Team knowledge multi-modal perception and gap traceability prediction method based on reinforcement learning

Abstract

The application discloses a team knowledge multi-modal perception and gap traceability prediction method based on reinforcement learning, and comprises the following steps: S1, multi-modal knowledge data of each cooperation stage in a team history cooperation process is collected, which relates to the technical field of artificial intelligence and knowledge management, the team knowledge manifold including knowledge state points and neighborhood similarity correlation is constructed by fusing and reducing the dimension of the multi-modal knowledge data of the team in advance, then the reinforcement learning strategy network is trained in combination with historical cooperation strategy actions and cooperation benefit values, then the team knowledge manifold is reduced to a two-dimensional grid and a manifold strategy consistency atlas is generated through Shannon entropy mapping, and finally the design of segmenting the atlas to identify a topological defect area is realized, so that effective detection of topological defects in the team knowledge manifold is realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge management technology, and in particular to a method for team knowledge multimodal perception and gap cause prediction based on reinforcement learning. Background Technology

[0002] In complex team collaboration environments, knowledge is often modeled in non-Euclidean spaces, such as high-dimensional spaces embedded through manifold learning techniques. Reinforcement learning is used to optimize policies in this space to achieve multimodal knowledge perception and gap-based abductive prediction.

[0003] Existing technologies largely focus on optimizing manifold learning or reinforcement learning algorithms themselves. However, when a team introduces new members with extremely heterogeneous cognitive patterns, such as AI agents with counterintuitive cognitive patterns, the existing knowledge manifold develops topological defects, such as black hole-like singularities. This causes a break in the transmission of reinforcement learning policies on the manifold, resulting in a manifold tearing problem driven by topological defects: geometrically similar points in the state space may produce drastically different policy actions. The boundaries of these topological defects are difficult to detect effectively. This directly disrupts the continuity and stability of team knowledge collaboration. Summary of the Invention

[0004] To address the technical problems existing in the background art, this invention proposes a team knowledge multimodal perception and gap cause prediction method based on reinforcement learning.

[0005] The proposed method for team knowledge multimodal perception and gap cause prediction based on reinforcement learning includes the following steps: S1. Collect multimodal knowledge data, historical collaboration strategies and actions, and historical collaboration results data for each collaboration stage in the team's historical collaboration process; fuse the multimodal knowledge data of each collaboration stage into Euclidean space feature vectors; obtain the Euclidean space feature vectors of all collaboration stages to form a set of Euclidean space feature vectors; use a manifold learning algorithm to reduce the dimensionality of each Euclidean space feature vector in the set of Euclidean space feature vectors to obtain multiple low-dimensional manifold vectors; use the low-dimensional manifold vectors as low-dimensional manifold features; and construct the team knowledge manifold based on the multiple low-dimensional manifold features. S2. Based on the team's knowledge manifold, generate a reinforcement learning policy network. The reinforcement learning policy network takes knowledge state points as input and outputs the probability distribution of the collaborative policy action corresponding to the knowledge state point. S3. Generate manifold feature images based on the team knowledge manifold and reinforcement learning policy network; S4. Perform image segmentation processing on the manifold strategy consistency map, and identify regions with gray intensity values ​​below a preset threshold as topological defect regions.

[0006] Preferably, in S1, the multimodal knowledge data includes text-based knowledge data, audio-based knowledge data, behavioral knowledge data, and image-based knowledge data; Historical collaboration results data include task completion time, task output quality indicators, and collaboration cost indicators.

[0007] Preferably, in S1, the multimodal knowledge data of each collaborative stage is fused into a Euclidean space feature vector, as follows: The TF-IDF algorithm is used to extract features from text-based knowledge data to generate TF-IDF feature vectors. Principal component analysis is then used to reduce the dimensionality of the TF-IDF feature vectors to 256 dimensions, resulting in 256-dimensional semantic feature vectors. The Mel-Cepstral Coefficient algorithm is used to extract time-frequency domain features from audio knowledge data; then, a one-dimensional convolutional neural network is used to encode the time-frequency domain features into a 128-dimensional acoustic feature vector. The behavioral knowledge data is preprocessed using min-max normalization, and then the features are mapped and encoded using a fully connected neural network to obtain a 64-dimensional behavioral feature vector. A convolutional neural network is used to encode image-type knowledge data into a 256-dimensional visual feature vector. A multilayer perceptron is used to map and fuse 256-dimensional semantic feature vectors, 128-dimensional acoustic feature vectors, 64-dimensional behavioral feature vectors, and 256-dimensional visual feature vectors into a 128-dimensional Euclidean space feature vector.

