Multi-UAV path planning method and system based on collaborative tasks

Abstract

The present invention provides a multi-UAV path planning method and system based on collaborative tasks, which relates to the technical field of UAV path planning. First, the global UAV data is determined and embedded into the system to form a UAV embedding vector. Second, the UAV embedding vector is deeply mined to form a UAV depth vector. Finally, based on the UAV depth vector and the operation area vectors of the unfinished operation areas of the multiple UAVs, the target position of each UAV is generated. This method can improve the relatively low reliability of multi-UAV path planning in existing technologies.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) path planning, and in particular to a multi-UAV path planning method and system based on collaborative tasks. Background Art

[0002] The goal of multi-UAV collaborative path planning is typically to use multiple UAVs to efficiently complete tasks in a target area, such as crop growth monitoring, precision spraying of pesticides or fertilizers, and soil analysis. Each UAV needs to perform tasks in a complex environment and collaborate with other UAVs to ensure that the tasks are completed efficiently and without conflict. However, in the prior art, the corresponding staff generally assigns an operating area to each UAV based on their experience, and then each UAV performs the corresponding operation in the assigned area. However, the inventors' research has found that, on the one hand, due to the limitations of the staff's experience, and on the other hand, because UAVs may be interfered with by other factors during the operation process, operating according to the path plan formed before the operation may not meet the actual operation requirements. In other words, in the prior art, there is a problem of relatively low reliability of multi-UAV path planning. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a multi-UAV path planning method and system based on collaborative tasks to improve the problem of relatively low reliability of multi-UAV path planning in the prior art.

[0004] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0005] A multi-UAV path planning method based on collaborative tasks, comprising:

[0006] Determining drone global data and embedding the drone global data to form a drone embedding vector, wherein the drone global data is used to reflect at least the current position, current power level, and current material reserve level of each of the multiple drones during operation, and the drone embedding vector is used to reflect shallow semantic information of the drone global data;

[0007] Deep mining the drone embedding vector to form a drone depth vector, wherein the drone depth vector is used to reflect deep semantic information of the drone global data;

[0008] Based on the drone depth vector and the operation area vectors of the unfinished operation areas of the multiple drones, a target position of each drone is generated, wherein each target position belongs to the unfinished operation area, and the target position serves as a moving target for the drone to move from the current position to complete path planning, and the drone completes the corresponding operation task based on the stored materials in the process of moving from the current position to the target position.

[0009] In some preferred embodiments, in the above-mentioned multi-UAV path planning method based on collaborative tasks, the step of performing deep mining on the UAV embedding vector to form a UAV depth vector includes:

[0010] In a first mapping subunit included in the semantic encoding unit, the drone embedding vector is subjected to a first mapping operation to form an initial depth vector, the initial depth vector is subjected to a second mapping operation to form an intermediate depth vector, and the intermediate depth vector is subjected to a third mapping operation to form an encoding vector of the first mapping subunit, wherein the first mapping operation, the second mapping operation, and the third mapping operation are different from each other;

[0011] Loading the encoding vector of the previous mapping subunit into the next mapping subunit to perform multiple levels of mapping operations until the encoding vector of the last mapping subunit is formed, and using the encoding vector as the drone encoding vector of the semantic encoding unit, wherein the semantic encoding unit is configured with multiple mapping subunits connected in sequence;

[0012] A semantic decoding unit is used to perform a semantic decoding operation on the drone encoding vector of the semantic encoding unit to form a drone decoding vector, and a drone depth vector is obtained based on the drone decoding vector, wherein the semantic decoding unit is configured with multiple mapping subunits connected in sequence.

[0013] In some preferred embodiments, in the above-mentioned multi-UAV path planning method based on collaborative tasks, the step of performing a first mapping operation on the UAV embedding vector to form an initial depth vector includes:

[0014] Mapping the vector parameters in the drone embedding vector to a target interval to form a drone normalized vector, wherein the lower limit value of the target interval is equal to 0 and the upper limit value is equal to 1;

[0015] Performing a first mapping operation on the normalized vector of the UAV to form a first mapping vector, wherein the first mapping operation includes a plurality of mapping stages, a link with a jump connection is configured between two adjacent mapping stages, and a transmission control parameter is set on the link, wherein the transmission control parameter is used to reflect whether the output of the previous mapping stage is connected to the output of the next mapping stage;

[0016] Perform a full connection operation on the first mapping vector to form an initial depth vector.

[0017] In some preferred embodiments, in the above-mentioned multi-UAV path planning method based on collaborative tasks, the step of performing a first mapping operation on the UAV normalized vector to form a first mapping vector includes:

[0018] Performing sliding window processing on the drone normalized vector to form multiple drone sliding window vectors;

[0019] In the first mapping stage, a self-attention operation is performed on the first drone sliding window vector to form a corresponding drone attention vector;

[0020] In each of the second and subsequent mapping stages, based on the drone attention vector of the previous mapping stage, a cross-attention operation is performed on the drone sliding window vector corresponding to the current mapping stage to form the corresponding drone cross vector;

[0021] Based on the similarity between the drone cross vector corresponding to the current mapping stage and the drone attention vector of the previous mapping stage, the transmission control parameters of the current stage are determined;

[0022] When the transmission control parameter is equal to 0, the drone attention vector of the previous mapping stage and the drone cross vector of the current mapping stage are added bit by bit to form the drone attention vector corresponding to the current stage; when the transmission control parameter is equal to 1, the drone cross vector of the current mapping stage is determined as the drone attention vector corresponding to the current stage;

[0023] The drone attention vector corresponding to the last mapping stage is used as the first mapping vector.

[0024] In some preferred embodiments, in the above-mentioned multi-UAV path planning method based on collaborative tasks, the step of determining the transmission control parameters of the current stage based on the similarity between the UAV cross vector corresponding to the current mapping stage and the UAV attention vector of the previous mapping stage includes:

[0025] Determine the similarity between the drone cross vector corresponding to the current mapping stage and the drone attention vector of the previous mapping stage, and determine the ratio between the number of transmission control parameters equal to 0 and the number of transmission control parameters equal to 1 in each previous mapping stage;

[0026] Based on the quantity ratio, the similarity is updated to form an updated similarity;

[0027] The updated similarity is rounded to form the transmission control parameter of the current stage.

[0028] In some preferred embodiments, in the above-mentioned multi-UAV path planning method based on collaborative tasks, the step of performing a second mapping operation on the initial depth vector to form an intermediate depth vector includes:

[0029] extracting a first characterization vector and a second characterization vector from the initial depth vector;

[0030] performing a nonlinear activation operation on the first representation vector to form a nonlinear activation vector, wherein each vector parameter in the nonlinear activation vector is less than or equal to 1 and greater than or equal to 0;

[0031] The nonlinear activation vector and the second representation vector are bitwise multiplied, and a fully connected operation is performed on the result of the bitwise multiplication operation to form an intermediate depth vector.

