A device and method for extracting deep confidence network intelligence in cognitive information confrontation

Abstract

The present invention provides a deep confidence network intelligence extraction device and method in cognitive information confrontation, which relates to the field of cognitive electronic warfare technology. The device of the present invention includes a DBN network deep learning platform, a front-end multi-source sensor, and a data storage device. The front-end multi-source sensor collects a large amount of target feature data in real time and inputs it into the DBN network input layer, and uploads the original information through the input layer; based on the DBN network deep learning architecture, the weight of the generation model is obtained by training through an unsupervised greedy layer-by-layer method, and the value is passed to the hidden layer through it. In turn, the formed weight will be randomly selected to reconstruct the original input data; through multiple iterations of the bidirectional data layer and the hidden layer, the feature true expression of the massive reconnaissance data is obtained; through the classification and fusion of the top-level data feature vector, high-precision analysis intelligence information of the massive reconnaissance data is formed; the high-dimensional features are projected into the low-dimensional space, and the similarity of the encoded training data is verified to express the true features of the massive reconnaissance data.

Description

Technical Field

[0001] The present invention relates to the field of cognitive electronic warfare technology, and in particular to a device and method for extracting deep confidence network intelligence in cognitive information confrontation. Background Art

[0002] In the war of air-ground integrated information confrontation, both sides attempt to seize the initiative on the battlefield by intercepting, interfering, deceiving, and controlling information. In many information confrontation scenarios, the reconnaissance party cannot obtain prior information on the reconnaissance data. At the same time, it is also necessary to identify multiple access methods, estimate signal parameters, identify modulation methods, capture frequency hopping patterns in communication reconnaissance, identify radar parameters, fingerprint features, working modes, etc. in radar reconnaissance, and even obtain intelligence information such as the target's military purpose, information deception, behavioral characteristics, cooperation methods, and tactical intentions. Therefore, the ability to accurately analyze, identify, and form accurate intelligence on massive reconnaissance data without prior knowledge is of great significance to modern air-ground integrated information confrontation and military confrontation, and is also the bottleneck and core of information intelligence acquisition, battlefield cognition, and battlefield command and control.

[0003] Traditional data analysis methods are no longer able to cope with the complex and diverse massive reconnaissance data. Therefore, there is an urgent need for corresponding efficient methods to analyze massive reconnaissance data, improve the analysis capability and accuracy of massive data, and ensure stable and reliable intelligence features analyzed under harsh channel conditions, while mining valuable intelligence information from massive and complex reconnaissance data.

[0004] For the analysis of massive reconnaissance data, the deep learning method first extracts the primary features of the data from the complex and massive reconnaissance data, and then goes to the intermediate area to perform target recognition, frequency hopping pattern analysis, parameter estimation, radar parameter extraction, and image feature extraction, and then goes to the advanced area to perform modulation mode recognition, parameter estimation adjustment, multiple access mode recognition, radar operating mode, and image feature recognition, and finally to the more advanced area of ​​military intelligence, target behavior, support characteristics, interference intentions, information deception, etc., to dig out deeper combat intelligence information.

[0005] At present, data analysis methods are divided into two categories: statistical pattern recognition methods based on feature extraction and maximum likelihood hypothesis testing methods based on decision theory. The former first extracts feature vectors from the data, and then uses the pattern recognition system to determine the identification classification; the latter is a multi-hypothesis likelihood problem, which is characterized by observing the form of the data to be identified, setting it as a candidate method, and then determining the specific method through similarity judgment.

[0006] The maximum likelihood hypothesis test method based on decision theory requires more prior knowledge, such as the form of the distribution function and the mean, variance, and signal-to-noise ratio parameters. For massive reconnaissance data, there is almost no prior knowledge available. Secondly, the existence of unknown parameters makes the statistical parameter expression of the likelihood ratio classification very complex, the amount of calculation is relatively large, and it is difficult to process in real time. If the likelihood ratio function is simplified, the classification information is lost and the classification performance is reduced. Therefore, there is a trade-off between the simplification method of the likelihood ratio function and the loss of classification performance. At the same time, the likelihood ratio method is sensitive to model mismatch and parameter deviation.

[0007] In terms of noise resistance, the statistical pattern recognition method based on feature extraction uses eigenvalues ​​with good noise resistance such as high-order cumulants, but the effect of eigenvalues ​​on noise suppression is still very limited. In the design of the recognizer, most algorithms use separate classifiers and separate recognizers, which cannot meet the recognition problem of large changes in data eigenvalues. The extracted data features are not fused and judged into corresponding intelligence information through recognition.