[0008] Preferably, in S1, a manifold learning algorithm is used to reduce the dimensionality of each Euclidean space feature vector in the set of Euclidean space feature vectors, resulting in multiple low-dimensional manifold vectors, as follows: The manifold learning algorithm is used to reduce the dimensionality of each Euclidean space feature vector in the set of Euclidean space feature vectors to 8, 12 or 16 dimensions, resulting in multiple low-dimensional feature vectors.

[0009] Preferably, in S1, a team knowledge manifold is constructed based on multiple low-dimensional manifold features, as follows: The geodesic distance between each low-dimensional manifold feature is obtained through a manifold learning algorithm. The reciprocal of the geodesic distance is used as the similarity quantification value. Two low-dimensional manifold features with similarity quantification values ​​greater than a preset similarity threshold are regarded as neighborhood feature pairs. All neighborhood feature pairs and their corresponding similarity quantification values ​​are integrated into a structured association relationship to generate neighborhood similarity associations between features. All low-dimensional manifold features and their neighborhood similarity associations are integrated into a holistic topological structure, which is then considered the team's knowledge manifold. Each low-dimensional manifold feature corresponds to a point on the team's knowledge manifold, serving as a knowledge state point. Each knowledge state point uniquely corresponds to the team's knowledge state in a single collaborative phase.

[0010] Preferably, in S2, a reinforcement learning policy network is generated based on the team knowledge manifold, as follows: Obtain the historical collaboration strategy actions corresponding to each knowledge state point in the team's knowledge manifold; The task completion time, task output quality index, and collaboration cost index in the historical collaboration result data are normalized. Based on the preset weights of task completion time, task output quality data, and collaboration cost data, a linear weighted scoring algorithm is used to obtain the collaboration benefit value of the normalized task completion time, task output quality index, and collaboration cost index, which is used as the collaboration benefit value corresponding to each knowledge state point. As an explanation, the preset values ​​for the weights of task completion time, task output quality data, and collaboration cost data can be set based on the subjective evaluation of the importance of task completion time, task output quality, and collaboration cost by team managers or business experts, and the weights of each importance can be normalized. The preset values ​​for the weights of task completion time, task output quality data, and collaboration cost data can also be set by the system administrator through the system configuration interface according to different business scenarios. The training sample set is obtained by using each knowledge state point on the team knowledge manifold, the historical collaboration strategy action corresponding to each knowledge state point, and the collaboration benefit value corresponding to each knowledge state point as training samples. Based on the training sample set, a neural network algorithm is used to generate a neural policy network. The neural policy network takes knowledge state points as input and outputs the probability distribution of the cooperative policy action corresponding to the knowledge state point. Based on the training sample set and the preset team collaboration goal, with the objective of maximizing the cumulative collaboration benefit value under the preset team collaboration goal, the network parameters of the neural policy network are iteratively updated using a reinforcement learning algorithm to obtain the reinforcement learning policy network. The reinforcement learning policy network takes knowledge state points as input and outputs the probability distribution of the collaboration policy action corresponding to the knowledge state point.

