Cross-domain information reasoning method based on data restriction

Abstract

The application discloses a cross-domain information reasoning method based on data limitation, comprising the following steps: acquiring a missing data set of multi-domain information; S2, using the acquired missing data set and adopting maximum information entropy to construct an underwater cross-domain information reasoning network model to obtain a target equation of the whole underwater cross-domain information reasoning network model, and training the target equation to obtain an optimal underwater cross-domain information reasoning network model; S3, using the obtained optimal underwater cross-domain information reasoning network model to perform missing data reasoning on underwater environment data detected by a detection device, and outputting missing data of the underwater environment detection data. The application takes the missing data as an optimization target, adopts the maximum information entropy to construct a cross-domain information reasoning network model and a target equation of the model, trains the target equation to obtain an optimal model, and can realize high-quality prediction of missing information.

Description

Technical Field

[0001] This invention relates to the field of information reasoning technology, and in particular to a cross-domain information reasoning method based on data constraints. Background Technology

[0002] Due to the complex and ever-changing marine environment, it is difficult to simultaneously collect relatively complete information about the same target in multiple domains, which poses a significant challenge to exploring the consistency and complementarity between different domains. Therefore, supplementing missing information based on the correlation and consistency between data from different domains in the missing data is of great significance for improving the representational ability of multi-domain data and increasing the information content of the data. The problem of missing or corrupted data in some domains in the underwater environment has brought severe challenges to the environmental perception of underwater robots. For example, sonar can usually measure targets at greater distances, but the hydrostatic pressure effect of ocean currents, interference from ship movement, and uneven pressure distribution can all cause problems such as ghosting and low resolution in the data, resulting in missing data acquired by current sensors. Under the same conditions, data acquired from different domains of the same environment will have varying degrees of data loss. How to combine these incomplete multi-domain data to improve the perception ability of underwater robots under heavy loads is a core problem that urgently needs to be solved.

[0003] Currently, common reasoning methods include: transfer learning and generative adversarial learning;

[0004] 1. Transfer learning: Applying knowledge learned from one domain (source domain) to another related but different domain (target domain) to solve cross-domain information reasoning problems. This method is particularly suitable for situations where the target domain data is limited. However, its limitations are also obvious. The effectiveness of transfer learning is highly dependent on the correlation between the source domain and the target domain. If the two domains are too different, the effect of direct transfer is not good.

[0005] 2. Generative Adversarial Learning: Generative Adversarial Networks (CAN) can generate synthetic data that closely resembles the distribution of real data. In cross-domain information reasoning, it can be used to generate data in missing domains to assist in model learning. However, its limitations mainly lie in the instability of the training process of Generative Adversarial Networks and the quality problems of the generated data, especially in the case of highly complex data distributions. Summary of the Invention

[0006] The technical problem to be solved by this invention is: in order to solve the problem that existing cross-domain information reasoning methods have large limitations and poor reasoning results, this invention provides a cross-domain information reasoning method based on data constraints. By improving the cross-domain information reasoning method, a cross-domain information reasoning network model is constructed by maximizing information entropy, so as to achieve high-quality prediction of missing information.

[0007] The technical solution adopted by this invention to solve its technical problem is: a cross-domain information reasoning method based on data constraints, comprising the following steps:

[0008] S1. Obtain the missing dataset with multi-domain information;

[0009] S2. Using the missing dataset obtained in step S1, and by maximizing information entropy, construct an underwater cross-domain information reasoning network model to obtain the overall objective equation of the underwater cross-domain information reasoning network model, and train the objective equation to obtain the optimal underwater cross-domain information reasoning network model.

[0010] S3. Using the optimal underwater cross-domain information reasoning network model obtained in step S2, perform missing data reasoning on the underwater environmental data detected by the detection equipment, and output the missing data of the underwater environmental detection data. Thus, the missing data is used as the optimization objective, and a cross-domain information reasoning network model and its objective equation are constructed by maximizing information entropy. The objective equation is trained to obtain the optimal model, which can achieve high-quality prediction of missing information, thereby improving the environmental perception capability of the detection equipment in complex environments.