[0032] In some preferred embodiments, in the above-mentioned multi-UAV path planning method based on collaborative tasks, the step of extracting the first characterization vector and the second characterization vector from the initial depth vector includes:

[0033] Mapping the vector parameters in the initial depth vector to a target interval to form a normalized depth vector, wherein the lower limit value of the target interval is equal to 0 and the upper limit value is equal to 1;

[0034] The normalized depth vector is convolved through two convolutional network layers to form a first representation vector and a second representation vector of the same size, respectively, wherein the two convolutional network layers have the same convolution kernel architecture and different convolution kernel parameters.

[0035] In some preferred embodiments, in the above-mentioned multi-UAV path planning method based on collaborative tasks, the step of performing a third mapping operation on the intermediate depth vector to form a coding vector of the first mapping subunit includes:

[0036] performing a normalization operation on the intermediate depth vector to form an intermediate normalized vector;

[0037] Performing a pooling operation on the intermediate normalized vector through two pooling network layers, respectively, to form a first pooling vector and a second pooling vector of the same size, wherein the two pooling network layers have the same pooling window and pooling step size, but different pooling modes;

[0038] Performing a nonlinear activation operation on the first pooled vector, and performing a bitwise multiplication operation on the second pooled vector based on a result of the nonlinear activation operation to form an encoding vector of a first mapping subunit;

[0039] In which, in the semantic decoding unit, each mapping sub-unit is used to perform a first mapping operation, a second mapping operation and a third mapping operation in sequence, and the first mapping operation and the second mapping operation have the same mapping method as the first mapping operation and the second mapping operation in the semantic encoding unit. The third mapping operation is based on the third mapping operation in the semantic encoding unit, and performs a full connection operation on the result of the bitwise multiplication operation.

[0040] In some preferred embodiments, in the above-mentioned multi-UAV path planning method based on collaborative tasks, the step of generating the target position of each UAV based on the UAV depth vector and the operation area vector of the unfinished operation area of the multiple UAVs includes:

[0041] Determining operation area vectors of unfinished operation areas of the plurality of drones, wherein, in the operation area vectors, coordinates representing positions corresponding to parameters equal to 1 belong to unfinished operation areas, and coordinates representing positions corresponding to parameters equal to 0 belong to completed operation areas;

[0042] Fusing the UAV depth vector and the operation area vector to form a UAV operation vector;

[0043] For the first UAV, obtaining a corresponding position probability distribution based on the UAV operation vector mapping, and determining the target position corresponding to the UAV based on the position probability distribution, wherein the position probability distribution is used to reflect the probability that each position is the target position of the UAV;

[0044] For the second and subsequent drones, the output vector corresponding to the current drone is determined based on the embedding vector of the target position corresponding to the previous drone and the output vector corresponding to the previous drone. The corresponding position probability distribution is obtained based on the output vector mapping, and the target position corresponding to the drone is determined based on the position probability distribution. The output vector corresponding to the first drone is the drone operation vector.

[0045] An embodiment of the present invention also provides a multi-UAV path planning system based on collaborative tasks, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the above-mentioned multi-UAV path planning method based on collaborative tasks.

[0046] The multi-UAV path planning method and system based on collaborative tasks provided by embodiments of the present invention first determines global UAV data and embeds it to form a UAV embedding vector. Second, deep mining is performed on the UAV embedding vector to form a UAV depth vector. Then, based on the UAV depth vector and the operation area vectors of the unfinished operation areas of the multiple UAVs, a target position for each UAV is generated. Based on the above, on the one hand, the embedding and deep mining process can obtain latent semantic information from the UAV global data, enabling more reliable path planning (the basis for path planning can be more reliable and richer) compared to conventional human operations based on experience. On the other hand, because the UAV global data reflects the current position, current battery level, and current material reserve of the UAV during the operation process, it fully considers dynamic information during the operation process. For example, the current battery level and current material reserve can affect whether the UAV can perform effective operations. This ensures that the determined target position is a position where the UAV can actually perform effective operations, ensuring the reliability of path planning, thereby improving the relatively low reliability of multi-UAV path planning in the prior art.

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a structural block diagram of a multi-UAV path planning system based on collaborative tasks provided in an embodiment of the present invention.

[0049] Figure 2 A flowchart of the steps included in the multi-UAV path planning method based on collaborative tasks provided in an embodiment of the present invention.

[0050] Figure 3 A schematic diagram of deep mining provided by an embodiment of the present invention.

[0051] Figure 4 A schematic diagram of a first mapping operation provided by an embodiment of the present invention.

[0052] Figure 5 Another schematic diagram of the first mapping operation provided by an embodiment of the present invention.

[0053] Figure 6A schematic diagram of an operation area provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0055] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0056] like Figure 1 As shown, an embodiment of the present invention provides a multi-UAV path planning system based on collaborative tasks. The multi-UAV path planning system based on collaborative tasks may include a memory and a processor.

[0057] Specifically, the memory and processor are directly or indirectly electrically connected to each other to enable data transmission or interaction. For example, the electrical connection may be achieved via one or more communication buses or signal lines. The memory may store at least one software function module (computer program) in the form of software or firmware. The processor may be configured to execute the executable computer program stored in the memory, thereby implementing the collaborative task-based multi-UAV path planning method provided in an embodiment of the present invention (described below).

[0058] That is to say, the multi-UAV path planning system based on collaborative tasks can be used for:

[0059] Determining drone global data and embedding the drone global data to form a drone embedding vector, wherein the drone global data is used to reflect at least the current position, current power level, and current material reserve level of each of the multiple drones during operation, and the drone embedding vector is used to reflect shallow semantic information of the drone global data;

[0060] Deep mining the drone embedding vector to form a drone depth vector, wherein the drone depth vector is used to reflect deep semantic information of the drone global data;

[0061] Based on the drone depth vector and the operation area vectors of the unfinished operation areas of the multiple drones, a target position of each drone is generated, wherein each target position belongs to the unfinished operation area, and the target position serves as a moving target for the drone to move from the current position to complete path planning, and the drone completes the corresponding operation task based on the stored materials in the process of moving from the current position to the target position.

[0062] Optionally, the memory may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0063] Optionally, the processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), a system on chip (SoC), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0064] and, Figure 1 The structure shown is for illustration only. The multi-UAV path planning system based on collaborative tasks may also include Figure 1 More or fewer components than shown, or with Figure 1 The different configurations shown, for example, may include a communication unit for exchanging information with other devices (such as drones) to obtain the current position, current power level, and current material reserve of the drone during operation.