[0011] Preferably, in S3, a manifold feature image is generated based on the team knowledge manifold and the reinforcement learning policy network, as follows: S3.1 Project the low-dimensional manifold features in the team knowledge manifold to a two-dimensional Euclidean space using dimensionality reduction techniques to obtain the corresponding two-dimensional point cloud; divide the region where the two-dimensional point cloud is located in the two-dimensional Euclidean space into a regular grid of M×N, with each grid serving as a pixel grid; M and N are both positive integers greater than 1, and M and N may be equal or unequal. S3.2 For each pixel grid, count all the knowledge state points corresponding to the two-dimensional point cloud that fall within it, and use them as pixel knowledge state points. The probability distribution of the cooperative policy action corresponding to the pixel knowledge state point is obtained by using a reinforcement learning policy network, which is then used as the probability distribution of the pixel cooperative policy action. Obtain the Shannon entropy of the probability distribution of pixel cooperative strategy actions, and use it as the Shannon entropy corresponding to the pixel knowledge state point. Aggregate the Shannon entropy corresponding to all prime knowledge state points within the pixel grid to obtain the Shannon entropy value corresponding to the pixel grid. S3.3 Obtain the negative value of Shannon entropy based on the Shannon entropy value corresponding to the pixel grid; The negative values ​​of the Shannon entropy corresponding to the pixel grid are normalized and mapped to the gray value range of 0-255. The normalized gray value is used as the gray intensity value of the corresponding pixel grid. After traversing all pixel grids, an M×N grayscale image is obtained, which is used as the manifold policy consistency map.

[0012] As an illustration, the Shannon entropy corresponding to all elemental knowledge state points within the pixel grid is aggregated, for example, by taking the average value.

[0013] Preferably, in S3.1, the dimensionality reduction technique includes an autoencoder; the autoencoder is used to map the low-dimensional manifold features in the team knowledge manifold to a two-dimensional Euclidean space.

[0014] Preferably, in step S3.3, the normalization process uses the minimum-maximum normalization method to linearly map the negative value of Shannon entropy to the gray value range of 0-255.

[0015] Preferably, in S4, the manifold strategy consistency map is processed by image segmentation, and regions with grayscale intensity values ​​below a preset threshold are identified as topological defect regions, as follows: Based on a preset threshold, pixel grids with grayscale intensity values ​​lower than the preset threshold are marked as abnormal pixel grids; By forming a connected region from adjacent abnormal pixel grids, one or more connected regions can be obtained; Connected regions are treated as topological defect regions.

[0016] The proposed reinforcement learning-based multimodal perception and gap-cause prediction method for team knowledge has the following beneficial technical effects: 1. This application constructs a team knowledge manifold by first fusing and reducing the dimensionality of the team's multimodal knowledge data, including knowledge state points and neighborhood similarity associations. Then, it trains a reinforcement learning policy network by combining historical collaboration strategy actions and collaboration reward values. Subsequently, it reduces the dimensionality of the team knowledge manifold to a two-dimensional grid and generates a manifold policy consistency map through Shannon entropy mapping. Finally, it segments the map to identify topological defect regions. This design achieves effective detection of topological defects in the team knowledge manifold, alleviating the technical problem of topological defects generated in the knowledge manifold when introducing new members with extremely heterogeneous cognitive patterns. The boundaries of these topological defects are difficult to detect effectively, which disrupts the continuity and stability of team knowledge collaboration.

[0017] 2. This application transforms the problem of topological defect detection on high-dimensional manifolds into a two-dimensional image segmentation problem. It indirectly detects topological defects by monitoring the behavioral consistency of reinforcement learning strategies, avoiding complex manifold geometric analysis. By utilizing Shannon entropy to quantify strategy inconsistency, it can reflect discontinuous regions on the manifold, thus enabling subsequent defect segmentation based on mature image processing techniques, which is computationally reliable and efficient.

[0018] 3. This application uses a reinforcement learning policy network and Shannon entropy to quantify policy consistency, associating each knowledge state point on the team's knowledge manifold with collaborative policy actions. Shannon entropy is used to reflect the consistency of policy actions within the same region. The lower the Shannon entropy value, the greater the policy divergence, corresponding to the manifold tear point. This transforms the abstract policy transmission break problem into quantifiable grayscale image features, alleviating the manifold tear problem that traditional reinforcement learning only focuses on policy optimization and cannot perceive the significant differences in policies at geometrically similar points. Attached Figure Description