[0011] Furthermore, in step S1, if the probability of event x is p(x), then the formula for calculating the information content I(x) of the underwater environment is:

[0012] I(x) = -log p(x);

[0013] If the task performed by the detection equipment in the ocean is U, then the formula for calculating the probability I(x, U) of the information I(x) provided by the environment in which event x occurs is:

[0014] I(x, U) = -log p(x| U );

[0015] The total amount of information F collected by the detection device from all events in the i-th domain i ∈i s×t The calculation formula is:

[0016] F i =I(X) i ,U);

[0017] In the event space X∈i m×s×t In this scenario, the formula for calculating the amount of information θ obtained by the detection device is:

[0018]

[0019] Where: m represents the number of data domains collected by the detection device, A represents the attention matrix, and X... iLet x represent the event in the i-th domain, s represent spatial information, t represent temporal information, ⊙ represent the dot product of matrices, and D represent the amount of data in the current underwater environment space X. Therefore, the amount of information from the detection device in the current scene is modeled, and an information representation model is constructed.

[0020] Further, step S2 includes the following steps:

[0021] S2-1, Original low-dimensional representation based on graph structure;

[0022] S2-2, Cross-domain information correlation learning;

[0023] S2-3, Learning to maintain inter-domain consistency;

[0024] S2-4, Inference of Missing Domain Samples. This enables high-quality prediction of missing data.

[0025] Further, in step S2-1, the original optical image G 1 Represented as:

[0026] G 1 ={A 1 ,X};

[0027] K-nearest neighbor algorithm constructs feature map G 2 Represented as:

[0028] G 2 ={A 2 ,X};

[0029] Node x i and x j Similarity S between ij The calculation formula is:

[0030]

[0031] The output H of layer l+1 (l+1) The calculation formula is:

[0032] H (l+1) =ReLU(D -1 / 2 AD -1 / 2 H (l) W (l) );

[0033] Among them: A 1 ∈i N×N Let be the adjacency matrix of the data, containing N nodes, ReLU be the activation function, D be the degree matrix of the feature map, and W be the number of nodes. (l) For the training weights of the network, map H to two domains. 1 and H 2To through the mapping function g(H) 1 ;δ)→Z 1 and g(H) 2 ;δ)→Z 2 Represented as Z in low-dimensional space 1 and Z 2 δ is the encoding parameter. Therefore, the original structural information of the data can be preserved, effectively maintaining the graph structure of the data.

[0034] Furthermore, in step S2-2, the cross-domain information loss function L is obtained through the mutual information between the two domains. co The maximum correlation is represented as:

[0035]

[0036] Where α is the regularization parameter, I is the mutual information between the two domains, and H is the information entropy. Therefore, by maximizing the mutual information of different data in the two domains, we can uncover the correlation and complementarity between the domains.

[0037] Furthermore, in steps S2-3, the underwater cross-domain information reasoning network model maximizes the mutual information I(Z) between the original graph and other domain low-dimensional representations. 1 A 2 The consistency between domains is represented as follows:

[0038]

[0039] Based on the relationship between entropy and mutual information, the above equation can be decomposed as follows:

[0040] I(Z 1 A 2 )=H(A 2 )-H(A 2 |Z 1 )S;

[0041] Where: H(A) 2 |Z 1 H(A) is the conditional entropy. 2 ) is A 2 The entropy. Therefore, we construct entropy and mutual information representations to explore consistent representations of different domains in a low-dimensional representation space.

[0042] Furthermore, in step S2-3, the objective equation is solved, and I(Z) is used to solve the objective equation. 1 A 2 The expression is transformed into:

[0043]

[0044] The loss function that maximizes the low-dimensional embedding of the spectral image and the original visible light image is expressed as:

[0045]

[0046] The consistency preservation loss function between different underwater domains is expressed as:

[0047]

[0048] Therefore, by designing different loss functions, we can obtain consistent representation information between the two domains to minimize the differences between the domains.

[0049] Furthermore, in steps S2-4, the samples conform to a Gaussian distribution, and the objective equation is constructed as follows:

[0050]

[0051] The overall objective equation of the underwater cross-domain information reasoning network model is expressed as:

[0052] L = L co +L uni +L fo .

[0053] Therefore, by combining the loss functions of the above three parts, consistency constraints are constructed for spectral image domain and visible light domain data to achieve accurate inference of missing data.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] By taking missing data as the optimization target and constructing a cross-domain information reasoning network model and its objective equation using the method of maximizing information entropy, the objective equation is trained to obtain the optimal model, which can achieve high-quality prediction of missing information, thereby improving the environmental perception capability of detection equipment in complex environments. Attached Figure Description

[0056] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0057] Figure 1 The flowchart shows the cross-domain information reasoning method based on data constraints of the present invention.

