An unmanned aerial vehicle software information matching method, device, equipment and storage medium

Abstract

The application discloses a kind of unmanned vehicle software information matching method, device and equipment and storage medium, it is related to data matching technical field, comprising: extracting the to-be-matched character of to-be-matched software information in to-be-matched software list and pre-set standard information base, construct the first feature for representing the pronunciation feature of to-be-matched character, the second feature for representing the structure feature of to-be-matched character and the third feature for representing the semantic feature of to-be-matched character of to-be-matched character, target feature is obtained using pre-set full connection network splicing first feature, second feature and third feature, and based on target feature, construct feature sequence, the feature sequence of to-be-matched character is matched, the similarity of different to-be-matched character is obtained, and the matching result of to-be-matched software information is determined based on similarity. By extracting three-dimensional features of sound, shape and meaning, the core features of characters are captured, and then complete character features are obtained by full connection network fusion, which significantly improves the matching accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of data matching technology, and in particular to a method, apparatus, device, and storage medium for matching software information of unmanned aerial vehicles (UAVs). Background Technology

[0002] In fields such as e-government, logistics and distribution, and financial services, there is a significant need to match non-standard addresses with standard address databases. This is particularly true in the software management of large drones, where the rapid expansion of software functionality due to drone platformization—with hundreds of software programs available for a single drone model—also necessitates the efficient matching of non-standard physical data lists with information from standard software libraries.

[0003] Traditional methods primarily rely on string edit distance, pinyin similarity, or single feature matching. Because they only consider surface-level string features, they lack effective handling of polyphonic Chinese characters and fail to capture subtle character-level differences. Therefore, they struggle to effectively address unique Chinese character features and various input errors. Thus, improving the ability to distinguish between polyphonic and similar-looking characters during Chinese character similarity calculations is a problem that needs to be solved in this field. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, device, and storage medium for matching UAV software information. By extracting three-dimensional features of sound, form, and meaning, the core features of characters are captured, and then the complete character features are obtained through fusion via a fully connected network, significantly improving the matching accuracy and reliability. The specific solution is as follows:

[0005] Firstly, this application provides a method for matching UAV software information, including:

[0006] The software to be matched list and the software information to be matched in the preset standard information database are determined respectively, and the characters to be matched in the software information are extracted to construct the first feature, the second feature and the third feature corresponding to the characters to be matched; the first feature is used to characterize the pronunciation feature of the characters to be matched, the second feature is used to characterize the structural feature of the characters to be matched, and the third feature is used to characterize the semantic feature of the characters to be matched.

[0007] The target feature is obtained by concatenating the first feature, the second feature, and the third feature using a preset fully connected network.

[0008] Based on the target features, a corresponding feature sequence is constructed, and the feature sequences of different characters to be matched are matched to obtain the similarity between the characters to be matched in the list of software to be matched and the characters to be matched in the preset standard information database, so as to determine the matching result corresponding to the software information to be matched based on the similarity.

[0009] Optionally, constructing the first feature corresponding to the character to be matched includes:

[0010] Determine several pronunciations corresponding to the character to be matched, and construct an initial sequence based on the several pronunciations and the character to be matched;

[0011] The first hidden state features of each pronunciation and the character to be matched in the initial sequence are extracted using a pre-trained language model.

[0012] Based on the self-attention mechanism, a first fusion feature is constructed between each of the pronunciations according to the first hidden state feature; and based on the cross-attention mechanism, a cross feature is constructed between the pronunciation and the character to be matched according to the first hidden state feature.

[0013] The first fusion feature and the cross feature are fused based on the first preset weight to obtain the first feature corresponding to the character to be matched.

[0014] Optionally, construct a second feature corresponding to the character to be matched, including:

[0015] A set of initial character images corresponding to the characters to be matched in several character formats are obtained, and the initial character images are preprocessed based on a preset image enhancement strategy to obtain the target character image; the character image is a grayscale image.

[0016] The multi-scale fusion features of the target character image are extracted using a preset deep neural network.

[0017] Global average pooling and global max pooling are performed on the multi-scale fusion features respectively to obtain the corresponding pooling results;

[0018] The second feature corresponding to the character to be matched is determined based on the pooling result.

[0019] Optionally, a third feature corresponding to the character to be matched is constructed, including:

[0020] Determine the preset window radius corresponding to the character to be matched, and construct a window sequence based on the character to be matched and the context content corresponding to the character to be matched according to the preset window radius;

[0021] The second hidden state features of each character in the window sequence are extracted using a pre-trained language model;

[0022] Determine the target position of the character to be matched in the window sequence, and extract the fourth feature of the target position from the second hidden state features;

[0023] Based on the second preset weight, the second hidden state features corresponding to each character in the window sequence are fused to obtain the fifth feature;

[0024] The location information corresponding to the target location is encoded to obtain the sixth feature;

[0025] The fourth feature, the fifth feature, and the sixth feature are combined to obtain the third feature corresponding to the character to be matched.

[0026] Optionally, the step of using a preset fully connected network to concatenate the first feature, the second feature, and the third feature to obtain the target feature includes:

[0027] The second fused feature is obtained by directly concatenating the first feature, the second feature, and the third feature.

[0028] The second fused feature is normalized to obtain the target vector;

[0029] The target vector is reduced in dimensionality by using two fully connected layers of the preset fully connected network to obtain the target features.

[0030] Optionally, the step of constructing a corresponding feature sequence based on the target feature and matching the feature sequences of different characters to be matched to obtain the similarity between the characters to be matched in the software list and the characters to be matched in the preset standard information database includes:

[0031] Construct a floating-point number sequence based on the target features corresponding to each character in the characters to be matched;

[0032] Determine the first floating-point number sequence corresponding to the character to be matched in the software list to be matched, and the second floating-point number sequence corresponding to the character to be matched in the preset standard information database, and construct a dynamic programming matrix based on the sequence lengths of the first floating-point number sequence and the second floating-point number sequence;

[0033] Traverse the dynamic programming matrix, extract the maximum similarity sum in the dynamic programming matrix, and normalize the maximum similarity sum to obtain the similarity between the character to be matched in the software list and the character to be matched in the preset standard information database.

[0034] Optionally, the step of constructing a corresponding feature sequence based on the target feature and matching the feature sequences of different characters to be matched to obtain the similarity between the characters to be matched in the software list and the characters to be matched in the preset standard information database includes:

[0035] Construct a binary encoding sequence based on the target features corresponding to each character in the characters to be matched;

[0036] A first position sequence and a second position sequence are constructed based on the binary encoded sequence; the first position sequence is used to represent the position of 1 in the binary encoded sequence, and the second position sequence is used to represent the position of 0 in the binary encoded sequence.

[0037] The first position sequence and the second position sequence of the character to be matched in the software list to be matched are respectively matched with the first position sequence and the second position sequence of the character to be matched in the preset standard information database, so as to determine the similarity between the character to be matched in the software list to be matched and the character to be matched in the preset standard information database based on the obtained matching results.