[0019] Figure 1 This is a flowchart of the reinforcement learning-based team knowledge multimodal perception and gap cause prediction method of the present invention. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] like Figure 1 The reinforcement learning-based team knowledge multimodal perception and gap cause prediction method shown includes the following steps: S1. Collect multimodal knowledge data for each collaboration stage in the team's historical collaboration process, historical collaboration strategies and actions for each collaboration stage, and historical collaboration result data corresponding to historical collaboration strategies and actions; In an optional embodiment, in S1, the multimodal knowledge data includes text-based knowledge data, audio-based knowledge data, behavioral knowledge data, and image-based knowledge data; Historical collaboration results data include task completion time, task output quality indicators, and collaboration cost indicators; Multimodal knowledge data from each collaborative stage are fused into a Euclidean space feature vector; In an optional embodiment, in S1, the multimodal knowledge data of each collaborative stage is fused into a Euclidean space feature vector, as follows: The TF-IDF algorithm is used to extract features from text-based knowledge data to generate TF-IDF feature vectors. Principal component analysis is then used to reduce the dimensionality of the TF-IDF feature vectors to 256 dimensions, resulting in 256-dimensional semantic feature vectors. The Mel-Cepstral Coefficient algorithm is used to extract time-frequency domain features from audio knowledge data; then, a one-dimensional convolutional neural network is used to encode the time-frequency domain features into a 128-dimensional acoustic feature vector. The behavioral knowledge data is preprocessed using min-max normalization, and then the features are mapped and encoded using a fully connected neural network to obtain a 64-dimensional behavioral feature vector. A convolutional neural network is used to encode image-type knowledge data into a 256-dimensional visual feature vector. A multilayer perceptron is used to map and fuse 256-dimensional semantic feature vectors, 128-dimensional acoustic feature vectors, 64-dimensional behavioral feature vectors, and 256-dimensional visual feature vectors into a 128-dimensional Euclidean space feature vector.

[0022] Obtain the Euclidean space feature vectors of all collaborative stages to form a set of Euclidean space feature vectors; The manifold learning algorithm is used to reduce the dimensionality of each Euclidean space feature vector in the set of Euclidean space feature vectors to obtain multiple low-dimensional manifold vectors. The low-dimensional manifold vectors are used as low-dimensional manifold features, and a team knowledge manifold is constructed based on the multiple low-dimensional manifold features. In an optional embodiment, in S1, a manifold learning algorithm is used to reduce the dimensionality of each Euclidean space feature vector in the set of Euclidean space feature vectors, resulting in multiple low-dimensional manifold vectors, as follows: The manifold learning algorithm is used to reduce the dimensionality of each Euclidean space feature vector in the Euclidean space feature vector set to 8, 12, or 16 dimensions, resulting in multiple low-dimensional feature vectors. In an optional embodiment, in S1, a team knowledge manifold is constructed based on multiple low-dimensional manifold features, as follows: The geodesic distance between each low-dimensional manifold feature is obtained through a manifold learning algorithm. The reciprocal of the geodesic distance is used as the similarity quantification value. Two low-dimensional manifold features with similarity quantification values ​​greater than a preset similarity threshold are regarded as neighborhood feature pairs. All neighborhood feature pairs and their corresponding similarity quantification values ​​are integrated into a structured association relationship to generate neighborhood similarity associations between features. All low-dimensional manifold features and neighborhood similarity associations between features are integrated into an overall topological structure, which is used as the team knowledge manifold. Each low-dimensional manifold feature corresponds to a point on the team knowledge manifold, which serves as a knowledge state point. Each knowledge state point uniquely corresponds to the team's knowledge state in a single collaboration phase.