[0065] In an alternative example, the multi-UAV path planning system based on collaborative tasks may be a server or server cluster with data processing capabilities.

[0066] Combine Figure 2The embodiment of the present invention also provides a multi-UAV path planning method based on collaborative tasks, which can be applied to the multi-UAV path planning system based on collaborative tasks. The method steps defined in the process related to the multi-UAV path planning method based on collaborative tasks can be implemented by the multi-UAV path planning system based on collaborative tasks (hereinafter referred to as the path planning system). Figure 2 The specific process shown is explained in detail.

[0067] In step S110 , the drone global data is determined and the drone global data is embedded to form a drone embedding vector.

[0068] In an embodiment of the present invention, the path planning system can determine global drone data and embed this data to form a drone embedding vector. The global drone data is used to reflect at least the current location (e.g., longitude, latitude, and altitude), current battery level (i.e., remaining battery charge), and current material reserves (e.g., remaining fertilizer or pesticide) of each of the multiple drones during operation. The drone embedding vector is used to reflect the shallow semantic information of the global drone data. In other words, through preliminary embedding processing, the global drone data can be embedded in a vector space for representation.

[0069] Step S120 : performing deep mining on the drone embedding vector to form a drone depth vector.

[0070] In an embodiment of the present invention, after forming the drone embedding vector, the path planning system can perform deep mining on the drone embedding vector to form a drone depth vector. The drone depth vector is used to reflect the deep semantic information of the drone's global data. In other words, since the drone embedding vector represents shallow semantic information formed through embedding, further deep mining can be performed to further improve the representation accuracy of the semantic information, thereby obtaining high-level, abstract semantic information.

[0071] Step S130 : generating a target position of each drone based on the drone depth vector and the operation area vectors of the unfinished operation areas of the multiple drones.

[0072] In an embodiment of the present invention, after forming the drone depth vector, the path planning system can generate a target position for each drone based on the drone depth vector and the operation area vectors of the unfinished operation areas of the multiple drones. Each of the target positions belongs to the unfinished operation area, and the target position serves as the moving target of the drone starting from the current position to complete the path planning, and the drone completes the corresponding operation task based on the stored materials in the process of moving from the current position to the target position. In other words, when performing path planning, not only the current position, current power level, and current material reserve of the drone should be considered, but also the actual unfinished operation area, that is, the target position of the drone should be effectively constrained, thereby improving the reliability of the determined target position.

[0073] Based on the above, on the one hand, because the latent semantic information in the drone's global data can be obtained through embedded processing and deep mining, more reliable path planning can be performed compared to conventional operations based on human experience (the basis for path planning can be more reliable and richer). On the other hand, because the drone's global data is used to reflect the drone's current position, current battery level, and current material reserves during the operation, it fully considers dynamic information during the operation. For example, the current battery level and current material reserves will affect whether the drone can perform effective operations. This ensures that the determined target position is a position where the drone can actually perform effective operations, ensuring the reliability of path planning, thereby improving the relatively low reliability of multi-drone path planning existing in existing technologies.

[0074] In the first part, it is necessary to further explain step S110 that the specific method of forming the drone embedding vector is not limited and can be selected according to actual needs.

[0075] For example, in a specific embodiment, the drone global data may be text data, such as the current location, current power level, and current material reserve of drone 1, and the current location, current power level, and current material reserve of drone 2. Based on this, after obtaining the drone global data, the drone global data may be word embedded (Word Embedding, by mapping words to a continuous vector space so that words with similar semantics are closer in the vector space) through a trained word embedding model (such as a Word2Vec model, etc.) to obtain the corresponding drone embedding vector.

[0076] In the second part, it is necessary to further explain step S120 that the specific method of deep mining the drone embedding vector is not limited and can be selected according to actual needs.

[0077] For example, in a specific embodiment, the drone embedding vector can be processed by Deep Convolutional Neural Networks (DCNN) to achieve deep mining, thereby forming a corresponding drone depth vector.

[0078] For example, in another specific embodiment, considering that the current position, current power and current material reserve of the drone during operation have no direct semantic relationship, but there is a potential correlation. For example, as the drone continues to move, the position will continue to change, the corresponding power will also decrease, and the material storage will also decrease. Therefore, it is necessary to capture these potential correlations. Based on this, the above step S120 can further include step S121, step S122 and step S123. The specific implementation process of each step is as follows (combined with Figure 3 shown).

[0079] Step S121: In the first mapping subunit included in the semantic coding unit, the drone embedding vector is subjected to a first mapping operation to form an initial depth vector, and the initial depth vector is subjected to a second mapping operation to form an intermediate depth vector, and the intermediate depth vector is subjected to a third mapping operation to form a coding vector of the first mapping subunit.

[0080] In an embodiment of the present invention, deep mining of the drone embedding vector can be achieved through a semantic encoding unit and a semantic decoding unit. During the encoding process of the semantic encoding unit, as the network depth gradually increases, the size of the corresponding output vector can gradually decrease. For example, it can be 1 / 2, 1 / 4, 1 / 8, 1 / 16, 1 / 32 of the size of the drone embedding vector. During the decoding process of the semantic decoding unit, as the network depth gradually increases, the size of the corresponding output vector can gradually increase. For example, it can be 1 / 16, 1 / 8, 1 / 4, 1 / 2, 1 times the size of the drone embedding vector. In other words, during the encoding process, the vector is gradually compressed, and during the decoding process, the vector is gradually expanded. Specifically, in a first mapping subunit included in the semantic encoding unit, the drone embedding vector is subjected to a first mapping operation to form an initial depth vector, the initial depth vector is subjected to a second mapping operation to form an intermediate depth vector, and the intermediate depth vector is subjected to a third mapping operation to form the encoding vector of the first mapping subunit. The first mapping operation, the second mapping operation, and the third mapping operation are different from each other, that is, three different mapping operations are used to capture or fit some potential nonlinear correlations in the drone embedding vector.

[0081] Step S122, load the encoding vector of the previous mapping subunit into the next mapping subunit for multiple levels of mapping operations until the encoding vector of the last mapping subunit is formed, and use the encoding vector as the drone encoding vector of the semantic encoding unit.