[0058] Figure 2 This is a flowchart of step S2 of the present invention;

[0059] Figure 3 This is a diagram of the cross-domain information reasoning model of the present invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, features defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0062] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0063] like Figures 1 to 3 The diagram shows the preferred embodiment of the present invention. This embodiment's cross-domain information reasoning method based on data constraints includes the following steps:

[0064] S1. Obtain the missing dataset with multi-domain information;

[0065] S2. Using the missing dataset obtained in step S1, and by maximizing information entropy, construct an underwater cross-domain information reasoning network model to obtain the overall objective equation of the underwater cross-domain information reasoning network model, and train the objective equation to obtain the optimal underwater cross-domain information reasoning network model.

[0066] S3. Using the optimal underwater cross-domain information reasoning network model obtained in step S2, perform missing data reasoning on the underwater environmental data detected by the detection equipment, and output the missing data of the underwater environmental detection data. Thus, the missing data is used as the optimization objective, and a cross-domain information reasoning network model and its objective equation are constructed by maximizing information entropy. The objective equation is trained to obtain the optimal model, which can achieve high-quality prediction of missing information, thereby improving the environmental perception capability of the detection equipment in complex environments.

[0067] For example, the testing equipment uses an underwater robot.

[0068] In this embodiment, in step S1, the probability of event x is p(x), and the formula for calculating the information content I(x) of the underwater environment is:

[0069] I(x) = -log p(x);

[0070] If the task performed by the detection equipment in the ocean is U, then the formula for calculating the probability I(x, U) of the information I(x) provided by the environment in which event x occurs is:

[0071] I(x, U) = -log p(x| U );

[0072] The total amount of information F collected by the detection device from all events in the i-th domain i ∈i s×t The calculation formula is:

[0073] F i =I(X) i ,U);

[0074] In the event space X∈i m×s×t In this scenario, the formula for calculating the amount of information θ obtained by the detection device is:

[0075]

[0076] Where: m represents the number of data domains collected by the detection device, A represents the attention matrix, and X... i Let x represent the event in the i-th domain, s represent spatial information, t represent temporal information, ⊙ represent the dot product of matrices, and D represent the amount of data in the current underwater environment space X. Therefore, the amount of information from the detection device in the current scene is modeled, and an information representation model is constructed.

[0077] In this embodiment, step S2 includes the following steps:

[0078] S2-1, Original low-dimensional representation based on graph structure;

[0079] S2-2, Cross-domain information correlation learning;

[0080] S2-3, Learning to maintain inter-domain consistency;

[0081] S2-4, Inference of Missing Domain Samples. This enables high-quality prediction of missing data.

[0082] In this embodiment, in step S2-1, the original optical image G 1 Represented as:

[0083] G 1 ={A1 ,X};

[0084] K-nearest neighbor algorithm constructs feature map G 2 Represented as:

[0085] G 2 ={A 2 ,X};

[0086] Node x i and x j Similarity S between ij The calculation formula is:

[0087]

[0088] The output H of layer l+1 (l+1) The calculation formula is:

[0089] H (l+1) =ReLU(D -1 / 2 AD -1 / 2 H (l) W (l) );

[0090] Among them: A 1 ∈i N×N Let be the adjacency matrix of the data, containing N nodes, ReLU be the activation function, D be the degree matrix of the feature map, and W be the number of nodes. (l) For the training weights of the network, map H to two domains. 1 and H 2 To through the mapping function g(H) 1 ;δ)→Z 1 and g(H) 2 ;δ)→Z 2 Represented as Z in low-dimensional space 1 and Z 2 δ is the encoding parameter. Therefore, the original structural information of the data can be preserved, effectively maintaining the graph structure of the data.

[0091] In this embodiment, in step S2-2, the cross-domain information loss function L is obtained through the mutual information between the two domains. co The maximum correlation is represented as:

[0092]

[0093] Where α is the regularization parameter, I is the mutual information between the two domains, and H is the information entropy. Therefore, by maximizing the mutual information of different data in the two domains, we can uncover the correlation and complementarity between the domains.

[0094] In this embodiment, in steps S2-3, the underwater cross-domain information reasoning network model maximizes the mutual information I(Z) between the original graph and the low-dimensional representations of other domains. 1 A 2 The consistency between domains is represented as follows:

[0095]

[0096] Based on the relationship between entropy and mutual information, the above equation can be decomposed as follows:

[0097] I(Z 1 A 2 )=H(A 2 )-H(A 2 |Z 1 )S;

[0098] Introducing the solution of the objective equation, I(Z) 1 A 2 The expression is transformed into:

[0099]

[0100] The loss function that maximizes the low-dimensional embedding of the spectral image and the original visible light image is expressed as:

[0101]

[0102] The consistency preservation loss function between different underwater domains is expressed as:

[0103]

[0104] Where: H(A) 2 |Z 1 H(A) is the conditional entropy. 2 ) is A 2 The entropy is then calculated. Therefore, entropy and mutual information representations are constructed to explore the consistent representations of different domains in a low-dimensional representation space. Different loss functions are designed to obtain consistent representation information between two domains, minimizing the differences between domains.