[0038] Secondly, this application provides a drone software information matching device, comprising:

[0039] The feature construction module is used to determine the software information to be matched in the list of software to be matched and the preset standard information database, respectively, and extract the characters to be matched in the software information to be matched, and construct the first feature, the second feature and the third feature corresponding to the characters to be matched; the first feature is used to characterize the pronunciation feature of the characters to be matched, the second feature is used to characterize the structural feature of the characters to be matched, and the third feature is used to characterize the semantic feature of the characters to be matched.

[0040] The feature splicing module is used to splice the first feature, the second feature, and the third feature using a preset fully connected network to obtain the target feature;

[0041] The feature matching module is used to construct a corresponding feature sequence based on the target feature, and to match the feature sequences of different characters to be matched to obtain the similarity between the characters to be matched in the list of software to be matched and the characters to be matched in the preset standard information database, so as to determine the matching result corresponding to the software information to be matched based on the similarity.

[0042] Thirdly, this application provides an electronic device, which includes a processor and a memory; wherein the memory is used to store a computer program, which is loaded and executed by the processor to implement the aforementioned UAV software information matching method.

[0043] Fourthly, this application provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the aforementioned UAV software information matching method.

[0044] This application first determines the list of software to be matched and the software information to be matched in a preset standard information database, and extracts the characters to be matched from the software information. It then constructs a first feature to represent the pronunciation of the characters to be matched, a second feature to represent the structural features of the characters to be matched, and a third feature to represent the semantic features of the characters to be matched. The first, second, and third features are then concatenated using a preset fully connected network to obtain the target features. Based on the target features, a corresponding feature sequence is constructed. The feature sequences of different characters to be matched are then matched to obtain the similarity between different characters to be matched, so as to determine the matching result corresponding to the software information to be matched based on the similarity.

[0045] Based on the above technical solution, this application first determines the list of software to be matched and the software information in the preset standard information database, extracts the core characters to be matched, and then extracts three types of features for each character: phonetic features, morphological features, and semantic features. Then, a fully connected network is used to concatenate and fuse the three-dimensional features of the same character, integrating multi-dimensional features and avoiding the limitations of single features. Afterwards, based on the target features of each character, a corresponding feature sequence is constructed. By comparing the feature sequences of different software information, the similarity is calculated, and finally, the matching of the list of software to be matched and the software information in the preset standard information database is completed based on the similarity. In this way, by extracting the three-dimensional features of phonetic, morphological, and semantic features, accurate character recognition and matching can be achieved, solving the problem of missed or incorrect matching caused by single-feature matching. Furthermore, by extracting the three-dimensional features of characters, the core features of each character can be captured. Then, by fusing them through a fully connected network, the target features contain complete character features, improving the robustness of matching in complex scenarios and significantly improving the matching accuracy and reliability. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0047] Figure 1 A flowchart of a method for matching software information of unmanned aerial vehicles (UAVs) provided in this application;

[0048] Figure 2 A framework diagram of a similarity calculation engine provided in this application;

[0049] Figure 3 A flowchart for sound feature vector conversion provided in this application;

[0050] Figure 4A flowchart of shape feature vector transformation provided in this application;

[0051] Figure 5 A flowchart of a semantic feature vector transformation is provided in this application;

[0052] Figure 6 A flowchart of a comprehensive representation vector transformation is provided for this application;

[0053] Figure 7 A flowchart for calculating the similarity of comprehensive representation vectors is provided for this application;

[0054] Figure 8 A flowchart for calculating the similarity of comprehensive characterization codes is provided for this application;

[0055] Figure 9 A schematic diagram of a UAV software information matching device provided in this application;

[0056] Figure 10 This application provides a structural diagram of an electronic device. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0059] In the software management of large unmanned aerial vehicles (UAVs), the rapid expansion of software functions due to the platformization of UAVs creates a need for efficient matching of non-standard physical data lists with standard software library information. Traditional methods mainly rely on string edit distance, pinyin similarity, or single feature matching, which are difficult to effectively handle the unique features of Chinese characters and various input errors. This application, however, extracts three-dimensional features of sound, form, and meaning to achieve accurate character recognition and matching, solving the problem of missed or mismatched matching caused by single feature matching. Furthermore, by fusing three-dimensional features through a fully connected network, the target features contain complete character features, improving the robustness of matching in complex scenarios and significantly enhancing matching accuracy and reliability.

[0060] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] See Figure 1 As shown, this embodiment of the invention discloses a method for matching UAV software information, including:

[0062] Step S11: Determine the software to be matched list and the software information to be matched in the preset standard information database respectively, and extract the characters to be matched from the software information to be matched, and construct the first feature, the second feature and the third feature corresponding to the characters to be matched; the first feature is used to characterize the pronunciation feature of the characters to be matched, the second feature is used to characterize the structural feature of the characters to be matched, and the third feature is used to characterize the semantic feature of the characters to be matched.

[0063] First of all, it should be pointed out that, such as Figure 2 As shown, in this embodiment, the complex string similarity calculation problem can be decomposed into three levels: character-level feature extraction, constructing Chinese character feature representations from the three dimensions of sound, form, and meaning; feature fusion optimization, effectively fusing multi-dimensional features into a unified representation; and sequence similarity calculation, realizing string-level similarity measurement based on feature sequences.

[0064] It can be understood that in the software management of large unmanned aerial vehicles (UAVs), due to the platformization of UAVs, the software functions have expanded rapidly. There is a need to efficiently match the non-standard physical data list with the information in the standard software library. Therefore, in this embodiment, the software list to be matched and the software information to be matched in the preset standard information library can be determined respectively first, and the characters to be matched in the software information to be matched are extracted, and the first feature, the second feature and the third feature corresponding to the characters to be matched are constructed. Among them, the above-mentioned first feature is the phonetic feature, which is used to represent the pronunciation feature of the characters to be matched, the second feature is the morphological feature, which is used to represent the structural feature of the characters to be matched, and the third feature is the semantic feature, which is used to represent the semantic feature of the characters to be matched. It can be understood that the above-mentioned characters to be matched usually refer to character strings, which include several characters, and in this embodiment, the above-mentioned characters are usually Chinese characters.

[0065] Specifically, when constructing the first feature corresponding to the characters to be matched, first determine several pronunciations corresponding to the characters to be matched, and construct an initial sequence based on the several pronunciations and the characters to be matched. Then use the pre-trained language model to extract the first hidden state features of each pronunciation and the characters to be matched in the initial sequence. Then, based on the self-attention mechanism, construct the first fusion feature between each pronunciation according to the first hidden state features, and based on the cross-attention mechanism, construct the cross feature between the pronunciation and the characters to be matched according to the first hidden state features. After that, fuse the first fusion feature and the cross feature based on the first preset weight to obtain the first feature corresponding to the characters to be matched. . That is to say, as Figure 3 shown, in this embodiment, the polyphonic characters can be uniformly encoded to obtain the phonetic feature of the character string. First, construct the input sequence, that is, the initial sequence, and define the unified input format for multiple pronunciations as: [CLS] Chinese character [SEP] pinyin1 pinyin2... pinyink [SEP]. For example: '重' [CLS] 重 [SEP]chong2 zhong4 [SEP]; '中' [CLS] 中 [SEP] zhong1 [SEP]. Assume that the character has multiple pronunciations , then the input sequence is:

[0066] .