[0008] Current feature extraction, classification, regression and other methods can be used as shallow structure algorithms. Their limitations are that they have limited ability to represent complex functions under limited samples and computing units, and their generalization ability for complex classification problems is subject to certain constraints. Summary of the invention

[0009] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and to provide a deep belief network intelligence extraction device and method in cognitive information confrontation. Based on the deep learning architecture of DBN (Deep Belief Network), the device and method can obtain the true expression of the features of massive reconnaissance data and form high-precision and high-reliability fused intelligence information of massive reconnaissance data.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0011] On the one hand, the present invention provides a deep confidence network intelligence extraction device in cognitive information confrontation, including a DBN network deep learning platform, a front-end multi-source sensor, and a data storage device; the DBN network deep learning platform is implemented by a digital chip Xilinx FPGA V6 chip XC6VHX255T-2FFG1155C, which is used to complete high-speed multi-information processing and fusion; the front-end multi-source sensor is a 1×N sensor array, including a plurality of image sensors and a plurality of laser radar sensors, which are used to obtain target features; the data storage device adopts FLASH AT24C02, which is an externally expandable storage chip of the DBN network deep learning platform, and is used to store the information collected by the front-end multi-source sensor; the front-end multi-source sensor is connected through the electrical port and the 10 Gigabit network optical port of the Aurora expandable by the XC6VHX255T-2FFG1155C chip;

[0012] Through the ALU unit of the FPGA deep learning platform of the DBN network, a network module with characteristic weight function is formed, and through the communication transmission between modules, a general IP soft core of the deep learning network is formed; through the Aurora electrical port and 10G network optical port expandable by the XC6VHX255T-2FFG1155C chip, high-speed data stream data of the target characteristics is obtained to form a high-speed seamless data connection; through the external expandable FLASH storage chip of the DBN network deep learning platform, a knowledge base with characteristics is formed to form fast information knowledge recognition access.

[0013] Furthermore, the image sensor adopts Hikvision DS-2CE16C3T-IT3 model, and the lidar sensor adopts VelodyneHDL-64E model.

[0014] On the other hand, the present invention also provides a method for extracting deep confidence network intelligence in cognitive information confrontation, which is implemented by using the above-mentioned deep confidence network intelligence extraction device in cognitive information confrontation, and using the software program in the DBN network deep learning platform to realize deep confidence network intelligence extraction in cognitive information confrontation. The specific method is as follows:

[0015] Step 1: The front-end multi-source sensor collects massive target feature data in real time; the front-end multi-source sensor is connected to the DBN network input layer and uploads the original information through the input layer;

[0016] Step 2: For data flow processing, a Bayesian belief network (Deep Boltzmann Machine, DBM) is used near the bottom layer of massive data, i.e. a directed graph model with no connection between nodes in the layer, and a restricted Boltzmann machine (RBM) is used at the farthest part;

[0017] Step 3: By combining DBM and RBM, increasing the number of hidden layers, we get DBN. Multi-layer RBM and DBM form DBN. DBN is a probabilistic generative model. The generative model is to establish a joint distribution between massive data and features. The bottom layer of DBN is composed of multiple RBM layers. The DBN network is "restricted" to one data layer and one hidden layer. There are connections between layers, but there are no connections between units within a layer.

[0018] Step 4: Establish DBM network;

[0019] The hidden units are trained to capture the correlation of high-order data presented in the data layer. This part is composed of DBM, which bidirectionally characterizes data features through probability transfer. The top network constitutes data correlation and fusion performance. The connection of a DBM is guided by the top-down generation of weights.

[0020] Step 5: Train the weights of the probability generation model DBN;

[0021] The DBN deep learning network obtains the weights of the generated DBN model through unsupervised greedy layer-by-layer pre-training, and obtains the preliminary features of the massive reconnaissance data, including signal-to-noise ratio, channel model, modulation method, radar parameters, and visual parameters. In this training phase, a vector is generated in the massive data layer, through which the feature values ​​of the multi-source data are passed to the hidden layer; in turn, the weights formed will be randomly selected to reconstruct the original input data.

[0022] Step 6: Update and obtain weights. The expression feature activation unit of the new multi-source data will transfer the activation unit to the hidden layer network to obtain new weights.

[0023] Through deep weight training, during the training process, the initial feature vector value is first passed to the hidden layer unit; then the hidden layer unit reconstructs the feature vector; these reconstructed feature vectors are mapped to the hidden layer unit again to obtain new hidden layer units; by increasing the number of network layers, the probability of training data is changed, and through multiple iterations of the bidirectional data layer and the hidden layer, a true expression that is closer and closer to the characteristics of massive data is obtained, thereby obtaining high-level information features;

[0024] Step 7: Obtain data output decision;

[0025] In the top two layers of the data flow, the data feature vectors are classified and fused, and the adjacent feature information is fused through the fusion function of the top layer; the massive data learning feature label set is obtained through training data, and the massive data learning feature label set is attached to the top layer. At the same time, DBN uses the correlation relationship of the neighborhood to achieve the feature transformation invariance of the generated model, and it is easier to transform to high-dimensional features, forming high-precision analysis intelligence information of massive reconnaissance data.

