Method and system for distributing intelligence on demand

Abstract

The present invention discloses a method and system for on-demand intelligence distribution, comprising the following steps: (1) receiving superior telegraphic information to obtain action tasks and intelligence information; (2) extracting features: commander features, action task features, and intelligence features; (3) error processing and model optimization; and (4) outputting intelligence in order of interest based on the optimized model. The present invention preprocesses task and intelligence text using an improved linear discriminant analysis method, encodes task and intelligence data using a sparse edge denoising autoencoder, and uses a matrix decomposition algorithm to remove blank portions of the matrix for dimensionality reduction, thereby effectively improving the accuracy of the training model, while also increasing the precision of on-demand intelligence distribution and enhancing the command efficiency of commanders.

Description

Technical Field

[0001] The present invention belongs to the field of command and control technology, and specifically relates to an on-demand intelligence distribution method and system. Background Art

[0002] In military, civil air defense, and emergency rescue operations, securing the information high ground is a key factor in successful operational command. With the rapid development of situational awareness technology, commanding actual operational missions requires the collection and processing of massive amounts of intelligence data and information. However, humans are limited in the amount of information they can process in a given timeframe. These limitations restrict commanders' ability to perceive the on-site situation and make command decisions. Furthermore, commanders with different responsibilities and duties under different operational missions focus on different intelligence information during the command process. To quickly and accurately deliver appropriate intelligence information to commanders with different tasks and responsibilities, we need to build deep learning model algorithms based on the responsibilities and operational characteristics of different commanders, as well as their historical operational behaviors with different intelligence. This algorithm can then mine the intelligence information required by different commanders under different operational missions and deliver it precisely on demand.

[0003] Existing research on intelligence distribution methods primarily focuses on user-based publish / subscribe models or rule-based intelligent intelligence distribution algorithm models. These methods suffer from issues such as extensive user subscription models, complex rule-making, and insufficient flexibility, resulting in a significant gap between the intelligence distributed and the intelligence commanders need. Furthermore, there is insufficient consideration of how to identify users' potential needs based on historical missions and distribute intelligence on demand. Summary of the Invention

[0004] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a method and system for on-demand intelligence distribution to solve the problem that traditional intelligence distribution cannot fully tap the interest and potential needs of commanders with different responsibilities for different intelligence under different operational tasks, thereby resulting in inaccurate and inefficient intelligence distribution.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] The method for distributing intelligence on demand of the present invention comprises the following steps:

[0007] (1) Receive superior telegraphic information to obtain operational tasks and intelligence information;

[0008] (2) Feature extraction: command personnel features, operational mission features, and intelligence features;

[0009] (3) Error processing and model optimization;

[0010] (4) According to the optimized model, the information is sorted and outputted according to the interest level.

[0011] Furthermore, in the step (1), the action tasks issued by the superior are obtained through the document information; and the intelligence information in the intelligence system is obtained in real time.

[0012] Furthermore, the feature extraction in step (2) includes: extracting command personnel features from the command system, extracting action task features from document information, and extracting intelligence features from intelligence information.

[0013] Furthermore, the commander characteristics include: the command organization to which the commander belongs and the commander's seat.

[0014] Furthermore, the action task characteristics include: action task type, action task level, action task area, action task content, action task means, target type, target level, and target scale involved in the action task.

[0015] Furthermore, the intelligence features include: intelligence type, intelligence region, intelligence generation time, and commander's operational behavior on the intelligence; divided by intelligence type: environmental information, disaster information, troop deployment, rescue force, important targets, supplies and equipment, medical information, population information, traffic information, and communication information; divided by commander's operational behavior on the intelligence, i.e., interest level: no operation, selection, marking, and output.