[0067] By constructing Chinese characters and all their pronunciations into a unified format for model understanding, preparation is made for subsequent fusion of multi-pronunciation information. Then, the pre-trained language model BERT (Bidirectional Encoder Representations from Transformers) is used for unified feature extraction, first obtaining the features of the first hidden state. :

[0068] ;

[0069] in, For BERT pre-trained models with whole word masking, Let be the dimension representation of the matrix. 768 represents the sequence length and the hidden layer dimension of the BERT model.

[0070] Then determine the position of each character in the string, for example: Chinese characters. Corresponding location: Pinyin sequence position: Then, cross-pronunciation attention fusion is performed to construct the first fusion feature between each pronunciation and the cross feature between the pronunciation and the character to be matched. Specifically, first, self-attention between pinyin is performed, and the pinyin feature matrix is... for:

[0071] ;

[0072] in, To emerge from hidden state The feature vector of each pronunciation is extracted; then self-attention fusion is performed to obtain the fused feature. :

[0073]

[0074]

[0075] in, , , The weight matrix is ​​a learnable matrix; Q (query), K (key), and V (value) are the three core matrices for self-attention, used to calculate the similarity between pronunciations; Used to calculate the similarity between each pronunciation and all pronunciations; This is the scaling factor; This is a normalization function that converts similarity into weights of 0-1. This is the attention weight matrix.

[0076] Then, character-pinyin cross-attention is performed, with the query vector being the hidden state at the position of the Chinese character. The key-value pair represents the hidden state of all pinyin characters. Cross-attention at this point:

[0077] ;

[0078] ;

[0079] in, This is a weight matrix used to transform the features of Chinese characters and Pinyin; , Let be the feature vectors of the i-th and j-th pronunciations; Used to calculate the matching degree between Chinese character features and the i-th pronunciation feature; It is an exponential function, which amplifies the differences in matching degree; Let be the weight of the i-th pronunciation.

[0080] Then, attention-enhanced feature pooling is performed. In this embodiment, a multi-level pooling strategy is adopted, wherein, for character-level representation: directly taking... Pinyin level representation: Average pooling; cross representation: ; Fusion representation yields fusion features :

[0081] ;

[0082] in, This is a characteristic of Chinese characters themselves; Average pooling for pinyin fusion features; It is a feature of the intersection of Chinese character pinyin; For the fusion weight matrix; This is the bias vector. Then, attention-gated pooling is performed, first by assigning importance weights. , , calculate:

[0083]

[0084] in, Use the Sigmoid activation function; Weight matrix For bias; Weighting the importance of the inherent features of Chinese characters; Weights for the importance of pinyin fusion features; These are the importance weights for the cross-features. Then, a weighted fusion is performed based on these weights:

[0085]

[0086] in, This represents element-wise multiplication, where the weights and corresponding features are multiplied element by element. It is an L1 norm; The final weighted and fused features are then used. Feature dimensionality reduction is then performed, with the gated fused features as input to the dimensionality reduction network. Attention shift:

[0087]

[0088] Then, dimensionality reduction is achieved through a dimensionality reduction network:

[0089] ;

[0090] ;

[0091] ;

[0092] in, ; Presentation layer normalization operation; This is the weight matrix; Indicates bias; The transformed features; For activation functions; This is the weight matrix, used to gradually achieve dimensionality reduction; For the bias of the corresponding dimension; This is an intermediate dimensionality reduction feature; This results in the final 8-dimensional phonetic feature vector. Therefore, in this embodiment, the use of a fusion attention mechanism to enhance the learning of polyphonic weights during feature extraction of polyphonic characters enables the calculation of more effective feature vectors, achieving unified processing of polyphonic characters and improving adaptability to complex Chinese characters.

[0093] Then, the second feature corresponding to the character to be matched is constructed by character image generation and multi-font integration. First, initial character images corresponding to the characters to be matched in several character formats are obtained, and the initial character images are preprocessed based on a preset image enhancement strategy to obtain the target character image; the above character images are grayscale images. Then, the output multi-scale fusion features of the target character image are extracted using a preset deep neural network, and global average pooling and global max pooling are performed on the multi-scale fusion features respectively to obtain the corresponding pooling results, and the second feature corresponding to the character to be matched is determined based on the pooling results. .

[0094] Specifically, such as Figure 4As shown, multiple font sets can be constructed first. The main fonts are Source Han Sans Regular and Bold variants; the auxiliary fonts are Source Han Serif, Microsoft YaHei, and Founder Fangsong. Based on this, a font rendering strategy can be constructed, generating five font variants for each character to enhance the model's generalization ability. The rendering parameters include: canvas size: 72×72 pixels, providing edge buffering; center cropping to 64×64 pixels for training; color mode: single-channel grayscale image, reducing computational complexity; anti-aliasing level: 4x oversampling to ensure smooth stroke edges. Afterwards, multi-font feature fusion can be performed. First, a multi-font image set, i.e., the initial character image, needs to be generated for each character. :

[0095] ;

[0096] in, Five different font images of the Chinese character 'c'; For image rendering functions; Five preset fonts are available; These are rendering parameters.

[0097] Next, image preprocessing and data augmentation are performed. The standardization process includes the following steps:

[0098] Illumination normalization:

[0099] ;

[0100] The image is the normalized version, i.e., the target character image; This is the original rendered image; The average brightness of a batch of images; The standard deviation of brightness for a batch of images; It is the numerical stability constant;

[0101] Local contrast enhancement: Apply CLAHE (ContrastLimited Adaptive Histogram Equalization) to improve the visibility of stroke details.

[0102] Data augmentation strategies include geometric transformation augmentation and appearance transformation augmentation. Geometric transformation augmentation includes: random affine transformations, such as rotation ±8° and scaling by 0.9-1.1 times; elastically deformable meshes to simulate natural handwriting deformation; and perspective transformations, such as slight 3D rotation to enhance viewpoint invariance. Appearance transformation augmentation includes: morphological operations, such as dilation and erosion to simulate stroke thickness variations; and Gaussian blur. Simulates different levels of clarity; salt and pepper noise, such as 1% density, improves noise robustness.

[0103] Then, feature extraction is performed using a pre-defined deep neural network ResNet-18 (ResNet, i.e., Residual Neural Network) backbone. In this embodiment, an optimized ResNet-18 is used, and its specific structure is as follows:

[0104] First, network structure optimization is performed, including input layer adaptation: Input: 1×64×64 grayscale image; Stem layer: 7×7 convolution, 64 channels, stride=2, padding=3, and 3×3 max pooling, stride=2; Residual block configuration optimization: Stage1: [3×3, 64]×2, output 28×28×64; Stage2: [3×3, 128]×2, stride=2, output 14×14×128; Stage3: [3×3, 256]×2, stride=2, output 7×7×256; Stage4: [3×3, 512]×2, stride=1, output 7×7×512; Feature pyramid construction: multi-scale features from Stage2, Stage3, and Stage4 are fused, and upsampled to the same size after unifying the number of channels through 1×1 convolution. Furthermore, channel attention enhancement is performed, primarily through SE module integration: a compression-excitation module is added after each residual block.