[0026] The beneficial effects of adopting the above technical solution are: the invention provides a deep belief network intelligence extraction device and method in cognitive information confrontation, proposes a deep learning architecture based on DBN network, and obtains the weights of the generation model through unsupervised greedy layer-by-layer method training, through which the value is passed to the hidden layer, and in turn, the formed weights will be randomly selected to reconstruct the original input data; through multiple iterations of the bidirectional data layer and the hidden layer, the true expression of the features of the massive reconnaissance data is obtained. And through the classification and fusion of the top-level data feature vectors, high-precision analysis intelligence information of the massive reconnaissance data is formed; the high-dimensional features are projected into the low-dimensional space, and the similarity of the encoded training data is verified to express the true characteristics of the massive reconnaissance data; the invention can improve the reliability of obtaining the intelligence information of the massive reconnaissance data, so as to obtain the optimal fusion intelligence information of the massive reconnaissance data. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic diagram of a device for extracting deep confidence network intelligence in cognitive information confrontation provided by an embodiment of the present invention;

[0028] Figure 2 A schematic diagram of a deep learning architecture based on a DBN network provided in an embodiment of the present invention;

[0029] Figure 3 A schematic diagram of deep learning decision-making based on fuzzy functions provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0031] A deep confidence network intelligence extraction device in cognitive information confrontation, such as Figure 1 As shown, it includes a DBN network deep learning platform, a front-end multi-source sensor, and a data storage device. The DBN network deep learning platform is implemented using the digital chip XilinxFPGAV6 chip XC6VHX255T-2FFG1155C, which is used to complete high-speed multi-information processing and fusion. The front-end multi-source sensor is a 1×N sensor array, including several image sensors and several laser radar sensors, which are used to obtain target radiation and receive high-frequency electromagnetic waves. The image sensor uses the Hikvision DS-2CE16C3T-IT3 model, and the laser radar sensor uses the VelodyneHDL-64E model. The data storage device is an externally expandable storage chip FLASHAT24C02 of the DBN network deep learning platform, which is used to store the information collected by the front-end multi-source sensor. The front-end multi-source sensor is connected through the electrical port and 10G network optical port of the Aurora that can be expanded by the XC6VHX255T-2FFG1155C chip.

[0032] Through the ALU unit of the FPGA deep learning platform of the DBN network, a network module with characteristic weight function is formed, and through the communication transmission between modules, a general IP soft core of the deep learning network is formed, forming a portable feature; through the Aurora 10G network optical port that can be expanded by the XC6VHX255T-2FFG1155C chip, high-speed data stream data from front-end multi-source sensors is obtained to form a high-speed seamless data connection; through the external expandable FLASH storage chip of the DBN network deep learning platform, a knowledge base with characteristics is formed to form fast information knowledge recognition access.

[0033] A method for extracting deep confidence network intelligence in cognitive information confrontation is implemented by using the above-mentioned deep confidence network intelligence extraction device in cognitive information confrontation, and a software program stored in the DBN network deep learning platform is used to realize deep confidence network intelligence extraction in cognitive information confrontation. The specific method is as follows:

[0034] The front-end multi-source sensors collect massive target feature data in real time; the front-end multi-source sensors are connected to the DBN network input layer, and the original information is uploaded through the input layer; first, for the processing of data streams, the Bayesian belief network (Deep Boltzmann Machine, DBM), that is, a directed graph model, is used near the bottom layer of massive data. There is no link between the nodes in the layer, and the restricted Boltzmann machine (Restricted Boltzmann Machine, RBM) is used in the farthest part. By combining DBM and RBM, the number of hidden layers is increased to obtain DBN. DBN is a probabilistic generative model. The generative model is to establish a joint distribution between massive data and features. The bottom layer of DBN consists of multiple RBM layers. The DBN network is "restricted" to one data layer and one hidden layer. There are connections between layers, but there are no connections between units within the layer.

[0035] Then, through the top two layers of the data flow, through data feature fusion, the higher-level output will provide an ideal reference, so that the top layer will fuse the adjacent feature information.

[0036] Finally, the feature label set learned from massive data will be attached to the top layer. At the same time, DBN uses the correlation relationship of the neighborhood to achieve the feature transformation invariance of the generated model, and it is easier to transform to high-dimensional features.

[0037] The deep learning architecture based on DBN network used in this embodiment is as follows Figure 2 shown.