[0016] Furthermore, the feature extraction steps in step (2) include:

[0017] (21) Encode the command personnel's organization and command personnel's seat to form a matrix M i 、N i , M i N represents the coding matrix of the organization to which the commander belongs in the i-th data label, i The coding matrix representing the commander's seat in the i-th data tag;

[0018] (22) The improved linear discriminant analysis method is used to preprocess the action task text and intelligence text to determine the importance of each word and obtain the preprocessed word vector; and the sparse edge denoising autoencoder is used to perform text encoding on the preprocessed word vector to obtain the matrix P i , Q i , P i represents the encoding matrix of the action task text in the i-th data label, Q i Represents the encoding matrix of the intelligence text in the i-th data label;

[0019] (23) According to the commander's operation behavior on intelligence, the interest degree of intelligence is assigned and coded to form a matrix Ri =[α], R i is the interest in intelligence in the i-th data label, where α = 0, 1, 2, 3, representing no operation, selection, labeling, and output, respectively;

[0020] (24) Use matrix decomposition algorithm to remove the matrix (M i ,N i ,P i ) and the matrix (Q i ,R i ) in the blank space.

[0021] Furthermore, the improved linear discriminant analysis method is specifically as follows:

[0022] (221) Solve the intra-class dispersion matrix S of the linear discriminant analysis method w and the inter-class dispersion matrix S b :

[0023]

[0024]

[0025] Where x m represents the mth class sample, N m represents the number of samples at m, N represents the total number of samples, x mn represents the nth sample in the mth class sample, c represents the number of sample categories, and T represents the transpose of the matrix;

[0026] (222) The Fisher criterion is changed to the following formula:

[0027]

[0028] Where, is an n-dimensional column vector, is the minimum scatter matrix difference criterion function;

[0029] (223) Solve the following characteristic equation to find the eigenvectors corresponding to the d smallest eigenvalues:

[0030] (S w -S b )α k =λ k α k ,k=1,2,…,d

[0031] Where λ k is the kth smallest eigenvalue, α k is the eigenvalue λ k The corresponding eigenvector.

[0032] Furthermore, the step (24) specifically includes:

[0033] (241) According to the certain correlation between the intelligence concerned by the same commander under the same action task, the co-occurrence matrix R of the action task text and the intelligence text is constructed. st}=(P s ,Q t ), where r st represents the number of times the intelligence text t appears in the sliding window centered on the action task s; each row of the matrix R is truncated to retain only the K non-zero r with the largest value st , the rest r st Set to zero;

[0034] (242) Note x s is the vector when the action task s is located at the center of the sliding window, y t The vector representing the intelligence text t when it is not in the center of the sliding window can be constructed using the following matrix decomposition model formula:

[0035]

[0036] Where c st =αlog(1+r st ), α and λ are hyper parameters, and the action task matrix P s =[x1,x2,...x s ...,x M ], intelligence text matrix Q t =[y1,y2,...y t ...,y M ], solve the objective function above and obtain the similarity between the action task text and the intelligence text.

[0037] Furthermore, the step (3) specifically includes:

[0038] (31) The model consists of three convolutional layers, three pooling layers, and one fully connected layer. The ReLU function is used as the activation function in the convolutional layer, and the last layer of the model is a fully connected layer, which divides the predicted intelligence results into four categories for output.

[0039] (32) Command personnel and operational tasks are used as the input of the model, and intelligence is used as the output of the model; a multi-layer sparse edge denoising autoencoder network is formed by stacking, and each layer of the sparse edge denoising autoencoder is trained hierarchically, and then the parameters of the entire model are adjusted to achieve model optimization;

[0040] (33) All historical task sample data are put into the model for training to obtain the optimal model; there is no initial error when the model is trained for the first time, so an initial error is given. After one training, the error between the model output and the actual situation is calculated and used as the correction error for the next training. It is continuously iterated to converge within the specified accuracy, and the optimal parameters are obtained, that is, the optimal model is obtained.

[0041] Furthermore, the error processing model optimization method is as follows:

[0042] (321) The autoencoder is a three-layer neural network consisting of an encoder and a decoder, which are respectively an input layer h, a hidden layer w, and an output layer v. If the encoding function is f and the decoding function is g, the encoder f maps the input layer h to the hidden layer w, which can be expressed as:

[0043] w=f(h)=s f (Wh+b)

[0044] Where s f is the encoder activation function, W is the encoder weight matrix, and b is the encoder bias vector; the decoder g maps the hidden layer w to the output layer v, which can be expressed as:

[0045] v=g(w)=s g (W′w+b′)

[0046] Where s g is the decoder activation function, W′ is the decoder weight matrix, and b′ is the decoder bias vector; the reconstruction error L(h,v) measures the similarity between the input layer h and the output layer v:

[0047] L(h,v)=||hv|| 2

[0048] The overall loss function of the autoencoder is:

[0049]

[0050] Where s is the training sample set;

[0051] (322) The loss function of the sparse autoencoder (SAE) is:

[0052]

[0053] In the formula, β is the weight coefficient, ρ is the parameter, is the activation degree of the neuron;

[0054] (323)The loss function of the edge denoising autoencoder is:

[0055]

[0056] Where, is the variance of the interference term of the dth dimension of the input data;

[0057] (324) The loss function formula of the sparse edge denoising autoencoder is as follows:

[0058] SmDAE (θ)=L(h,v)+J SAE +J mDAE +Y

[0059]

[0060] Where y i is the total number of real intelligence under the i-th type of action mission, is the total number of accurately predicted intelligence under the i-th action task, r is the feedback correction factor, and the loss function J of the sparse edge denoising autoencoder is optimized by training the network. SmDAE .

[0061] Furthermore, the step (4) generates a recommendation list by sorting all the acquired intelligence according to interest, with intelligence with high interest being arranged at the front; and pushes the recommendation list to the corresponding commander.

[0062] The present invention also provides an intelligence on-demand distribution system, which comprises the following steps:

[0063] The acquisition module is used to receive superior document and telegraphic information to obtain action tasks and intelligence information;

[0064] Feature extraction module, used for command personnel features, action mission features, and intelligence features;

[0065] Model optimization module, used for error processing and model optimization;

[0066] The output module is used to output the intelligence in order of interest based on the optimized model.

[0067] Beneficial effects of the present invention:

[0068] The present invention preprocesses task and intelligence texts through an improved linear discriminant analysis method (ILDA), encodes task and intelligence data using a sparse edge denoising autoencoder (SmDAE), and uses a matrix decomposition algorithm to remove the blank parts of the matrix for dimensionality reduction, thereby effectively improving the accuracy of the training model, while improving the accuracy of on-demand intelligence distribution and enhancing the command efficiency of commanders. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 Schematic diagram of the method of the present invention;

[0070] Figure 2 This is the overall architecture diagram of the training model. DETAILED DESCRIPTION

[0071] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and drawings. The contents mentioned in the embodiments are not intended to limit the present invention.

[0072] Reference Figure 1 As shown, the method for distributing intelligence on demand of the present invention comprises the following steps:

[0073] (1) Receive superior documents and telegrams to obtain action tasks and intelligence information; specifically, obtain action tasks issued by superiors through documents and telegrams; obtain intelligence information from the intelligence system in real time.

[0074] (2) Feature extraction: command personnel features, operational mission features, and intelligence features;

[0075] The feature extraction in step (2) includes: extracting command personnel features from the command system background, extracting action task features from document information, and extracting intelligence features from intelligence information.

[0076] In a specific example, the commander characteristics include: the command organization to which the commander belongs and the commander's seat.

[0077] The characteristics of the action task include: action task type, action task level, action task area, action task content, action task means, target type, target level, and target scale involved in the action task.

[0078] The intelligence features include: intelligence type, intelligence region, intelligence generation time, and commanders' operational behavior on the intelligence; intelligence type is divided into: environmental information, disaster information, troop deployment, rescue force, important targets, supplies and equipment, medical information, population information, traffic information, and communication information; intelligence operational behavior, i.e., interest level, is divided into: no operation, selection, marking, and output.

[0079] The feature extraction steps in step (2) include:

[0080] (21) Encode the command personnel's organization and command personnel's seat to form a matrix M i 、N i , M i N represents the coding matrix of the organization to which the commander belongs in the i-th data label, i The coding matrix representing the commander's seat in the i-th data tag;

[0081] (22) The improved linear discriminant analysis (ILDA) method is used to preprocess the action task text and intelligence text to determine the importance of each word and obtain the preprocessed word vector; and the sparse edge denoising autoencoder (SmDAE) is used to encode the preprocessed word vector to obtain the matrix P i , Q i , P i represents the encoding matrix of the action task text in the i-th data label, Q i Represents the encoding matrix of the intelligence text in the i-th data label;

[0082] (23) According to the commander's operation behavior on intelligence, the interest degree of intelligence is assigned and coded to form a matrix R i =[α], R i is the interest in intelligence in the i-th data label, where α = 0, 1, 2, 3, representing no operation, selection, labeling, and output, respectively;

[0083] (24) Use matrix decomposition algorithm to remove the matrix (M i ,N i ,P i ) and the matrix (Q i ,R i ) in the blank space.