[0082] In an embodiment of the present invention, in the mapping subunits after the first one included in the semantic coding unit, the encoding vector of the previous mapping subunit can be loaded into the next mapping subunit for multiple levels of mapping operations until the encoding vector of the last mapping subunit is formed, and the encoding vector is used as the drone encoding vector of the semantic coding unit. The semantic coding unit is configured with multiple mapping subunits connected in sequence. That is, in the second mapping subunit, the encoding vector of the first mapping subunit can be sequentially subjected to the first mapping operation, the second mapping operation, and the third mapping operation as described above (the logic of each mapping operation in the first mapping subunit is the same, and the specific mapping parameters can be formed during training) to obtain the encoding vector of the second mapping subunit, and so on until the encoding vector of the last mapping subunit is obtained, completing the encoding. Among them, the size of the encoding vector of the first mapping subunit can be 1 / 2 of the size of the drone embedding vector, the size of the encoding vector of the second mapping subunit can be 1 / 4 of the size of the drone embedding vector, the size of the encoding vector of the third mapping subunit can be 1 / 8 of the size of the drone embedding vector, the size of the encoding vector of the fourth mapping subunit can be 1 / 16 of the size of the drone embedding vector, and the size of the encoding vector of the fifth mapping subunit can be 1 / 32 of the size of the drone embedding vector.

[0083] Step S123: Using a semantic decoding unit, a semantic decoding operation is performed on the drone encoding vector of the semantic encoding unit to form a drone decoding vector, and a drone depth vector is obtained based on the drone decoding vector.

[0084] In an embodiment of the present invention, after the encoding is completed, that is, after the drone encoding vector of the semantic encoding unit is obtained, the semantic decoding unit can be used to perform a semantic decoding operation on the drone encoding vector of the semantic encoding unit to form a drone decoding vector, and based on the drone decoding vector, a drone depth vector is obtained. Wherein, the semantic decoding unit is configured with multiple mapping subunits connected in sequence, that is, the first mapping subunit performs a semantic decoding operation on the drone encoding vector of the semantic encoding unit to obtain a corresponding decoding vector, the second mapping subunit performs a semantic decoding operation on the decoding vector corresponding to the first mapping subunit to obtain a corresponding decoding vector, and so on, to obtain the decoding vector corresponding to the last mapping subunit, and use the drone decoding vector as the corresponding drone decoding vector. Then, the drone decoding vector can be directly used as the corresponding drone depth vector, or the drone decoding vector can be further processed, such as attention processing, to obtain the drone depth vector. Among them, the size of the decoding vector of the first mapping subunit can be 1 / 16 of the size of the drone embedding vector, the size of the decoding vector of the second mapping subunit can be 1 / 8 of the size of the drone embedding vector, the size of the decoding vector of the third mapping subunit can be 1 / 4 of the size of the drone embedding vector, the size of the decoding vector of the fourth mapping subunit can be 1 / 2 of the size of the drone embedding vector, and the size of the decoding vector of the fifth mapping subunit can be 1 times the size of the drone embedding vector, that is, decoded to the size of the drone embedding vector.

[0085] Optionally, in the above step S121, the specific manner of performing the first mapping operation on the drone embedding vector is not limited. In order to fully mine the deep semantic information and consider the semantic distortion problem and computing resource consumption problem in the deep mining process, the above step S121 may further include step S121a, step S121b and step S121c. The specific implementation process of each step is as follows (combined with Figure 4 shown).

[0086] Step S121a: Map the vector parameters in the drone embedding vector to a target interval to form a drone normalized vector.

[0087] In this embodiment of the present invention, the vector parameters in the drone embedding vector can be first mapped (i.e., normalized) to a target interval to form a drone normalized vector. The target interval has a lower limit of 0 and an upper limit of 1. This can reduce the computational complexity of subsequent processing to a certain extent.

[0088] Step S121b: performing a first mapping operation on the drone normalized vector to form a first mapping vector.

[0089] In an embodiment of the present invention, after obtaining the drone normalized vector, the drone normalized vector may be subjected to a first mapping operation to form a first mapping vector. The first mapping operation includes multiple mapping stages, with a jump connection link configured between two adjacent mapping stages, and a transmission control parameter set on the link. The transmission control parameter is used to reflect whether the output of the previous mapping stage is connected to the output of the next mapping stage. For example, after the drone normalized vector is processed in the first mapping stage to obtain a first processing vector, the first processing vector may be processed in a second mapping stage. If the transmission control parameter between the first and second mapping stages is a first value, the first processing vector may be added to the processing result of the second mapping stage to obtain a second processing vector, thereby improving semantic distortion. If the transmission control parameter between the first and second mapping stages is a second value, the first processing vector may not be connected to the processing result of the second mapping stage. In this way, the processing result of the second mapping stage may be directly used as the second processing vector, thereby improving the problem of excessive computing resource consumption.

[0090] Step S121c: performing a full connection operation on the first mapping vector to form an initial depth vector.

[0091] In an embodiment of the present invention, after obtaining the first mapping vector, the first mapping vector can be fully connected to form an initial depth vector. It should be noted that in step S121c, after multiple stages of mapping processing, the size of the vector may change accordingly, such as decreasing. Based on this, in order to make the size of the formed initial depth vector equal to the size of the drone embedding vector, it can be achieved by performing a fully connected operation. For example, according to the rule of matrix multiplication, a vector of (m*n) is multiplied by a vector of (n*p) to obtain a vector of (m*p). Therefore, the first mapping vector and a fully connected matrix (the size can be (m*p), the specific parameters can be formed during the training process of the semantic coding unit, and initially, it can be a randomly generated matrix) can be multiplied to obtain the initial depth vector.

[0092] Optionally, in the above step S121b, the specific manner of performing the first mapping operation on the normalized vector of the drone is not limited. For example, in a specific embodiment, in order to fully capture the potential correlation through the first mapping operation, the above step S121b may further include step b1, step b2, step b3, step b4, step b5 and step b6. The specific implementation process of each step is as follows (combined with Figure 5 shown).

[0093] Step b1: performing sliding window processing on the drone normalized vector to form multiple drone sliding window vectors.

[0094] In an embodiment of the present invention, a sliding window process can be performed on the normalized drone vector to form multiple drone sliding window vectors. In this way, the size of each drone sliding window vector is equal to the window size of the sliding window process, such as a*b. Furthermore, the step size of the sliding window process can be equal to 1 to fully capture any potential correlations. Alternatively, the step size of the sliding window process can be greater than 1 to reduce the number of drone sliding window vectors formed, thereby reducing the consumption of computing resources.

[0095] Step b2: In the first mapping stage, a self-attention operation is performed on the first drone sliding window vector to form a corresponding drone attention vector.

[0096] In an embodiment of the present invention, after obtaining the multiple drone sliding window vectors, a self-attention operation (Self-Attention, allowing each vector element to interact with other parts according to its own different parts during the processing process, thereby dynamically adjusting the degree of attention to different parts to capture potential correlations) can be performed on the first drone sliding window vector in the first mapping stage to form a corresponding drone attention vector.

[0097] In step b3, in each of the second and subsequent mapping stages, based on the drone attention vector of the previous mapping stage, a cross-attention operation is performed on the drone sliding window vector corresponding to the current mapping stage to form a corresponding drone cross vector.

