the target region among the multiple regions; The edge side aggregates the model weights corresponding to all region states in the region state cluster and uses the self-attention mechanism module to obtain the spatio-temporal characteristics of all region states in the region state cluster; The edge side filters the aggregated model weights using the spatio-temporal characteristics to obtain the final model weights, and constructs a global model using the final model weights; The vehicle terminal in the target region receives the global model sent by the edge side and, based on federated learning, obtains a local model according to the global model; the local model is used to output a GPS region error based on the GPS data of the vehicle terminal; The vehicle terminal corrects its own position information using the GPS region error.

2. The vehicle cooperative positioning method according to claim 1, wherein The f-th region state S among the multiple region states f is as follows: S f =(id sf ,e f ,n f ,t f ,w f ,nu m f,a f ) where f = 1, 2,..., L, L represents the total number of the regional states, id sf represents the unique identifier of the f-th regional state, e f represents the longitude of the center of the region corresponding to the f-th regional state, n f represents the latitude of the center of the region corresponding to the f-th regional state, t f represents the time value, w f ∈ {0, 1, 2, 3}, w f represents the weather condition of the f-th regional state at t f time, 0 represents non-severe weather, 1 represents heavy rain, 2 represents lightning, 3 represents heavy snow, num f represents the number of vehicles in the region corresponding to the f-th regional state, a f ∈ {0, 1}, a f represents the GPS signal status of the f-th regional state, 0 represents unstable GPS signal status, 1 represents stable GPS signal status.

3. The vehicle collaborative positioning method according to claim 1, wherein The cloud clusters the states of multiple regions based on a multi-attribute spatio-temporal clustering algorithm to determine the region state cluster corresponding to the target domain, including: The cloud calculates the similarity between the target domain and other region states among the multiple region states through a multi-attribute spatio-temporal similarity calculation formula; The cloud takes the other region states similar to the target domain and the target domain as the region state cluster corresponding to the target domain.

4. The vehicle cooperative positioning method according to claim 3, characterized in that, The cloud calculates the similarity between the target domain and other region states among the multiple region states through a multi-attribute spatio-temporal similarity calculation formula, including: The cloud uses the formula: Calculate the target domain S i and the other area state S l to calculate the similarity between them; Among them, (Sp_T_A) il represents the similarity between the target domain S i and the other region state S l . (Sp_T_A) il = 1 indicates that the target domain S i is similar to the other region state S l . (Sp_T_A) il = 0 indicates that the target domain S i is not similar to the other region state S i . S i ∈(S1, S2,..., S L ), l = 1, 2,..., L, l ≠ i, where L represents the total number of region states. Sp il represents the result of whether the target domain S i is similar to the other region state S l in the spatial dimension. T il represents the result of whether the target domain S i is similar to the other region state S l in the temporal dimension. A represents the result of whether the target domain S i is similar to the other region state S l in other attributes.

5. The vehicle cooperative positioning method according to claim 4, wherein The target domain S i and the state of the other regions S l The judgment result Sp on whether they are similar in the spatial dimension il is as follows: Among them, Sp il = 1 indicates that the target domain S i is spatially similar to the other region state S l , Sp il = 0 indicates that the target domain S i is not spatially similar to the other region state S l , ΔSp il represents the spatial distance between the target domain S i and the other region state S l . ΔSp il ∈{-1, 0, 1}, -1 indicates that the target domain S i is completely negatively correlated spatially with the other region state S l , 0 indicates that the target domain S i is spatially independent of the other region state S l , 1 indicates that the target domain S i is completely correlated spatially with the other region state S l , where x i = R * β i , x i represents the projection result of β i , R represents the radius of the Earth, β i represents the longitude of the center of the target area, y i represents the projection result of γ i , γ i represents the latitude of the center of the target area, x l = R * β l , x l represents the projection result of β l , β l represents the longitude of the center of the area corresponding to the other region state S l , y l represents the projection result of γ l , γ l represents the latitude of the center of the area corresponding to the other region state S l ; spatial_threshold represents the spatial distance threshold; the target domain S i and the state of the other region S l the judgment result T on whether they are similar in the time dimension il is as follows: Among them, T il = 1 indicates that the target domain S i is similar to the other region state S l in terms of time, T il = 0 indicates that the target domain S i is not similar to the other region state S i in terms of time, ΔT il represents the time distance between the target domain S i and the other region state S l in terms of time, ΔT il = |t i - t l |, temporal_threshold represents the time distance threshold, t i represents the time value of the target domain S i and t l represents the time value of the other region state S l ; The target domain S i and the determination result A l on whether it is similar to the other area state S il is as follows: Among them, A il = 1 indicates that the target domain S i is similar to the other regional status S l in other attributes, A il = 0 indicates that the target domain S i is not similar to the other regional status S l in other attributes, ΔA il represents the distance between the target domain S i and the other regional status S l in other attributes, ΔA il _threshold represents the similarity threshold of other attributes.