[0084] More specifically, the improved linear discriminant analysis method is as follows:

[0085] (221) Solve the intra-class dispersion matrix S of the linear discriminant analysis method w and the inter-class dispersion matrix S b :

[0086]

[0087]

[0088] Where x m represents the mth class sample, N m represents the number of samples at m, N represents the total number of samples, x mn represents the nth sample in the mth class sample, c represents the number of sample categories, and T represents the transpose of the matrix;

[0089] (222) The Fisher criterion is changed to the following formula:

[0090]

[0091] Where, is an n-dimensional column vector, is the minimum scatter matrix difference criterion function;

[0092] (223) Solve the following characteristic equation to find the eigenvectors corresponding to the d smallest eigenvalues:

[0093] (S w -S b )α k =λ k α k ,k=1,2,…,d

[0094] Where λ k is the kth smallest eigenvalue, α k is the eigenvalue λ k The corresponding eigenvector.

[0095] The step (24) specifically includes:

[0096] (241) According to the certain correlation between the intelligence concerned by the same commander under the same action task, the co-occurrence matrix R of the action task text and the intelligence text is constructed. st}=(P s ,Q t ), where r st represents the number of times the intelligence text t appears in the sliding window centered on the action task s; each row of the matrix R is truncated to retain only the K non-zero r with the largest value st , the rest r st Set to zero;

[0097] (242) Note x s is the vector when the action task s is located at the center of the sliding window, y t The vector representing the intelligence text t when it is not in the center of the sliding window can be constructed using the following matrix decomposition model formula:

[0098]

[0099] Where c st =αlog(1+r st ), α and λ are hyper parameters, and the action task matrix P s =[x1,x2,…x s ...,x M ], intelligence text matrix Q t =[y1,y2,…y t ...,y M ], solve the objective function above and obtain the similarity between the action task text and the intelligence text (this method overcomes the limitation of the sparsity of user behavior data).

[0100] (3) Error processing and model optimization;

[0101] (31) The overall architecture of the model is as follows Figure 2As shown in the figure, the on-demand intelligence distribution model consists of three convolutional layers, three pooling layers, and a fully connected layer. The convolutional layer uses the ReLU function as the activation function (because the partial derivative of the ReLU function to the right of the origin is 1, it converges quickly and is efficient during backpropagation, thereby improving the efficiency of the entire model). The last layer of the model is a fully connected layer, which outputs the predicted intelligence results into four categories.

[0102] (32) Command personnel and operational tasks are used as the input of the model, and intelligence is used as the output of the model; a multi-layer sparse edge denoising autoencoder network is formed by stacking, and each layer of the sparse edge denoising autoencoder is trained hierarchically, and then the parameters of the entire model are adjusted to achieve model optimization;

[0103] (33) All historical task sample data are put into the model for training to obtain the optimal model; there is no initial error when the model is trained for the first time, so an initial error is given. After one training, the error between the model output and the actual situation is calculated and used as the correction error for the next training. It is continuously iterated to converge within the specified accuracy, and the optimal parameters are obtained, that is, the optimal model is obtained.

[0104] Among them, the error processing model optimization method (error processing) is as follows:

[0105] (321) The autoencoder is a three-layer neural network consisting of an encoder and a decoder, which are respectively an input layer h, a hidden layer w, and an output layer v. If the encoding function is f and the decoding function is g, the encoder f maps the input layer h to the hidden layer w, which can be expressed as:

[0106] w=f(h)=s f (Wh+b)

[0107] Where s f is the encoder activation function, W is the encoder weight matrix, and b is the encoder bias vector; the decoder g maps the hidden layer w to the output layer v, which can be expressed as:

[0108] v=g(w)=s g (W′w+b′)

[0109] Where s g is the decoder activation function, W′ is the decoder weight matrix, and b′ is the decoder bias vector; the reconstruction error L(h,v) measures the similarity between the input layer h and the output layer v:

[0110] L(h,v)=||hv|| 2

[0111] The overall loss function of the autoencoder is:

[0112]

[0113] Where s is the training sample set;

[0114] (322) The loss function of the sparse autoencoder (SAE) is:

[0115]

[0116] In the formula, β is the weight coefficient, ρ is the parameter, is the activation degree of the neuron;

[0117] (323) The loss function of the edge denoising autoencoder (mDAE) is:

[0118]

[0119] Where, is the variance of the interference term of the dth dimension of the input data;

[0120] (324) The loss function formula of the sparse edge denoising autoencoder (SmDAE) is as follows:

[0121] SmDAE (θ)=L(h,v)+J SAE +J mDAE +Y

[0122]

[0123] Where y i is the total number of real intelligence under the i-th type of action mission, is the total number of accurately predicted intelligence under the i-th action task, r is the feedback correction factor, and the loss function J of the sparse edge denoising autoencoder is optimized by training the network. SmDAE .

[0124] (4) All acquired intelligence is sorted according to interest level to generate a recommendation list, with intelligence with high interest level arranged first; and the recommendation list is pushed to the corresponding commander.

[0125] The present invention also provides an intelligence on-demand distribution system, which comprises the following steps:

[0126] The acquisition module is used to receive superior document and telegraphic information to obtain action tasks and intelligence information;

[0127] Feature extraction module, used for command personnel features, action mission features, and intelligence features;

[0128] Model optimization module, used for error processing and model optimization;

[0129] The output module is used to output the intelligence in order of interest based on the optimized model.

[0130] The present invention has many specific application paths. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the principles of the present invention. These improvements should also be considered as the scope of protection of the present invention.