[0098] In an embodiment of the present invention, in each of the second and subsequent mapping stages, a cross-attention operation can be performed on the drone sliding window vector corresponding to the current mapping stage based on the drone attention vector of the previous mapping stage to form a corresponding drone cross-vector. That is, a cross-attention operation (where the query vector is derived from the drone attention vector of the first mapping stage, and the key vector and value vector are derived from the drone sliding window vector corresponding to the second mapping stage) can be performed on the drone sliding window vector corresponding to the second mapping stage based on the drone attention vector of the first mapping stage to form a drone cross-vector for the second mapping stage; and a cross-attention operation can be performed on the drone sliding window vector corresponding to the third mapping stage based on the drone attention vector of the second mapping stage to form a drone cross-vector for the third mapping stage, and so on. Furthermore, it should be noted that there is a one-to-one correspondence between the multiple drone sliding window vectors and the multiple mapping stages, i.e., the first drone sliding window vector serves as the drone sliding window vector corresponding to the first mapping stage, the second drone sliding window vector serves as the drone sliding window vector corresponding to the second mapping stage, the third drone sliding window vector serves as the drone sliding window vector corresponding to the third mapping stage, and so on.

[0099] In step b4, the transmission control parameters of the current phase are determined based on the similarity between the drone cross vector corresponding to the current mapping phase and the drone attention vector of the previous mapping phase.

[0100] In embodiments of the present invention, the transmission control parameter for the current phase can also be determined based on the similarity between the drone cross vector corresponding to the current mapping phase and the drone attention vector from the previous mapping phase. For example, when the similarity is high (e.g., greater than 0.5), the transmission control parameter can be equal to the second value, indicating that the similarity is relatively low and the likelihood of semantic distortion is low, so no connection is required. When the similarity is low (e.g., less than 0.5), the transmission control parameter can be equal to the first value, indicating that the similarity is relatively low and the likelihood of semantic distortion is high, so connection is required.

[0101] Step b5, when the transmission control parameter is equal to 0, the drone attention vector of the previous mapping stage and the drone cross vector of the current mapping stage are bitwise added to form the drone attention vector corresponding to the current stage; when the transmission control parameter is equal to 1, the drone cross vector of the current mapping stage is determined as the drone attention vector corresponding to the current stage.

[0102] In an embodiment of the present invention, when the transmission control parameter is equal to 0 (i.e., the first numerical value), the drone attention vector of the previous mapping stage and the drone cross vector of the current mapping stage can be added bit by bit to form the drone attention vector corresponding to the current stage; when the transmission control parameter is equal to 1 (i.e., the second numerical value), the drone cross vector of the current mapping stage is determined as the drone attention vector corresponding to the current stage.

[0103] In step b6, the drone attention vector corresponding to the last mapping stage is used as the first mapping vector.

[0104] In an embodiment of the present invention, after obtaining the drone attention vector corresponding to the last mapping stage, the drone attention vector corresponding to the last mapping stage can be used as the first mapping vector. In this way, by performing cross-attention operations on two adjacent drone sliding window vectors in sequence, the correlation between the two adjacent drone sliding window vectors can be captured.

[0105] Optionally, in step b4 above, the transmission control parameters for the current stage can be determined based on the similarity between vectors. This will increase the consumption of computing resources to a certain extent. Therefore, in other specific implementations, each transmission control parameter can be used as a parameter of the semantic coding unit, that is, formed during the training process of the semantic coding unit. In this way, in actual applications, the parameters learned from the sample data can be directly used, thereby reducing the consumption of computing resources. Among them, in the aforementioned implementation, the specific method of determining the transmission control parameters based on similarity is also not limited. For example, in a specific implementation, it can include:

[0106] The first step is to determine the similarity (e.g., cosine similarity) between the drone cross vector corresponding to the current mapping stage and the drone attention vector of the previous mapping stage, and to determine the ratio between the number of transmission control parameters equal to 0 and the number of transmission control parameters equal to 1 in each previous mapping stage;

[0107] In the second step, the similarity is updated based on the quantity ratio to form an updated similarity. For example, the quantity ratio can be multiplied by the similarity.

[0108] In a third step, the updated similarity may be rounded (it may be rounded down, where if it is greater than 1, it may be updated to 1) to form the transmission control parameter of the current stage.

[0109] That is to say, when the number of transmission control parameters equal to 0 in each previous mapping stage is large (that is, the number of connections is large), the updated similarity obtained may be greater, and the transmission control parameter formed in the current stage is more likely to be 1, that is, it is more likely that no connection will be made in the current mapping stage, so as to reduce the number or proportion of connections.

[0110] Optionally, in the above-mentioned step S121, the specific method of performing the second mapping operation on the initial depth vector is not limited. In order to realize the screening of important semantic information during the mapping process, the above-mentioned step S121 can further include step S121d, step S121e and step S121f. The specific implementation process of each step is as follows.

[0111] Step S121d: extracting a first characterization vector and a second characterization vector from the initial depth vector.

[0112] In an embodiment of the present invention, a first representation vector and a second representation vector can be extracted from the initial depth vector, wherein the first representation vector and the second representation vector can both represent the global semantic information in the initial depth vector, but the focus of the representation can be different, that is, the extracted semantic information can be different.

[0113] Step S121e: performing a nonlinear activation operation on the first characterization vector to form a nonlinear activation vector.

[0114] In an embodiment of the present invention, after obtaining the first representation vector, a nonlinear activation operation may be performed on the first representation vector to form a nonlinear activation vector. Each vector parameter in the nonlinear activation vector is less than or equal to 1 and greater than or equal to 0. For example, the nonlinear activation operation may be performed on the first representation vector using a Sigmoid activation function.

[0115] Step S121f: performing a bitwise multiplication operation on the nonlinear activation vector and the second representation vector, and performing a full connection operation on the result of the bitwise multiplication operation to form an intermediate depth vector.

[0116] In an embodiment of the present invention, after obtaining the nonlinear activation vector, a bitwise multiplication operation can be performed on the nonlinear activation vector and the second representation vector, and a fully connected operation can be performed on the result of the bitwise multiplication operation to form an intermediate depth vector. Since each vector parameter in the nonlinear activation vector is less than or equal to 1 and greater than or equal to 0, after the bitwise multiplication operation, the parameters in the second representation vector corresponding to the parameters of the nonlinear activation vector equal to 0 can be intercepted, the parameters in the second representation vector corresponding to the parameters of the nonlinear activation vector equal to 1 can be retained, and the parameters corresponding to the parameters between 0 and 1 can be updated and adjusted, thereby achieving different levels of focus. In addition, since the first and second representation vectors are extracted from the initial depth vector, their sizes may be reduced. Therefore, a further fully connected operation can be performed to maintain the same size, that is, the size of the intermediate depth vector is equal to that of the initial depth vector.