[0105] ;

[0106] ;

[0107] ;

[0108] in, The output is the glyph feature map of the convolutional layer; GAP is global average pooling. The 1D channel features after GAP processing; This is the weight matrix; Activate ReLU; Activated for Sigmoid; This indicates that the weights are multiplied element-wise by the original feature map; This is the enhanced feature map.

[0109] Next, we can perform shape feature extraction. First, we use a feature pyramid network to perform multi-scale feature fusion:

[0110] ;

[0111] in This represents the feature maps of ResNet-18 at different stages. Small-scale features It is a mesoscale feature. It is a large-scale feature; Represents a 1×1 convolution; This is the fused multi-scale feature map; This indicates bilinear upsampling to the same spatial size.

[0112] Next, global context modeling is performed to enhance global average pooling: In this embodiment, a global max pooling branch, GMP (Global Max Pooling), is added to the standard GAP to achieve two-stream feature concatenation.

[0113] ;

[0114] After the splicing and adjustment, progressive feature dimensionality reduction is achieved through a dimensionality reduction network architecture, which is as follows:

[0115] Hierarchical feature compression:

[0116] Input: 1024 dimensions (GAP+GMP splicing);

[0117] FC1:1024 256, ReLU+BatchNorm;

[0118] FC2:256 64, ReLU+BatchNorm;

[0119] FC3:64 16, ReLU+BatchNorm;

[0120] FC4:16 8. Linear output;

[0121] Mathematical expression:

[0122] ;

[0123] ;

[0124] ;

[0125] ;

[0126] in, The input is a 1024-dimensional global feature vector; The weight matrix has dimensions of 256×1024, 64×256, 16×64, and 8×16, with the dimensions being gradually compressed. This is the bias vector for the corresponding dimension; That is, batch normalization; The intermediate dimensionality reduction features are 256-dimensional, 64-dimensional, and 16-dimensional, respectively; This is the final 8-dimensional feature vector, also known as the second feature. This process removes redundant information, extracts glyph features, and through four layers of progressive compression, retains key information, outputting an 8-dimensional vector that facilitates subsequent fusion with phonetic and semantic features.

[0127] Then as Figure 5 As shown, when constructing the third feature corresponding to the character to be matched, the preset window radius corresponding to the character to be matched can be determined, and a window sequence can be constructed based on the character to be matched and the context content corresponding to the character to be matched according to the preset window radius. Then, the second hidden state features of each character in the window sequence are extracted using the pre-trained language model BERT-base-Chinese. It determines the target position of the character to be matched in the window sequence and extracts the fourth feature of the target position from the second hidden state features. Based on the second hidden state features corresponding to each character in the second preset weighted fusion window sequence, the fifth feature is obtained. The location information corresponding to the target location is encoded to obtain the sixth feature. Then, the fourth, fifth, and sixth features can be combined to obtain the third feature corresponding to the character to be matched. Specifically, regarding window size, the basic window consists of 4 characters before and after the target character, forming a 9-character window (target character + 4 characters before and after). For important characters, an expanded window is used, which can be extended to 8 characters before and after. It should also be noted that this embodiment can dynamically adjust the window size based on the importance of the character within the vocabulary. For window boundary processing, the beginning of the text (i.e., the preceding context) is filled with [PAD] markers; the end of the text (i.e., the following context) is also filled with [PAD] markers. This maintains window structure consistency and ensures correct positional encoding. Therefore, for the target character... In position The context window is:

[0128] ;

[0129] in Target Chinese characters; That is, the position of the target Chinese character in the original string; Base window radius; The context character for the target character. And if the target character is at the beginning or end of the string, it is padded with [PAD] markers.

[0130] It should be noted that in this embodiment, the base model selected for pre-training is BERT-base-Chinese, with 110M parameters, 768 hidden layer dimensions, 12 layers, and 12 attention heads. The input is represented as a [CLS] context window sequence of Chinese characters [SEP], and absolute positional encoding is used to ensure accurate capture of character position information. The corresponding feature extraction process is forward propagation:

[0131] ;

[0132] Next, target character location is performed to determine the target character. Position in the sequence And extract the hidden state at the corresponding position: ;

[0133] in The sequence length is (window size + special markers). Refers to the BERT-base-Chinese pre-trained model; A sequence of context windows containing the special tags [CLS] and [SEP]; The hidden state matrix output by the model; Dimensions are matrices; The index of the target character's position in the sequence; This is the semantic feature vector of the target character.

[0134] Furthermore, this embodiment employs multi-level feature extraction to comprehensively capture the semantic dimension, avoiding the limitations of single features. First, character-level semantic features, i.e., basic semantic representations, are determined, and the character-level semantic vector is obtained directly using the BERT output of the target character position. :

[0135] ;

[0136] Furthermore, it should be noted that when processing sub-word information, for Chinese characters that are split into multiple sub-words, the first-sub-word representation or average pooling strategy is adopted. Then, the context-aware features are determined, and the local context encoding uses a weighted fusion of the hidden states before and after the target character to obtain the context-fused semantic vector. :

[0137] ;

[0138] Among them, weight Based on the distance decay calculation to the target character, The position of the character in the context window; For the first in the context window The hidden state of each character. This allows us to incorporate the semantic influence of surrounding characters on the target character and weight it accordingly. Ensure that neighboring characters have a greater impact.

[0139] Next, positional enhancement features are determined, and relative positional encoding is used to capture the relative positional information of the target character within the window to obtain a positional semantic vector. :

[0140] ;

[0141] in This is a positional encoding function that converts the relative positions of characters within a window into vectors. This represents the relative position of the target character within the window. For example, when k=4, the target character is in the center of the window, with a relative position of 4. This can solve the problem of semantically dependent positions. For example, in "I like apples" and "Apples like me," the position of the target character "apple" is different, resulting in completely different semantic relationships. This embodiment can capture this difference through positional encoding.

[0142] Then feature fusion and enhancement can be performed to concatenate multiple features into a concatenated feature vector. :

[0143] ;

[0144] Then, a feature fusion layer is used to fuse multi-source features using linear transformation to obtain a linearly fused feature vector. :

[0145] ;

[0146] in This is the weight matrix. ; This is the flattening operation, which converts the spliced ​​multidimensional vector into a one-dimensional vector to facilitate linear computation. This is the bias vector.

[0147] Furthermore, this embodiment introduces lightweight self-attention during the feature fusion stage:

[0148] ;

[0149] ;

[0150] ;

[0151] ;

[0152] ;

[0153] in, is the weight matrix; Q, K, and V are the self-attention core vectors used to calculate semantic importance; Used to calculate semantic relevance; This is the scaling factor; This is the attention weight matrix; This is the semantic vector after attention enhancement.