Claims

1. A method for distributing intelligence on demand, characterized in that: Here are the steps: (1) Receive superior telegraphic information to obtain operational tasks and intelligence information; (2) Feature extraction: command personnel features, operational mission features, and intelligence features; (3) Error processing and model optimization; (4) Outputting the intelligence in order of interest based on the optimized model; The specific steps of feature extraction in step (2) include: (21) Encode the command personnel's organization and command personnel's seat to form a matrix M i 、N i , M i N represents the coding matrix of the organization to which the commander belongs in the i-th data label, i The coding matrix representing the commander's seat in the i-th data tag; (22) The improved linear discriminant analysis method is used to preprocess the action task text and intelligence text to determine the importance of each word and obtain the preprocessed word vector; and the sparse edge denoising autoencoder is used to perform text encoding on the preprocessed word vector to obtain the matrix P i , Q i , P i represents the encoding matrix of the action task text in the i-th data label, Q i Represents the encoding matrix of the intelligence text in the i-th data label; (23) According to the commander's operation behavior on intelligence, the interest degree of intelligence is assigned and coded to form a matrix R i =[α], R i is the interest level of the intelligence in the i-th data label, where α = 0, 1, 2, 3, representing no operation, selection, labeling, and output, respectively; (24) Use matrix decomposition algorithm to remove the matrix (M i , N i , P i ) and the matrix (Q i , R i ) in the blank space; The improved linear discriminant analysis method is specifically: (221) Solve the intra-class dispersion matrix S of the linear discriminant analysis method w and the inter-class dispersion matrix S b : Where x m represents the mth class sample, N m represents the number of samples at m, N represents the total number of samples, x mn represents the nth sample in the mth class sample, c represents the number of sample categories, and T represents the transpose of the matrix; (222) The Fisher criterion is changed to the following formula: Where, is an n-dimensional column vector, is the minimum scatter matrix difference criterion function; (223) Solve the following characteristic equation to find the eigenvectors corresponding to the d smallest eigenvalues: (S w -S b )a k =λ k a k ,k=1,2,…,d Where λ k is the kth smallest eigenvalue, α k is the eigenvalue λ k The corresponding eigenvector; The step (24) specifically includes: (241) According to the certain correlation between the intelligence concerned by the same commander under the same action task, the co-occurrence matrix R of the action task text and the intelligence text is constructed. st }=(P s , Q t ), where r st represents the number of times the intelligence text t appears in the sliding window centered on the action task s; each row of the matrix R is truncated to retain only the K non-zero r with the largest value st , the rest r st Set to zero; (242) Note x s is the vector when the action task s is located at the center of the sliding window, y t The vector representing the intelligence text t when it is not in the center of the sliding window can be constructed using the following matrix decomposition model formula: Where c st =αlog(1+r st ), α and λ are hyper parameters, and the action task matrix P s =[x1, x2, ...x s ..., x M ], intelligence text matrix Q t =[y1,y2,...y t ..., y M ], solve the objective function above and obtain the similarity between the action task text and the intelligence text; The step (3) specifically includes: (31) The model consists of three convolutional layers, three pooling layers, and one fully connected layer. The ReLU function is used as the activation function in the convolutional layer, and the last layer of the model is a fully connected layer, which divides the predicted intelligence results into four categories for output. (32) Command personnel and operational tasks are used as the input of the model, and intelligence is used as the output of the model; a multi-layer sparse edge denoising autoencoder network is formed by stacking, and each layer of the sparse edge denoising autoencoder is trained hierarchically, and then the parameters of the entire model are adjusted to achieve model optimization; (33) All historical task sample data are put into the model for training to obtain the optimal model; there is no initial error when the model is first trained, so an initial error is given. After one training, the error between the model output and the actual situation is calculated and used as the correction error for the next training. It is continuously iterated to converge within the specified accuracy, and the optimal parameters are obtained, that is, the optimal model is obtained; The error processing model optimization method is as follows: (321) The autoencoder is a three-layer neural network consisting of an encoder and a decoder, which are respectively an input layer h, a hidden layer w, and an output layer v. If the encoding function is f and the decoding function is g, the encoder maps the input layer h to the hidden layer w, and the feature representation is: w=f(h)=s f (Wh+b) Where s f is the encoder activation function, W is the encoder weight matrix, and b is the encoder bias vector; the decoder g maps the hidden layer w to the output layer v as follows: v=g(w)=s g (W′w+b′) Where s g is the decoder activation function, W′ is the decoder weight matrix, and b′ is the decoder bias vector; the reconstruction error L(h, v) measures the similarity between the input layer h and the output layer v: L(h,v)=||h-v|| 2 The overall loss function of the autoencoder is: Where s is the training sample set; (322) The loss function of the sparse autoencoder is: In the formula, β is the weight coefficient, ρ is the parameter, is the activation degree of the neuron; (323)The loss function of the edge denoising autoencoder is: Where, is the variance of the interference term of the dth dimension of the input data; (324) The loss function formula of the sparse edge denoising autoencoder is as follows: SmDAE(θ)=L(L,v)+J SAE +J mDAE +Y Where y i is the total number of real intelligence under the i-th type of action mission, is the total number of accurately predicted intelligence under the i-th action task, r is the feedback correction factor, and the loss function J of the sparse edge denoising autoencoder is optimized by training the network. SmDAE .

2. The method for distributing information on demand according to claim 1, wherein: The feature extraction in step (2) includes: extracting command personnel features from the command system, extracting action task features from document information, and extracting intelligence features from intelligence information.