[0117] Optionally, in the above step S121d, the specific method of extracting the first representation vector and the second representation vector from the initial depth vector is not limited. For example, in a specific embodiment, in order to ensure that the first representation vector and the second representation vector have the same size and can both represent global semantic information, so that the semantic information of one can be controlled based on the semantic information of the other, the above step S121d may further include:

[0118] In the first step, the vector parameters in the initial depth vector may be mapped to a target interval to form a normalized depth vector, wherein the lower limit value of the target interval is equal to 0 and the upper limit value is equal to 1;

[0119] In the second step, the normalized depth vector can be convolved through two convolutional network layers to form a first representation vector and a second representation vector of the same size, wherein the two convolutional network layers have the same convolution kernel architecture and different convolution kernel parameters. For example, the first convolutional network layer may include convolution kernel 1, and the second convolutional network layer may include convolution kernel 2, wherein the size of convolution kernel 1 and convolution kernel 2 can both be 3*3, the step size can be 1, and the edges are not padded. Exemplarily, convolution kernel 1 can be:

[0120] ;

[0121] Convolution kernel 2 can be:

[0122] .

[0123] Optionally, in the above-mentioned step S121, the specific method of performing the third mapping operation on the intermediate depth vector is not limited. In order to further screen important semantic information during the mapping process, the above-mentioned step S121 may further include step S121g, step S121h and step S121i. The specific implementation process of each step is as follows.

[0124] Step S121g: performing a normalization operation on the intermediate depth vector to form an intermediate normalized vector.

[0125] In the embodiment of the present invention, the intermediate depth vector may be normalized first to form an intermediate normalized vector, thereby reducing the amount of calculation in subsequent processing.

[0126] Step S121h: performing a pooling operation on the intermediate normalized vector through two pooling network layers respectively to form a first pooling vector and a second pooling vector of the same size.

[0127] In an embodiment of the present invention, after forming the intermediate normalized vector, the intermediate normalized vector can be pooled through two pooling network layers to form a first pooling vector and a second pooling vector of the same size. The two pooling network layers have the same pooling window and pooling step size, but different pooling methods. For example, the size window corresponding to the first pooling network layer can be 3*3, the pooling step size can be 2, and the pooling method can be mean pooling. The size window corresponding to the second pooling network layer can be 3*3, the pooling step size can be 2, and the pooling method can be maximum pooling.

[0128] Step S121i: perform a nonlinear activation operation on the first pooled vector, and based on the result of the nonlinear activation operation, perform a bitwise multiplication operation on the second pooled vector to form a coding vector of the first mapping subunit.

[0129] In an embodiment of the present invention, after obtaining the first pooling vector and the second pooling vector, a nonlinear activation operation may be performed on the first pooling vector, and based on the result of the nonlinear activation operation, a bitwise multiplication operation may be performed on the second pooling vector to form an encoding vector of the first mapping subunit, as described above. It should be noted that in a mapping subunit, the first mapping operation and the second mapping operation can both achieve consistency in the size of the input and output through a full connection operation, and the third mapping operation may not perform a full connection operation. In this way, size compression can be achieved through the pooling operation in step S121h, and the specific compression ratio can be achieved by configuring the pooling step size as required, or, in step S121i, a further full connection operation may be performed on the result of the bitwise multiplication operation to achieve size adjustment.

[0130] Optionally, in the above-mentioned step S123, the specific manner of performing a semantic decoding operation on the drone encoding vector of the semantic coding unit is also not limited. For example, in a specific embodiment, in the semantic decoding unit, each mapping subunit is used to perform a first mapping operation, a second mapping operation and a third mapping operation in sequence, and the first mapping operation and the second mapping operation have the same mapping method as the first mapping operation and the second mapping operation in the semantic coding unit. The third mapping operation is based on the third mapping operation in the semantic coding unit, and a full connection operation is performed on the result of the bitwise multiplication operation to achieve the expansion of the vector size, such as from 1 / 32 of the size of the drone embedding vector to 1 / 16, 1 / 8, 1 / 4, 1 / 2, 1 times the size of the drone embedding vector, etc. The specific processing process can refer to the relevant description in the previous text and will not be repeated here.

[0131] In the second part, it is necessary to further explain step S120 that after deep mining of the drone embedding vector, the result of the deep mining can be further processed to obtain a drone depth vector with higher semantic representation accuracy.

[0132] Optionally, it has been found through research that if there is an abnormal change in the power level or the material reserve level at two adjacent locations, it may affect the prediction result of the subsequent target location. Therefore, the historical depth mining results can be further integrated to obtain the drone depth vector. For example, the above step S120 may include:

[0133] In the first step, deep mining can be performed on the drone embedding vector to obtain corresponding deep mining results (the deep mining method can be achieved by the relevant methods mentioned above);

[0134] In the second step, based on the historical deep mining results (i.e., the results of deep mining of the drone embedding vector formed by embedding the drone global data at the previous position) in the last target position prediction (i.e., when the current position is used as the target position), the deep mining results are updated to form the corresponding drone depth vector.

[0135] In one specific embodiment, the deep mining results may be updated by cross-attention processing. In another specific embodiment, in order to analyze the relationship between historical deep mining results and deep mining results from multiple perspectives, thereby providing a reliable basis for updating the deep mining results, the updating process may include:

[0136] In the first step, bitwise multiplication, bitwise addition, and bitwise subtraction operations may be performed on the historical deep mining result and the deep mining result to form corresponding multiplication analysis results, addition analysis results, and subtraction analysis results;

[0137] In a second step, adjacent difference calculations may be performed on the multiplication analysis result, the addition analysis result, and the subtraction analysis result, respectively, to obtain a multiplication difference result, an addition difference result, and a subtraction difference result. For example, an absolute difference calculation may be performed between the first parameter in the multiplication analysis result and the first parameter to obtain the first parameter in the multiplication difference result, and an absolute difference calculation may be performed between the second parameter in the multiplication analysis result and the first parameter to obtain the second parameter in the multiplication difference result, and an absolute difference calculation may be performed between the third parameter in the multiplication analysis result and the second parameter to obtain the third parameter in the multiplication difference result;

[0138] In a third step, the corresponding key point parameter distribution may be determined based on the volatility among the multiplication difference result, the addition difference result, and the subtraction difference result, wherein the position with greater volatility corresponds to a larger value of the key point parameter;

[0139] The fourth step is to multiply the key point parameter distribution and the depth mining result bit by bit to obtain the corresponding drone depth vector.