[0154] Then, an MLP (Multilayer Perceptron) projection dimensionality reduction network can be used to compress the 768-dimensional high-dimensional semantic features into an 8-dimensional standardized vector, removing redundancy while retaining the core semantics, facilitating fusion with sound and shape features. The hierarchical dimensionality reduction network architecture is as follows: input 768-dimensional fused features; FC1:768 256, ReLU activated + LayerNorm; Dropout: 0.1; FC2: 256 64, ReLU activation + LayerNorm; FC3:64 16, ReLU activation; FC4:16 8. Linear output. Mathematically, it is expressed as:

[0155] ;

[0156] ;

[0157] ;

[0158] ;

[0159] in The weight matrices have dimensions of 256×768, 64×256, 16×64, and 8×16, respectively. This is the bias vector for the corresponding dimension; This is an intermediate dimensionality reduction feature; This results in an 8-dimensional semantic feature vector as the final output. Through four layers of progressive dimensionality reduction, redundant semantic information is removed, retaining core features and providing accurate semantic basis for subsequent similarity calculations. Furthermore, this embodiment utilizes various feature extraction techniques to obtain feature representations in the three dimensions of sound, form, and meaning. Autoencoding fusion technology integrates these multi-dimensional feature information into a comprehensive representation vector. This multi-dimensional feature fusion of sound, form, and meaning more comprehensively captures the characteristics of Chinese characters.

[0160] Step S12: Use a preset fully connected network to splice the first feature, the second feature, and the third feature to obtain the target feature.

[0161] In this embodiment, a preset fully connected network can be used to concatenate the first feature, the second feature, and the third feature. Specifically, an 8-dimensional first feature can be directly concatenated. 8-dimensional second feature and 8-dimensional third feature The second fusion feature is obtained. The second fusion feature is then normalized to obtain the target vector. Then, the target vector is reduced in dimensionality by using two fully connected layers of a pre-defined fully connected network to obtain the target features. .

[0162] Specifically, such as Figure 6 As shown, the 8-dimensional feature vectors from the three sources are first concatenated in sequence to form a 24-dimensional fused feature vector:

[0163] ;

[0164] Then, batch normalization is performed, applying batch normalization to the concatenated 24-dimensional vector to obtain the target vector. To ensure numerical stability:

[0165] ;

[0166] in This is a vector of batch mean values; This is the batch standard deviation vector; It is the numerical stability constant, usually 1. ; These are learnable scaling and translation parameters.

[0167] Then, a two-layer fully connected network is used for processing. The first fully connected layer transforms the 24-dimensional input into a 16-dimensional hidden layer, resulting in... :

[0168] ;

[0169] in This is the weight matrix; This is the bias vector. Then, a nonlinear transformation is introduced using the GeLU activation function to obtain the activated 16-dimensional feature vector. :

[0170] ;

[0171] The GeLU function is defined as:

[0172] ;

[0173] in It is the cumulative distribution function of the standard normal distribution;

[0174] The second fully connected transformation further reduces dimensionality, compressing the 16-dimensional hidden layer into an 8-dimensional output layer to obtain the 8-dimensional feature vector after the second transformation. :

[0175] ;

[0176] in This is the weight matrix; This is the bias vector.

[0177] Then, the output can be normalized using the Sigmoid activation function. This can be achieved by using a range constraint and applying the Sigmoid activation function to ensure that all output values ​​are within the range [0, 1].

[0178] ;

[0179] in This is the final output 8-dimensional comprehensive representation vector, with values ​​between (0,1); This is the 8-dimensional vector output by the second fully connected layer. Then, the sigmoid function is applied element-wise, independently affecting each element of the vector.

[0180] ;

[0181] in This is the i-th element of the composite vector; This is an exponential function used to map numerical values ​​to (0,1).

[0182] Furthermore, to ensure output characteristics, range verification is required during output, which is achieved strictly through the mathematical properties of the Sigmoid function.

[0183] .

[0184] In this way, the range constraint can be used to force all feature element values ​​to fall within the (0,1) range through the Sigmoid function, avoiding the distortion of subsequent similarity calculations caused by excessively large or small values. Furthermore, standardization ensures that the comprehensive feature vectors of different characters are within the same numerical range. The final output 8-dimensional comprehensive representation vector contains pronunciation, character shape, and semantic information, and the standardized format helps to improve computational efficiency.

[0185] Step S13: Construct a corresponding feature sequence based on the target feature, and match the feature sequences of different characters to be matched to obtain the similarity between the characters to be matched in the list of software to be matched and the characters to be matched in the preset standard information database, so as to determine the matching result corresponding to the software information to be matched based on the similarity.

[0186] In this embodiment, a corresponding feature sequence can be constructed based on the target features, and the feature sequences of different characters to be matched can be matched to obtain the similarity between the characters to be matched in the list of software to be matched and the characters to be matched in the preset standard information database, so as to determine the matching result corresponding to the software information to be matched based on the similarity.

[0187] Furthermore, in one specific embodiment, determining the similarity between characters to be matched can be achieved by constructing a floating-point sequence based on the target features corresponding to each character in the characters to be matched. It determines the first floating-point sequence corresponding to the characters to be matched in the software list and the second floating-point sequence corresponding to the characters to be matched in the preset standard information database, and constructs a dynamic programming matrix based on the sequence lengths of the first and second floating-point sequences. Then, the dynamic programming matrix is ​​traversed, the maximum similarity sum in the dynamic programming matrix is ​​extracted, and the maximum similarity sum is normalized to obtain the similarity between the characters to be matched in the software list and the characters to be matched in the preset standard information database.

[0188] like Figure 7 As shown, the process first converts the string to a floating-point number sequence. When obtaining the character-level representation vector, this embodiment employs a traversal processing mechanism. That is, for each input string, it processes the characters sequentially to obtain the corresponding 8-dimensional comprehensive representation vector for each character. For example, for the string... Obtain vector sequence Each of them It is a character The sound-form-meaning fusion feature vector is then used. Then, during the sequence concatenation process, the 8-dimensional vectors of all characters are flattened sequentially and concatenated into a continuous floating-point sequence. :

[0189] ;

[0190] The resulting floating-point sequence The length is ; Let be the j-th dimension feature value of the i-th character. And in the above sequence, the length of the floating-point sequence of string 1 is: ; Length of the floating-point sequence of string 2: Furthermore, each of the eight consecutive floating-point numbers in the sequence corresponds to one character in the original string.

[0191] Furthermore, in this embodiment, for two floating-point values... and Their similarity distance Defined as:

[0192] ;

[0193] Among them, when hour: , indicates perfect similarity; when hour: This indicates complete dissimilarity. And... .

[0194] Furthermore, the aforementioned similarity distance, i.e., the complement of the absolute difference, converts the absolute difference between two values ​​into a similarity metric; the smaller the difference, the higher the similarity. And since each dimension of the comprehensive representation vector is within the range [0,1], therefore... This ensures the reasonableness of the similarity distance.