[0023] S2. Based on the team's knowledge manifold, generate a reinforcement learning policy network. The reinforcement learning policy network takes knowledge state points as input and outputs the probability distribution of the collaborative policy action corresponding to the knowledge state point. In an optional embodiment, in S2, a reinforcement learning policy network is generated based on the team knowledge manifold, as follows: Obtain the historical collaboration strategy actions corresponding to each knowledge state point in the team's knowledge manifold; The task completion time, task output quality index, and collaboration cost index in the historical collaboration result data are normalized. Based on the preset weights of task completion time, task output quality data, and collaboration cost data, a linear weighted scoring algorithm is used to obtain the collaboration benefit value of the normalized task completion time, task output quality index, and collaboration cost index, which is used as the collaboration benefit value corresponding to each knowledge state point. As an explanation, the preset values ​​for the weights of task completion time, task output quality data, and collaboration cost data can be set based on the subjective evaluation of the importance of task completion time, task output quality, and collaboration cost by team managers or business experts, and the weights of each importance can be normalized. The preset values ​​for the weights of task completion time, task output quality data, and collaboration cost data can also be set by the system administrator through the system configuration interface according to different business scenarios. The training sample set is obtained by using each knowledge state point on the team knowledge manifold, the historical collaboration strategy action corresponding to each knowledge state point, and the collaboration benefit value corresponding to each knowledge state point as training samples. Based on the training sample set, a neural network algorithm is used to generate a neural policy network. The neural policy network takes knowledge state points as input and outputs the probability distribution of the cooperative policy action corresponding to the knowledge state point. Based on the training sample set and the preset team collaboration goal, with the objective of maximizing the cumulative collaboration benefit value under the preset team collaboration goal, the reinforcement learning algorithm is used to iteratively update the network parameters of the neural policy network to obtain the reinforcement learning policy network. The reinforcement learning policy network takes knowledge state points as input and the output of the reinforcement learning policy network is the probability distribution of the collaboration policy action corresponding to the knowledge state point. S3. Generate manifold feature images based on the team knowledge manifold and reinforcement learning policy network; In an optional embodiment, in S3, a manifold feature image is generated based on the team knowledge manifold and the reinforcement learning policy network, as follows: S3.1 Project the low-dimensional manifold features in the team knowledge manifold to a two-dimensional Euclidean space using dimensionality reduction techniques to obtain the corresponding two-dimensional point cloud; divide the region where the two-dimensional point cloud is located in the two-dimensional Euclidean space into a regular grid of M×N, with each grid serving as a pixel grid; M and N are both positive integers greater than 1, and M and N may be equal or unequal. In an optional embodiment, in S3.1, the dimensionality reduction technique includes an autoencoder; the autoencoder is used to map the low-dimensional manifold features in the team knowledge manifold to a two-dimensional Euclidean space; S3.2 For each pixel grid, count all the knowledge state points corresponding to the two-dimensional point cloud that fall within it, and use them as pixel knowledge state points. The probability distribution of the cooperative policy action corresponding to the pixel knowledge state point is obtained by using a reinforcement learning policy network, which is then used as the probability distribution of the pixel cooperative policy action. Obtain the Shannon entropy of the probability distribution of pixel cooperative strategy actions, and use it as the Shannon entropy corresponding to the pixel knowledge state point. Aggregate the Shannon entropy corresponding to all prime knowledge state points within the pixel grid to obtain the Shannon entropy value corresponding to the pixel grid. S3.3 Obtain the negative value of Shannon entropy based on the Shannon entropy value corresponding to the pixel grid; The negative values ​​of the Shannon entropy corresponding to the pixel grid are normalized and mapped to the gray value range of 0-255. The normalized gray value is used as the gray intensity value of the corresponding pixel grid. After traversing all pixel grids, an M×N grayscale image is obtained, which is used as the manifold policy consistency map.

[0024] As an illustration, the Shannon entropy corresponding to all elemental knowledge state points within the pixel grid is aggregated, for example, by taking the average value.

[0025] This application uses a reinforcement learning policy network and Shannon entropy to quantify policy consistency, associating each knowledge state point on the team's knowledge manifold with collaborative policy actions. Shannon entropy is used to reflect the consistency of policy actions within the same region. The lower the Shannon entropy value, the greater the policy divergence, corresponding to the manifold tear point. This transforms the abstract policy transmission break problem into quantifiable grayscale image features, alleviating the manifold tear problem that traditional reinforcement learning only focuses on policy optimization and cannot perceive the significant differences in policies at geometrically similar points.

[0026] In an optional embodiment, in step S3.3, the normalization process uses the minimum-maximum normalization method to linearly map the negative value of Shannon entropy to the gray value range of 0-255. S4. Perform image segmentation processing on the manifold strategy consistency map, and identify regions with gray intensity values ​​below a preset threshold as topological defect regions.