[0140] The calculation formula for the key point parameter distribution may include:

[0141] ;

[0142] On this basis, through the i By normalizing, we can get the corresponding key point parameters in the key point parameter distribution. Among them, Max() means taking the maximum value, A() means calculating the mean, B() means calculating the standard deviation, u i 、v i and z i Specifically, when volatility is large, the standard deviation will be larger and the natural logarithm function value will be smaller, so that the obtained Y i The larger the value of the corresponding key point parameter is, the larger the value is, that is, as a key point, so that through subsequent bit-by-bit multiplication operations, the parameter at the corresponding position in the deep mining result can be paid more attention.

[0143] In the third part, it is necessary to further explain step S130 that the specific method of generating the target position of each drone is not limited and can be selected according to actual needs.

[0144] For example, in a specific embodiment, the depth vector of the drone and the working area vectors of the unfinished working areas of the multiple drones can be spliced together, and then the spliced vectors can be processed to generate position coordinates (for example, this can be achieved through a decoder in a Transformer model) to form the target position of each of the drones.

[0145] For example, in another specific embodiment, in order to make the generation of the target position applicable to the entire operation process of the drone, the above step S130 may further include:

[0146] In the first step, the operation area vectors of the unfinished operation areas of the multiple drones can be determined, wherein, in the operation area vector, the position represented by the coordinate corresponding to the parameter equal to 1 belongs to the unfinished operation area, and the position represented by the coordinate corresponding to the parameter equal to 0 belongs to the completed operation area. In this way, it can be ensured that the size of the operation area vector in the prediction of the target position at each moment in the entire operation process is the same, so as to facilitate subsequent processing; in addition, in order to ensure that the solution provided by the present invention can be adapted to more operation areas, the operation area vector can be configured as a vector with a larger size. In this way, when the number of positions in an operation area is less than the size, the corresponding size can be guaranteed by filling the edges with 0 values, that is, other areas that do not belong to the operation area are regarded as completed operation areas. In this way, it does not affect the constraints on the unfinished operation area; in addition, the number of positions in the operation area can be formed by evenly dividing the operation area according to a certain unit area, such as Figure 6 As shown, the unit area can be determined based on the area of the effective action area of the drone at a location when performing operations, such as 4 square meters; based on this, the current position and target position of the drone belong to Figure 6 The center position of a small grid in the image is the longitude, latitude and altitude of the corresponding position. In other words, the drone can Figure 6 Work on the area corresponding to each small grid;

[0147] In the second step, the drone depth vector and the operation area vector may be fused to form a drone operation vector. For example, the drone depth vector and the operation area vector may be concatenated or added to form the drone operation vector. Alternatively, self-attention processing may be performed on the concatenation and addition results to form the drone operation vector.

[0148] In the third step, for the first UAV, a corresponding position probability distribution can be obtained based on the UAV operation vector mapping, and the target position corresponding to the UAV is determined based on the position probability distribution, wherein the position probability distribution is used to reflect the probability that each position belongs to the target position of the UAV (so, the position with the highest probability belongs to the target position), wherein the size of the position probability distribution can be equal to the size of the operation area vector, so that each parameter can correspond to a position; in addition, the UAV operation vector can be fully connected, and then the size is mapped to the size of the operation area vector, and then the result of the fully connected processing is mapped through the softmax function to obtain the corresponding position probability distribution;

[0149] In the fourth step, for each subsequent drone, the output vector corresponding to the current drone can be determined based on the embedding vector of the target location corresponding to the previous drone (the size of this embedding vector is equal to the size of the operation area vector, where only the parameter corresponding to the target location is equal to 1, and the parameters corresponding to other locations are equal to 0) and the output vector corresponding to the previous drone (e.g., concatenated, added, or the result of self-attention processing of the concatenation and addition), and the corresponding position probability distribution (as described above) mapped based on this output vector. The target location corresponding to the drone is then determined based on this position probability distribution, where the output vector corresponding to the first drone is the drone operation vector. It should be noted that the order of the multiple drones can be predetermined, such as drone a, drone b, drone c, drone d, drone e, etc., i.e., the first drone is drone a, the second drone is drone b, and the third drone is drone c. This enables information transfer or task collaboration during target location determination between drones. Each drone not only performs its task independently but also relies on the decision-making results of the previous drone, which can improve the collaborative nature of task execution.

[0150] In summary, the multi-UAV path planning method and system based on collaborative tasks provided by the present invention first determines global UAV data and embeds it to form a UAV embedding vector. Second, deep mining is performed on the UAV embedding vector to form a UAV depth vector. Then, based on the UAV depth vector and the operation area vectors of the unfinished operation areas of the multiple UAVs, a target position for each UAV is generated. Based on the above, on the one hand, because the embedding and deep mining can obtain latent semantic information in the UAV global data, more reliable path planning can be achieved compared to conventional human operations based on experience (the basis for path planning can be reliable and rich). On the other hand, because the UAV global data reflects the current position, current battery level, and current material reserve of the UAV during the operation process, dynamic information during the operation process is fully considered. For example, the current battery level and current material reserve can affect whether the UAV can perform effective operations. Therefore, the determined target position is the location where the UAV can actually perform effective operations, ensuring the reliability of path planning, thereby improving the relatively low reliability of multi-UAV path planning in the prior art.

[0151] In the several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions, and operations of the devices, methods, and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or part of the code, which contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the boxes can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, as well as the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified functions or actions, or can be implemented using a combination of dedicated hardware and computer instructions.

[0152] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0153] If the functions are implemented in the form of software modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, electronic device, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or device that includes a series of elements includes not only those elements but also other elements not explicitly listed, or also includes elements inherent to such process, method, article, or device. Without further constraints, an element defined by the phrase "comprises a..." does not preclude the existence of additional identical elements in the process, method, article or apparatus that includes the element.

[0154] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A multi-UAV path planning method based on collaborative tasks, characterized in that: include: Determining drone global data and embedding the drone global data to form a drone embedding vector, wherein the drone global data is used to reflect at least the current position, current power level, and current material reserve level of each of the multiple drones during operation, and the drone embedding vector is used to reflect shallow semantic information of the drone global data; In a first mapping subunit included in the semantic encoding unit, the drone embedding vector is subjected to a first mapping operation to form an initial depth vector, the initial depth vector is subjected to a second mapping operation to form an intermediate depth vector, and the intermediate depth vector is subjected to a third mapping operation to form an encoding vector of the first mapping subunit, wherein the first mapping operation, the second mapping operation, and the third mapping operation are different from each other; Loading the encoding vector of the previous mapping subunit into the next mapping subunit to perform multiple levels of mapping operations until the encoding vector of the last mapping subunit is formed, and using the encoding vector as the drone encoding vector of the semantic encoding unit, wherein the semantic encoding unit is configured with multiple mapping subunits connected in sequence; Using a semantic decoding unit, performing a semantic decoding operation on the drone encoding vector of the semantic encoding unit to form a drone decoding vector, and obtaining a drone depth vector based on the drone decoding vector, wherein the semantic decoding unit is configured with a plurality of mapping subunits connected in sequence, wherein the drone depth vector is used to reflect deep semantic information of the drone global data; Based on the drone depth vector and the operation area vectors of the unfinished operation areas of the multiple drones, a target position of each drone is generated, wherein each target position belongs to the unfinished operation area, and the target position serves as a moving target for the drone to move from the current position to complete path planning, and the drone completes the corresponding operation task based on the stored materials in the process of moving from the current position to the target position.