[0195] Furthermore, this embodiment calculates similarity based on a dynamic programming matrix. First, it initializes the matrix and sets boundary conditions: a matrix of size [size missing] is created. Dynamic programming matrix ,in:

[0196] ;

[0197] For all ;

[0198] For all ;

[0199] The core recursive formula is:

[0200] For each position ,in and :

[0201] ;

[0202] The calculation process is as follows: First, traverse the matrix in row-major or column-major order. For each position, calculate the payoff of three possible operations: Insertion operation: ignore the j-th element of sequence 2, and the similarity sum remains unchanged; Deletion operation: ignore the i-th element of sequence 1, and the similarity sum remains unchanged; Matching operation: match the i-th element of sequence 1 with the j-th element of sequence 2, increasing the similarity. Then select the operation with the highest return to update the current cell. Computational complexity: Time complexity Space complexity . Let i be the i-th floating-point feature value of sequence 1; The j-th floating-point feature value of sequence 2; the insertion operation ignores the j-th element of sequence 2, similarity and inheritance The deletion operation ignores the i-th element of sequence 1, similarity, and continuation. Matching operation comparison and Similarity sum = previous optimal value + Then, select the operation with the highest similarity sum among the three operations and update. .

[0203] After the dynamic programming calculation is completed, the element in the lower right corner of the matrix It contains the maximum similarity sum of two floating-point sequences:

[0204] ;

[0205] Then, the maximum similarity sum is normalized using the length of the longer sequence:

[0206] ;

[0207] Among them, the molecular range is: Denominator selection: Ensure the similarity score is within the range [0,1]. When two sequences are completely identical: When the two sequences are completely different: .

[0208] Furthermore, to ensure the rationality of normalization, a length penalty mechanism is implemented. This involves dividing by the length of the longer sequence, naturally penalizing length differences and ensuring the fairness of similarity scores. Regarding the value range: due to the similarity at each position... And can only match a maximum of There are 1 position, therefore:

[0209] .

[0210] In this way, the algorithm can convert a single character into an 8-bit floating-point number. By comparing the floating-point numbers, the similarity related to human understanding contained in the characters can be calculated, so that the similarity can penetrate traditional computer encoding and obtain results that are more in line with human intuition.

[0211] In another specific embodiment, determining the similarity between characters to be matched can be achieved by constructing a binary encoding sequence based on the target features corresponding to each character in the characters to be matched, and constructing a first position sequence and a second position sequence based on the binary encoding sequence; wherein the first position sequence is used to represent the position of 1 in the binary encoding sequence, and the second position sequence is used to represent the position of 0 in the binary encoding sequence; then, the characters to be matched in the software list to be matched are matched with the first position sequence and the second position sequence of the characters to be matched in the preset standard information database, respectively, so as to determine the similarity between the characters to be matched in the software list to be matched and the characters to be matched in the preset standard information database based on the obtained matching results.

[0212] Specifically, such as Figure 8 As shown, this embodiment requires the conversion of the comprehensive representation vector to binary code. First, a discretization encoding process is performed to discretize the vector, converting the 8-dimensional comprehensive representation vector, i.e., the floating-point number with each dimension in the range [0,1], into a multi-bit discrete binary code.

[0213] For each dimension of the vector Use uniform quantization:

[0214] ;

[0215] in Let i be a discrete integer representing the i-th feature, where i = 1 to 8, corresponding to each dimension of the 8-dimensional composite vector; The i-th dimension of the 8-dimensional comprehensive representation vector is a floating-point number. The number of bits allocated to each dimension, for a total code length of... Bit; To quantize precision, when k=4, the floating-point number in each dimension is divided into 16 levels; This is the floor function.

[0216] Then, binary encoding and concatenation are performed, converting the discrete values ​​of each dimension into binary and concatenating them:

[0217] ;

[0218] in Indicates using Bits represent integers. .

[0219] Furthermore, in this embodiment, the encoding table can be pre-calculated, that is, the comprehensive representation code of all characters in the character set is calculated in advance, and a mapping table from characters to binary codes is established. :

[0220] ;

[0221] For all Chinese character sets to be processed, It is an 8k-bit binary string. And the binary code of each character can be... Byte storage significantly reduces memory usage.

[0222] Then, a string-to-binary bit string conversion is performed. During the bit string generation process, the input string is... The binary code of each character is obtained by looking up a table. :

[0223] ;

[0224] Then, bit string concatenation is performed, concatenating the binary codes of all characters in order to form a continuous bit string:

[0225] ;

[0226] in This indicates a binary string concatenation operation. The length of a single character in the above bit string is... Bits, string bit length And each bit in the bit string Each bit corresponds to a character in the original string.

[0227] Next, the position information in the bit string is extracted. When extracting the Bit1 position list, a traversal scan is performed, scanning the bit string from left to right, recording the position indices of all bits with a value of 1:

[0228] ;

[0229] The position indexing rule is that the position starts counting from 0, corresponding to the leftmost end of the bit string.

[0230] When extracting the Bit0 position list, complementary processing is performed, and the bit string is scanned to record the position indices of all bits with a value of 0:

[0231] ;

[0232] In the list above, mutual exclusion and completeness are as follows:

[0233] ;

[0234] ;

[0235] Furthermore, both position lists are strictly increasing ordered sequences, laying the foundation for efficient subsequent computation.

[0236] Next, Bit1 similarity is calculated for the list of Bit1 positions of the two strings. and Initialize pointer Similarity and ;when and hour:

[0237] if :

[0238] ;

[0239] ;

[0240] Otherwise if :

[0241] ;

[0242] otherwise:

[0243] ;

[0244] Bit0 similarity calculation uses symmetric processing for the list of Bit0 positions. and Perform the same two-pointer algorithm to obtain the similarity and The algorithm complexity is as follows, calculating Bit1: Bit0 calculation: Total complexity: ,in Given the lengths of the two bit strings. Compared to dynamic programming for floating-point strings, the complexity is reduced from... Reduce to It achieves exponential acceleration and is suitable for large-scale data matching scenarios.

[0245] Then, the similarity scores can be summed. You can directly add the similarity scores of Bit1 and Bit0 together:

[0246] ;

[0247] The maximum similarity and the length of the shorter of the two bit strings are:

[0248] ;

[0249] Then, length normalization is performed, specifically using the length of the longer bit string for normalization:

[0250] ;

[0251] Minimum value: 0, i.e., no matching bits; Maximum value: This means that the shorter string length is completely matched. In this way, the floating-point vector can be converted into a binary encoded string through uniform quantization. Based on the characteristics of binary strings, a high-efficiency algorithm can be designed to significantly reduce space usage and improve computational efficiency while maintaining accuracy, thus meeting the requirements for similarity calculation scenarios of big data, long text, and streaming data.