6. The vehicle cooperative positioning method according to claim 1, wherein, The edge side aggregates the model weights corresponding to all region states in the region state cluster, including: The edge side aggregates the model weights corresponding to the area state clusters to obtain the aggregated model weights Among them, represents the total number of parameters in the aggregated model weights and is the j-th parameter of the aggregated model weights: Among them, W2 is the weight matrix, W2 ∈ R C×C , b2 is the bias term, b2 ∈ R C×C , R C×C represents the dimension; Among them, W1 is the weight matrix, W1 ∈ R C'×C , b1 is the bias term, b1 ∈ R C‘×C , R C'×C represents the dimension; where \(i = 1, 2, \cdots, k\), \(k\) represents the number of regional states in the regional state cluster, \(w\) ij represents the \(j\)-th parameter corresponding to the \(i\)-th regional state in the regional state cluster, \(W_0\) is the convolutional kernel, and \(b_0\) is the bias term.

7. The vehicle collaborative positioning method according to claim 1, characterized in that The self-attention mechanism module includes multiple self-attention units and a linear layer, and the output end of each self-attention unit is connected to the input end of the linear layer; The edge side uses the self-attention mechanism module to obtain the spatio-temporal characteristics of all region states in the region state cluster, including: The edge side inputs multiple regional states in the regional state cluster into the self-attention mechanism module and obtains the spatio-temporal characteristics of all regional states in the regional state cluster from the linear layer. The multiple regional states correspond one-to-one with multiple self-attention units. The expression of the self-attention unit is: Among them, Att TS (Q, K, V) represents the correlation degree between the region state corresponding to the self-attention unit and other region states in the region state cluster. Q, K, and V respectively represent the query input, key input, and value input, d k represents the dimension of K, △T represents the time gap between two region states, △S represents the space gap between two region states, Softmax(·) is the normalization exponential function, tanh(·) is the Tanh activation function, and Sig(·) is the Sigmoid activation function. represents the unique identifier of the spatio-temporal characteristics, represents the aggregated longitude, represents the aggregated latitude, represents the time value, represents the aggregated weather condition. 0 represents non-severe weather, 1 represents heavy rain, 2 represents lightning, and 3 represents heavy snow; represents the aggregated number of vehicles; represents the aggregated GPS signal state. 0 represents unstable GPS signal state, and 1 represents stable GPS signal state.

8. The vehicle cooperative positioning method according to claim 1, characterized in that, The weights of the final model are where w' j is the j-th parameter of the weights of the final model, represents the total number of parameters in the model weights, and the edge side obtains the parameter w' through the calculation formula , i = 1, 2,..., k, where k represents the number of regions, j is the j-th parameter of the second-layer linear layer, represents the element-wise product, represents the element-wise addition, F represents the forget gate, and I i represents the input gate; i ​ Among them, S i is the target domain, is the spatio-temporal characteristic, Sig(·) is the Sigmoid activation function, W if and W ii are both weight matrices of the target domain, W if ∈R 7×C and W ii ∈R 7×C , W sf and W si are both weight matrices of the spatio-temporal characteristic, W sf ∈R 7 ×C , W si ∈R 7×C , b f is the bias term of the forget gate, b f ∈R C , b i is the bias term of the input gate, b i ∈R C , W f is the parameter matrix of the forget gate, W f ∈R C , W i is the parameter matrix of the input gate, W i ∈R C , R 7×C and R C are both dimensions.

9. The vehicle collaborative positioning method according to claim 1, wherein When there are multiple target regions among the multiple regions, the multiple target regions correspond to multiple edge sides one by one.

Images