2. The multi-UAV path planning method based on collaborative tasks according to claim 1, characterized in that: The step of performing a first mapping operation on the drone embedding vector to form an initial depth vector includes: Mapping the vector parameters in the drone embedding vector to a target interval to form a drone normalized vector, wherein the lower limit value of the target interval is equal to 0 and the upper limit value is equal to 1; Performing a first mapping operation on the normalized vector of the UAV to form a first mapping vector, wherein the first mapping operation includes a plurality of mapping stages, a link with a jump connection is configured between two adjacent mapping stages, and a transmission control parameter is set on the link, wherein the transmission control parameter is used to reflect whether the output of the previous mapping stage is connected to the output of the next mapping stage; Perform a full connection operation on the first mapping vector to form an initial depth vector.

3. The multi-UAV path planning method based on collaborative tasks according to claim 2, characterized in that: The step of performing a first mapping operation on the drone normalized vector to form a first mapping vector includes: Performing sliding window processing on the drone normalized vector to form multiple drone sliding window vectors; In the first mapping stage, a self-attention operation is performed on the first drone sliding window vector to form a corresponding drone attention vector; In each of the second and subsequent mapping stages, based on the drone attention vector of the previous mapping stage, a cross-attention operation is performed on the drone sliding window vector corresponding to the current mapping stage to form the corresponding drone cross vector; Based on the similarity between the drone cross vector corresponding to the current mapping stage and the drone attention vector of the previous mapping stage, the transmission control parameters of the current stage are determined; When the transmission control parameter is equal to 0, the drone attention vector of the previous mapping stage and the drone cross vector of the current mapping stage are added bit by bit to form the drone attention vector corresponding to the current stage; when the transmission control parameter is equal to 1, the drone cross vector of the current mapping stage is determined as the drone attention vector corresponding to the current stage; The drone attention vector corresponding to the last mapping stage is used as the first mapping vector.

4. The multi-UAV path planning method based on collaborative tasks according to claim 3, characterized in that: The step of determining the transmission control parameters of the current stage based on the similarity between the drone cross vector corresponding to the current mapping stage and the drone attention vector of the previous mapping stage includes: Determine the similarity between the drone cross vector corresponding to the current mapping stage and the drone attention vector of the previous mapping stage, and determine the ratio between the number of transmission control parameters equal to 0 and the number of transmission control parameters equal to 1 in each previous mapping stage; Based on the quantity ratio, the similarity is updated to form an updated similarity; The updated similarity is rounded to form the transmission control parameter of the current stage.

5. The multi-UAV path planning method based on collaborative tasks according to claim 1, characterized in that: The step of performing a second mapping operation on the initial depth vector to form an intermediate depth vector comprises: extracting a first characterization vector and a second characterization vector from the initial depth vector; performing a nonlinear activation operation on the first representation vector to form a nonlinear activation vector, wherein each vector parameter in the nonlinear activation vector is less than or equal to 1 and greater than or equal to 0; The nonlinear activation vector and the second representation vector are bitwise multiplied, and a fully connected operation is performed on the result of the bitwise multiplication operation to form an intermediate depth vector.

6. The multi-UAV path planning method based on collaborative tasks according to claim 5, characterized in that: The step of extracting the first characterization vector and the second characterization vector from the initial depth vector comprises: Mapping the vector parameters in the initial depth vector to a target interval to form a normalized depth vector, wherein the lower limit value of the target interval is equal to 0 and the upper limit value is equal to 1; The normalized depth vector is convolved through two convolutional network layers to form a first representation vector and a second representation vector of the same size, respectively, wherein the two convolutional network layers have the same convolution kernel architecture and different convolution kernel parameters.

7. The multi-UAV path planning method based on collaborative tasks according to claim 1, characterized in that: The step of performing a third mapping operation on the intermediate depth vector to form a coding vector of the first mapping subunit comprises: performing a normalization operation on the intermediate depth vector to form an intermediate normalized vector; Performing a pooling operation on the intermediate normalized vector through two pooling network layers, respectively, to form a first pooling vector and a second pooling vector of the same size, wherein the two pooling network layers have the same pooling window and pooling step size, but different pooling modes; Performing a nonlinear activation operation on the first pooled vector, and performing a bitwise multiplication operation on the second pooled vector based on a result of the nonlinear activation operation to form an encoding vector of a first mapping subunit; In which, in the semantic decoding unit, each mapping sub-unit is used to perform a first mapping operation, a second mapping operation and a third mapping operation in sequence, and the first mapping operation and the second mapping operation have the same mapping method as the first mapping operation and the second mapping operation in the semantic encoding unit. The third mapping operation is based on the third mapping operation in the semantic encoding unit, and performs a full connection operation on the result of the bitwise multiplication operation.

8. The multi-UAV path planning method based on collaborative tasks according to claim 1, characterized in that: The step of generating a target position of each drone based on the drone depth vector and the operation area vectors of the unfinished operation areas of the multiple drones includes: Determining operation area vectors of unfinished operation areas of the plurality of drones, wherein, in the operation area vectors, coordinates representing positions corresponding to parameters equal to 1 belong to unfinished operation areas, and coordinates representing positions corresponding to parameters equal to 0 belong to completed operation areas; Fusing the UAV depth vector and the operation area vector to form a UAV operation vector; For the first UAV, obtaining a corresponding position probability distribution based on the UAV operation vector mapping, and determining the target position corresponding to the UAV based on the position probability distribution, wherein the position probability distribution is used to reflect the probability that each position is the target position of the UAV; For the second and subsequent drones, the output vector corresponding to the current drone is determined based on the embedding vector of the target position corresponding to the previous drone and the output vector corresponding to the previous drone. The corresponding position probability distribution is obtained based on the output vector mapping, and the target position corresponding to the drone is determined based on the position probability distribution. The output vector corresponding to the first drone is the drone operation vector.

9. A multi-UAV path planning system based on collaborative tasks, characterized by: It includes a processor and a memory, the memory is used to store a computer program, and the processor is used to execute the computer program to implement the multi-UAV path planning method based on collaborative tasks as described in any one of claims 1 to 8.

Images