[0252] Based on the above technical solution, this embodiment performs three-dimensional feature construction during the character feature construction stage. Phonetic feature construction is based on the BERT-wwm pre-trained model, employing a unified encoding strategy for multiple pronunciations. It fuses different pronunciation information through an attention mechanism, ultimately reducing the dimensionality to an 8-dimensional feature vector, effectively handling polyphonic characters and capturing the pronunciation characteristics of characters. Shape feature construction utilizes computer vision technology, extracting Chinese character structural features through multi-font character image generation and an improved ResNet-18 network architecture. Combined with channel attention mechanisms and multi-scale feature fusion, it ensures the discriminativeness and robustness of shape features. Semantic feature construction is based on a pre-trained language model, extracting character semantic information from the context window. Through multi-level feature extraction and attention enhancement, it captures the semantic meaning of characters in specific contexts, reducing the dimensionality to an 8-dimensional semantic feature vector. Then, a two-layer fully connected network is used to reduce the 24-dimensional phonetic-shape-semantic concatenation features to an 8-dimensional comprehensive representation vector. A Sigmoid activation function ensures the output value is within the range [0,1], providing standardized input for subsequent similarity calculations. Furthermore, the similarity calculation engine can dynamically adopt two modes: a precise calculation mode based on dynamic programming algorithms, performing fine-grained comparisons on floating-point vector sequences to ensure accuracy, suitable for scenarios with high precision requirements; and a fast calculation mode that optimizes binary encoding, reducing the complexity of similarity calculation from O(mn) to O(m+n), significantly improving computational efficiency, suitable for large-scale data matching scenarios. Therefore, this embodiment significantly improves matching accuracy, is particularly suitable for strings containing various human errors, and its computational efficiency meets the processing needs of millions of data points. The system is highly adaptable and can be extended to other short text similarity calculation scenarios. Moreover, through systematic multi-dimensional feature design and hierarchical optimization strategies, existing open-source data and corpora can be directly used, effectively improving accuracy, computational efficiency, and discrimination stability.

[0253] See Figure 9 As shown in the illustration, this application also discloses a drone software information matching device, comprising:

[0254] The feature construction module 11 is used to determine the software information to be matched in the list of software to be matched and the preset standard information database, respectively, and extract the characters to be matched in the software information to be matched, and construct the first feature, the second feature and the third feature corresponding to the characters to be matched; the first feature is used to characterize the pronunciation feature of the characters to be matched, the second feature is used to characterize the structural feature of the characters to be matched, and the third feature is used to characterize the semantic feature of the characters to be matched.

[0255] Feature splicing module 12 is used to splice the first feature, the second feature and the third feature using a preset fully connected network to obtain the target feature;

[0256] The feature matching module 13 is used to construct a corresponding feature sequence based on the target feature, and to match the feature sequences of different characters to be matched to obtain the similarity between the characters to be matched in the list of software to be matched and the characters to be matched in the preset standard information database, so as to determine the matching result corresponding to the software information to be matched based on the similarity.

[0257] This embodiment identifies the software list to be matched and the software information in a preset standard information database, extracts the core characters to be matched, and then extracts three types of features for each character: phonetic features, morphological features, and semantic features. A fully connected network is then used to concatenate and fuse the phonetic, morphological, and semantic features of the same character, integrating multi-dimensional features to avoid the limitations of single features. Based on the target features of each character, a corresponding feature sequence is constructed. By comparing the feature sequences of different software information, the similarity is calculated, and finally, the matching of the software list to be matched and the software information in the preset standard information database is completed based on the similarity. In this way, this embodiment can achieve accurate character recognition and matching by extracting phonetic, morphological, and semantic features, solving the problem of missed or incorrect matching caused by single-feature matching. Furthermore, by extracting three-dimensional character features, the core features of each character can be captured, and by fusing them through a fully connected network, the target features contain complete character features, improving the robustness of matching in complex scenarios and significantly improving the matching accuracy and reliability.

[0258] In some specific embodiments, the feature construction module 11 specifically includes:

[0259] A pronunciation sequence construction unit is used to determine several pronunciations corresponding to the character to be matched, and to construct an initial sequence based on the several pronunciations and the character to be matched;

[0260] The first feature extraction unit is used to extract the first hidden state features of each of the pronunciations and the characters to be matched in the initial sequence using a pre-trained language model;

[0261] The first feature fusion unit is used to construct a first fusion feature between each of the pronunciations based on the first hidden state features using a self-attention mechanism, and to construct a cross feature between the pronunciation and the character to be matched based on the first hidden state features using a cross-attention mechanism.

[0262] The second feature fusion unit is used to fuse the first fusion feature and the cross feature based on a first preset weight to obtain the first feature corresponding to the character to be matched.

[0263] In some specific embodiments, the feature construction module 11 specifically includes:

[0264] An image processing unit is used to acquire initial character images corresponding to the characters to be matched in several character formats, and to preprocess the initial character images based on a preset image enhancement strategy to obtain a target character image; the character image is a grayscale image.

[0265] The second feature extraction unit is used to extract the output multi-scale fusion features of the target character image using a preset deep neural network.

[0266] The feature pooling unit is used to perform global average pooling and global max pooling on the multi-scale fused features respectively to obtain the corresponding pooling results;

[0267] A feature determination unit is used to determine the second feature corresponding to the character to be matched based on the pooling result.

[0268] In some specific embodiments, the feature construction module 11 specifically includes:

[0269] A semantic sequence construction unit is used to determine the preset window radius corresponding to the character to be matched, and to construct a window sequence based on the character to be matched and the context content corresponding to the character to be matched according to the preset window radius;

[0270] The third feature extraction unit is used to extract the second hidden state features of each character in the window sequence using a pre-trained language model;

[0271] The fourth feature extraction unit is used to determine the target position of the character to be matched in the window sequence and extract the fourth feature of the target position from the second hidden state features;

[0272] Based on the second preset weight, the second hidden state features corresponding to each character in the window sequence are fused to obtain the fifth feature;

[0273] The third feature fusion unit is used to encode the location information corresponding to the target location to obtain the sixth feature;

[0274] The fourth feature fusion unit is used to fuse the fourth feature, the fifth feature and the sixth feature to obtain the third feature corresponding to the character to be matched.

[0275] In some specific embodiments, the feature splicing module 12 specifically includes:

[0276] A feature splicing unit is used to directly splice the first feature, the second feature, and the third feature to obtain a second fused feature;

[0277] The feature processing unit is used to normalize the second fused feature to obtain the target vector;

[0278] The feature dimensionality reduction unit is used to reduce the dimensionality of the target vector sequentially using two fully connected layers of the preset fully connected network to obtain the target features.

[0279] In some specific embodiments, the feature matching module 13 specifically includes:

[0280] The first sequence construction unit is used to construct a floating-point number sequence based on the target features corresponding to each character in the characters to be matched;

[0281] A matrix construction unit is used to determine the first floating-point number sequence corresponding to the character to be matched in the software list to be matched, and the second floating-point number sequence corresponding to the character to be matched in the preset standard information database, and to construct a dynamic programming matrix based on the sequence lengths of the first floating-point number sequence and the second floating-point number sequence;

[0282] The first similarity calculation unit is used to traverse the dynamic programming matrix, extract the maximum similarity sum in the dynamic programming matrix, and normalize the maximum similarity sum to obtain the similarity between the character to be matched in the software list and the character to be matched in the preset standard information database.

[0283] In some specific embodiments, the feature matching module 13 specifically includes:

[0284] The second sequence construction unit is used to construct a binary encoding sequence based on the target features corresponding to each character in the characters to be matched;

[0285] The third sequence construction unit is used to construct a first position sequence and a second position sequence based on the binary encoded sequence; the first position sequence is used to represent the position of 1 in the binary encoded sequence, and the second position sequence is used to represent the position of 0 in the binary encoded sequence.