3. An on-demand intelligence distribution system, characterized in that: Here are the steps: The acquisition module is used to receive superior document and telegraphic information to obtain action tasks and intelligence information; Feature extraction module, used for command personnel features, action mission features, and intelligence features; Model optimization module, used for error processing and model optimization; The output module is used to output the intelligence in order of interest based on the optimized model; The feature extraction module specifically includes: (21) Encode the command personnel's organization and command personnel's seat to form a matrix M i 、N i , M i N represents the coding matrix of the organization to which the commander belongs in the i-th data label, i The coding matrix representing the commander's seat in the i-th data tag; (22) The improved linear discriminant analysis method is used to preprocess the action task text and intelligence text to determine the importance of each word and obtain the preprocessed word vector; and the sparse edge denoising autoencoder is used to perform text encoding on the preprocessed word vector to obtain the matrix P i , Q i , P i represents the encoding matrix of the action task text in the i-th data label, Q i Represents the encoding matrix of the intelligence text in the i-th data label; (23) According to the commander's operation behavior on intelligence, the interest degree of intelligence is assigned and coded to form a matrix R i =[α], R i is the interest level of the intelligence in the i-th data label, where α = 0, 1, 2, 3, representing no operation, selection, labeling, and output, respectively; (24) Use matrix decomposition algorithm to remove the matrix (M i , N i , P i ) and the matrix (Q i , R i ) in the blank space; The improved linear discriminant analysis method is specifically: (221) Solve the intra-class dispersion matrix S of the linear discriminant analysis method w and the inter-class dispersion matrix S b : Where x m represents the mth class sample, N m represents the number of samples at m, N represents the total number of samples, x mn represents the nth sample in the mth class sample, c represents the number of sample categories, and T represents the transpose of the matrix; (222) The Fisher criterion is changed to the following formula: Where, is an n-dimensional column vector, is the minimum scatter matrix difference criterion function; (223) Solve the following characteristic equation to find the eigenvectors corresponding to the d smallest eigenvalues: (S w -S b )a k =λ k a k ,k=1,2,…,d Where λ k is the kth smallest eigenvalue, α k is the eigenvalue λ k The corresponding eigenvector; The step (24) specifically includes: (241) According to the certain correlation between the intelligence concerned by the same commander under the same action task, the co-occurrence matrix R of the action task text and the intelligence text is constructed. st }=(P s , Q t ), where r st represents the number of times the intelligence text t appears in the sliding window centered on the action task s; each row of the matrix R is truncated to retain only the K non-zero r with the largest value st , the rest r st Set to zero; (242) Note x s is the vector when the action task s is located at the center of the sliding window, y t The vector representing the intelligence text t when it is not in the center of the sliding window can be constructed using the following matrix decomposition model formula: Where c st =αlog(1+r st ), α and λ are hyper parameters, and the action task matrix P s =[x1, x2, ...x s ..., x M ], intelligence text matrix Q t =[y1,y2,...y t ..., y M ], solve the objective function above and obtain the similarity between the action task text and the intelligence text; The model optimization module specifically includes: (31) The model consists of three convolutional layers, three pooling layers, and one fully connected layer. The ReLU function is used as the activation function in the convolutional layer, and the last layer of the model is a fully connected layer, which divides the predicted intelligence results into four categories for output. (32) Command personnel and operational tasks are used as the input of the model, and intelligence is used as the output of the model; a multi-layer sparse edge denoising autoencoder network is formed by stacking, and each layer of the sparse edge denoising autoencoder is trained hierarchically, and then the parameters of the entire model are adjusted to achieve model optimization; (33) All historical task sample data are put into the model for training to obtain the optimal model; there is no initial error when the model is first trained, so an initial error is given. After one training, the error between the model output and the actual situation is calculated and used as the correction error for the next training. It is continuously iterated to converge within the specified accuracy, and the optimal parameters are obtained, that is, the optimal model is obtained; The error processing model optimization method is as follows: (321) The autoencoder is a three-layer neural network consisting of an encoder and a decoder, which are respectively an input layer h, a hidden layer w, and an output layer v. If the encoding function is f and the decoding function is g, the encoder maps the input layer h to the hidden layer w, and the feature representation is: w=f(h)=s f (Wh+b) Where S f is the encoder activation function, W is the encoder weight matrix, and b is the encoder bias vector; the decoder g maps the hidden layer w to the output layer v as follows: v=g(w)=s g (W′w+b′) Where s g is the decoder activation function, W′ is the decoder weight matrix, and b′ is the decoder bias vector; the reconstruction error L(h,v) measures the similarity between the input layer h and the output layer v: L(h,v)=||h-v|| 2 The overall loss function of the autoencoder is: Where s is the training sample set; (322) The loss function of the sparse autoencoder is: In the formula, β is the weight coefficient, ρ is the parameter, is the activation degree of the neuron; (323)The loss function of the edge denoising autoencoder is: Where, is the variance of the interference term of the dth dimension of the input data; (324) The loss function formula of the sparse edge denoising autoencoder is as follows: SmDAE(θ)=L(h,v)+J SAE +J mDAE +Y Where y i is the total number of real intelligence under the i-th type of action mission, is the total number of accurately predicted intelligence under the i-th action task, r is the feedback correction factor, and the loss function J of the sparse edge denoising autoencoder is optimized by training the network. SmDAE .

Images