[0038] Deep learning decision-making based on fuzzy functions solves the problem that a simple classifier cannot adapt to large-scale data feature jumps. This method adopts a new fuzzy density concept that combines accuracy and uncertainty to improve the reliability of recognition. Fuzzy integral fuses multiple classifiers to obtain a near-optimal fusion result, which combines the accuracy and uncertainty of the classifier to improve the reliability of the overall classification. Figure 3 As shown in the figure, this method uses the feature classification results of the classifier to train and select effective fuzzy density and determine the appropriate fuzzy function. Among them, f(x) is the different data features after classification, and x is the super-complete basis. According to the selected fuzzy function, multiple classifiers are reasonably selected to maximize the unity, which reduces the unevenness of the adaptive classification algorithm caused by the combined classifier. Feature output and training to select effective fuzzy density can be processed in parallel, and according to the adjustment of fuzzy function parameters, data features are merged to output fusion intelligence information.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for extracting intelligence from deep belief networks in cognitive information confrontation, Features: The method is implemented by a deep confidence network intelligence extraction device in cognitive information confrontation, which includes a DBN network deep learning platform, a front-end multi-source sensor, and a data storage device; the DBN network deep learning platform is implemented by a digital chip Xilinx FPGAV6 chip XC6VHX255T-2FFG1155C, which is used to complete high-speed multi-information processing and fusion; the front-end multi-source sensor is a 1×N sensor array, including a plurality of image sensors and a plurality of laser radar sensors, which are used to obtain target features; the data storage device FLASH AT24C02 is an externally expandable storage chip of the DBN network deep learning platform, which is used to store the information collected by the front-end multi-source sensor; the front-end multi-source sensor is connected through the electrical port and the 10G network optical port of the Aurora expandable by the XC6VHX255T-2FFG1155C chip; the image sensor adopts the Hikvision DS-2CE16C3T-IT3 model, and the laser radar sensor adopts the VelodyneHDL-64E model; Through the ALU unit of the FPGA deep learning platform of the DBN network, a network module with characteristic weight function is formed, and through the communication transmission between modules, a general IP soft core of the deep learning network is formed; through the Aurora electrical port and 10G network optical port that can be expanded by the XC6VHX255T-2FFG1155C chip, high-speed data stream data of target characteristics is obtained to form a high-speed seamless data connection; through the external scalable FLASH storage chip of the DBN network deep learning platform, a knowledge base with characteristics is formed to form fast information knowledge recognition access; The software program in the DBN network deep learning platform is used to realize the deep belief network intelligence extraction in cognitive information confrontation. The specific method is as follows: Step 1: The front-end multi-source sensor collects massive target feature data in real time; the front-end multi-source sensor is connected to the DBN network input layer and uploads the original information through the input layer; Step 2: For data flow processing, a Bayesian belief network (Deep Boltzmann Machine, DBM) is used near the bottom layer of massive data, i.e. a directed graph model with no connection between nodes in the layer, and a restricted Boltzmann machine (RBM) is used at the farthest part; Step 3: By combining DBM and RBM, the number of hidden layers is increased to obtain DBN. Multi-layer RBM and DBM form DBN. The DBN network is "restricted" to one data layer and one hidden layer. There are connections between layers, but there are no connections between units within a layer. Step 4: Establish DBM network; The hidden units are trained to capture the correlation of high-order data presented in the data layer. This part is composed of DBM, which bidirectionally characterizes data features through probability transfer. The top network constitutes data correlation and fusion performance. The connection of a DBM is guided by the top-down generation of weights. Step 5: Train the weights of the probability generation model DBN; The DBN deep learning network obtains the weights of the generated DBN model through unsupervised greedy layer-by-layer pre-training, and obtains the preliminary features of the massive reconnaissance data, including signal-to-noise ratio, frequency hopping characteristics, multiple access mode, modulation mode, radar 5 parameters, radar fingerprint, working mode, and visual parameters; in this training stage, a vector is generated in the massive data layer, through which the feature values ​​of the multi-source data are transferred to the hidden layer; in turn, the weights formed will be randomly selected to reconstruct the original input data; Step 6: Update and obtain weights. The expression feature activation unit of the new multi-source data will transfer the activation unit to the hidden layer network to obtain new weights. Through deep weight training, during the training process, the initial feature vector value is first passed to the hidden layer unit; then the hidden layer unit reconstructs the feature vector; these reconstructed feature vectors are mapped to the hidden layer unit again to obtain new hidden layer units; by increasing the number of network layers, the probability of training data is changed, and through multiple iterations of the bidirectional data layer and the hidden layer, a true expression that is closer and closer to the characteristics of massive data is obtained, thereby obtaining high-level information features; Step 7: Obtain data output decision; In the top two layers of the data flow, the data feature vectors are classified and fused, and the adjacent feature information is fused through the fusion function of the top layer; the massive data learning feature label set is obtained through training data, and the massive data learning feature label set is attached to the top layer. At the same time, DBN uses the correlation relationship of the neighborhood to achieve the feature transformation invariance of the generated model, and it is easier to transform to high-dimensional features, forming high-precision analysis intelligence information of massive reconnaissance data.

Images