[0286] The second similarity calculation unit is used to match the character to be matched in the software list to be matched with the first position sequence and the second position sequence of the character to be matched in the preset standard information database, respectively, so as to determine the similarity between the character to be matched in the software list to be matched and the character to be matched in the preset standard information database based on the obtained matching result.

[0287] Furthermore, embodiments of this application also disclose an electronic device, Figure 10 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0288] Figure 10 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the UAV software information matching method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0289] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0290] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0291] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the UAV software information matching method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include a computer program capable of performing other specific tasks.

[0292] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned UAV software information matching method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0293] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0294] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0295] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0296] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0297] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for matching software information of unmanned aerial vehicles (UAVs), characterized in that, include: The software to be matched list and the software information to be matched in the preset standard information database are determined respectively, and the characters to be matched in the software information are extracted to construct the first feature, the second feature and the third feature corresponding to the characters to be matched; The first feature is used to characterize the pronunciation feature of the character to be matched, the second feature is used to characterize the structural feature of the character to be matched, and the third feature is used to characterize the semantic feature of the character to be matched. The target feature is obtained by concatenating the first feature, the second feature, and the third feature using a preset fully connected network. Based on the target features, a corresponding feature sequence is constructed, and the feature sequences of different characters to be matched are matched to obtain the similarity between the characters to be matched in the list of software to be matched and the characters to be matched in the preset standard information database, so as to determine the matching result corresponding to the software information to be matched based on the similarity. The construction of the first feature corresponding to the character to be matched includes: Determine several pronunciations corresponding to the character to be matched, and construct an initial sequence based on the several pronunciations and the character to be matched; The first hidden state features of each pronunciation and the character to be matched in the initial sequence are extracted using a pre-trained language model. Based on the self-attention mechanism, a first fusion feature is constructed between each of the pronunciations according to the first hidden state feature; and based on the cross-attention mechanism, a cross feature is constructed between the pronunciation and the character to be matched according to the first hidden state feature. The first fusion feature and the cross feature are fused based on the first preset weight to obtain the first feature corresponding to the character to be matched; Furthermore, the step of constructing a corresponding feature sequence based on the target features and matching the feature sequences of different characters to be matched to obtain the similarity between the characters to be matched in the software list and the characters to be matched in the preset standard information database includes: Construct a binary encoding sequence based on the target features corresponding to each character in the characters to be matched; A first position sequence and a second position sequence are constructed based on the binary encoded sequence; the first position sequence is used to represent the position of 1 in the binary encoded sequence, and the second position sequence is used to represent the position of 0 in the binary encoded sequence. The first position sequence and the second position sequence of the character to be matched in the software list to be matched are respectively matched with the first position sequence and the second position sequence of the character to be matched in the preset standard information database, so as to determine the similarity between the character to be matched in the software list to be matched and the character to be matched in the preset standard information database based on the obtained matching results.

2. The UAV software information matching method according to claim 1, characterized in that, Construct the second feature corresponding to the character to be matched, including: A set of initial character images corresponding to the characters to be matched in several character formats are obtained, and the initial character images are preprocessed based on a preset image enhancement strategy to obtain the target character image; the character image is a grayscale image. The multi-scale fusion features of the target character image are extracted using a preset deep neural network. Global average pooling and global max pooling are performed on the multi-scale fusion features respectively to obtain the corresponding pooling results; The second feature corresponding to the character to be matched is determined based on the pooling result.

3. The UAV software information matching method according to claim 1, characterized in that, Construct the third feature corresponding to the character to be matched, including: Determine the preset window radius corresponding to the character to be matched, and construct a window sequence based on the character to be matched and the context content corresponding to the character to be matched according to the preset window radius; The second hidden state features of each character in the window sequence are extracted using a pre-trained language model; Determine the target position of the character to be matched in the window sequence, and extract the fourth feature of the target position from the second hidden state features; Based on the second preset weight, the second hidden state features corresponding to each character in the window sequence are fused to obtain the fifth feature; The location information corresponding to the target location is encoded to obtain the sixth feature; The fourth feature, the fifth feature, and the sixth feature are combined to obtain the third feature corresponding to the character to be matched.

4. The UAV software information matching method according to claim 1, characterized in that, The step of using a preset fully connected network to concatenate the first feature, the second feature, and the third feature to obtain the target feature includes: The second fused feature is obtained by directly concatenating the first feature, the second feature, and the third feature. The second fused feature is normalized to obtain the target vector; The target vector is reduced in dimensionality by using two fully connected layers of the preset fully connected network to obtain the target features.

5. A device for matching software information of unmanned aerial vehicles (UAVs), characterized in that, include: The feature construction module is used to determine the software information to be matched in the list of software to be matched and the preset standard information database, respectively, and extract the characters to be matched in the software information to be matched, and construct the first feature, the second feature and the third feature corresponding to the characters to be matched; The first feature is used to characterize the pronunciation feature of the character to be matched, the second feature is used to characterize the structural feature of the character to be matched, and the third feature is used to characterize the semantic feature of the character to be matched. The feature splicing module is used to splice the first feature, the second feature, and the third feature using a preset fully connected network to obtain the target feature; The feature matching module is used to construct a corresponding feature sequence based on the target feature, and to match the feature sequences of different characters to be matched to obtain the similarity between the characters to be matched in the list of software to be matched and the characters to be matched in the preset standard information database, so as to determine the matching result corresponding to the software information to be matched based on the similarity. The feature construction module includes: A pronunciation sequence construction unit is used to determine several pronunciations corresponding to the character to be matched, and to construct an initial sequence based on the several pronunciations and the character to be matched; The first feature extraction unit is used to extract the first hidden state features of each of the pronunciations and the characters to be matched in the initial sequence using a pre-trained language model; The first feature fusion unit is used to construct a first fusion feature between each of the pronunciations based on the first hidden state features using a self-attention mechanism, and to construct a cross feature between the pronunciation and the character to be matched based on the first hidden state features using a cross-attention mechanism. The second feature fusion unit is used to fuse the first fusion feature and the cross feature based on a first preset weight to obtain the first feature corresponding to the character to be matched; Furthermore, the feature matching module includes: The second sequence construction unit is used to construct a binary encoding sequence based on the target features corresponding to each character in the characters to be matched; The third sequence construction unit is used to construct a first position sequence and a second position sequence based on the binary encoded sequence; the first position sequence is used to represent the position of 1 in the binary encoded sequence, and the second position sequence is used to represent the position of 0 in the binary encoded sequence. The second similarity calculation unit is used to match the character to be matched in the software list to be matched with the first position sequence and the second position sequence of the character to be matched in the preset standard information database, respectively, so as to determine the similarity between the character to be matched in the software list to be matched and the character to be matched in the preset standard information database based on the obtained matching result.

6. An electronic device, characterized in that, The electronic device includes a processor and a memory; wherein the memory is used to store a computer program, which is loaded and executed by the processor to implement the UAV software information matching method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, Used to store a computer program, which, when executed by a processor, implements the UAV software information matching method as described in any one of claims 1 to 4.

Images