[0105] In this embodiment, in steps S2-4, the samples conform to a Gaussian distribution, and the objective equation is constructed as follows:

[0106]

[0107] The overall objective equation of the underwater cross-domain information reasoning network model is expressed as:

[0108] L = L co +L uni +L fo .

[0109] Therefore, by combining the loss functions of the above three parts, consistency constraints are constructed for spectral image domain and visible light domain data to achieve accurate inference of missing data.

[0110] In summary, this invention takes missing data as the optimization target and constructs a cross-domain information reasoning network model and its objective equation by maximizing information entropy. The objective equation is trained to obtain the optimal model, which can achieve high-quality prediction of missing information, thereby improving the environmental perception capability of detection equipment in complex environments.

[0111] The above description is based on the preferred embodiments of the present invention. Through the above description, those skilled in the art can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the contents of the specification, but must be determined by the scope of the claims.

Claims

1. A method for underwater cross-domain information reasoning based on data-constrained conditions, characterized in that, Includes the following steps: S1. Obtain the missing dataset with multi-domain information; S2. Using the missing dataset obtained in step S1, and by maximizing information entropy, construct an underwater cross-domain information reasoning network model to obtain the overall objective equation of the underwater cross-domain information reasoning network model, and train the objective equation to obtain the optimal underwater cross-domain information reasoning network model. S3. Using the optimal underwater cross-domain information reasoning network model obtained in step S2, perform missing data reasoning on the underwater environmental data detected by the detection equipment, and output the missing data of the underwater environmental detection data. Step S2 specifically includes the following steps: S2-1. The original low-dimensional representation based on graph structure includes: Original optical image G 1 Represented as: in, For event space; K-nearest neighbor algorithm constructs feature map G 2 Represented as: ； Node x i and x j Similarity S between ij The calculation formula is: ； The output H of layer l+1 （l+1） The calculation formula is: ； in: Let be the adjacency matrix of the data, containing N nodes, ReLU be the activation function, D be the degree matrix of the feature map, and W be the number of nodes. （l） For the training weights of the network, map H to two domains. 1 and H 2 to through mapping function and Represented as Z in low-dimensional space 1 and Z 2 δ is the encoding parameter; S2-2, Cross-domain information correlation learning, obtaining the cross-domain information loss function L through mutual information between two domains. co The maximum correlation is represented as: ； Where: α is the regularization parameter, I is the mutual information between the two domains, and H is the information entropy; S2-3, Learning to maintain inter-domain consistency; In steps S2-3, the underwater cross-domain information reasoning network model maximizes the mutual information I(Z) between the original graph and the low-dimensional representations of other domains. 1 A 2 To represent the consistency between domains as follows: ； Based on the relationship between entropy and mutual information, the above equation can be decomposed as follows: ； Where: H(A) 2 |Z 1 H(A) is the conditional entropy. 2 ) is A 2 Entropy; Introducing the solution of the objective equation, The expression is transformed into: ； The loss function that maximizes the low-dimensional embedding of the spectral image and the original visible light image is expressed as: ； The consistency preservation loss function between different underwater domains is expressed as: ； S2-4, Missing Domain Sample Inference: In step S2-4, the samples conform to a Gaussian distribution, and the objective equation is constructed as follows: ； The overall objective equation of the underwater cross-domain information reasoning network model is expressed as: 。 2. The underwater cross-domain information reasoning method based on data-constrained conditions according to claim 1, characterized in that, In step S1, if the probability of event x is p(x), then the formula for calculating the information content I(x) of the underwater environment is: ； If the task performed by the detection equipment in the ocean is U, then the probability I(x, U) of the information I(x) provided by the environment in which event x occurs is calculated as follows: ； The total amount of information collected by the detection device for all events in the i-th domain. The calculation formula is: ; In the event space In this scenario, the formula for calculating the amount of information θ obtained by the detection device is: ； Where: m represents the number of data domains collected by the detection device, A represents the attention matrix, and X... i Let x represent the event in the i-th domain, s represent spatial information, and t represent temporal information. This represents the dot product operation of matrices, and D represents the amount of data in the current underwater environment space X.

Images