[0027] This application constructs a team knowledge manifold by first fusing and reducing the dimensionality of the team's multimodal knowledge data, including knowledge state points and neighborhood similarity associations. Then, it trains a reinforcement learning policy network by combining historical collaboration strategy actions and collaboration reward values. Subsequently, it reduces the dimensionality of the team knowledge manifold to a two-dimensional grid and generates a manifold policy consistency map through Shannon entropy mapping. Finally, it segments the map to identify topological defect regions. This design enables the effective detection of topological defects in the team knowledge manifold, alleviating the technical problem of topological defects generated in the knowledge manifold when introducing new members with extremely heterogeneous cognitive patterns. The boundaries of these topological defects are difficult to detect effectively, which disrupts the continuity and stability of team knowledge collaboration.

[0028] In an optional embodiment, in S4, the manifold strategy consistency map is processed by image segmentation, and regions with grayscale intensity values ​​below a preset threshold are identified as topological defect regions, as follows: Based on a preset threshold, pixel grids with grayscale intensity values ​​lower than the preset threshold are marked as abnormal pixel grids; By forming a connected region from adjacent abnormal pixel grids, one or more connected regions can be obtained; Connected regions are treated as topological defect regions.

[0029] This application transforms the problem of topological defect detection on high-dimensional manifolds into a two-dimensional image segmentation problem. It indirectly detects topological defects by monitoring the behavioral consistency of reinforcement learning strategies, avoiding complex manifold geometric analysis. By utilizing Shannon entropy to quantify strategy inconsistency, it can reflect discontinuous regions on the manifold, thus enabling subsequent defect segmentation based on mature image processing techniques, which is computationally reliable and efficient.

[0030] For clarification, "acquisition" in this application refers to obtaining the required content or data using existing technical means.

[0031] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0032] In the embodiments provided by this invention, it should be understood that the disclosed system or method can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for instance, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0033] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0034] Furthermore, the functional modules in the various embodiments of this invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in a combination of hardware and software functional modules.

[0035] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the basic characteristics of the present invention.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A team knowledge multimodal perception and gap cause prediction method based on reinforcement learning, characterized in that, Includes the following steps: S1. Collect multimodal knowledge data, historical collaboration strategies and actions, and historical collaboration results data for each collaboration stage in the team's historical collaboration process; fuse the multimodal knowledge data of each collaboration stage into Euclidean space feature vectors; obtain the Euclidean space feature vectors of all collaboration stages to form a set of Euclidean space feature vectors; use a manifold learning algorithm to reduce the dimensionality of each Euclidean space feature vector in the set of Euclidean space feature vectors to obtain multiple low-dimensional manifold vectors; use the low-dimensional manifold vectors as low-dimensional manifold features; and construct the team knowledge manifold based on the multiple low-dimensional manifold features. S2. Generate a reinforcement learning policy network based on the team's knowledge manifold; S3. Generate manifold feature images based on the team knowledge manifold and reinforcement learning policy network; S4. Perform image segmentation processing on the manifold strategy consistency map, and identify regions with gray intensity values ​​below a preset threshold as topological defect regions.

2. The method for multimodal perception and gap-cause prediction of team knowledge based on reinforcement learning according to claim 1, characterized in that, In S1, multimodal knowledge data includes text-based knowledge data, audio-based knowledge data, behavioral knowledge data, and image-based knowledge data; Historical collaboration results data include task completion time, task output quality indicators, and collaboration cost indicators.

3. The team knowledge multimodal perception and gap cause prediction method based on reinforcement learning according to claim 2, characterized in that, In S1, the multimodal knowledge data from each collaborative stage is fused into a Euclidean space feature vector, as follows: The TF-IDF algorithm is used to extract features from text-based knowledge data to generate TF-IDF feature vectors. Principal component analysis is then used to reduce the dimensionality of the TF-IDF feature vectors to 256 dimensions, resulting in 256-dimensional semantic feature vectors. The Mel-Cepstral Coefficient algorithm is used to extract time-frequency domain features from audio knowledge data; The time-domain and frequency-domain features are then encoded into 128-dimensional acoustic feature vectors using a one-dimensional convolutional neural network. The behavioral knowledge data is preprocessed using min-max normalization, and then the features are mapped and encoded using a fully connected neural network to obtain a 64-dimensional behavioral feature vector. A convolutional neural network is used to encode image-type knowledge data into a 256-dimensional visual feature vector. A multilayer perceptron is used to map and fuse 256-dimensional semantic feature vectors, 128-dimensional acoustic feature vectors, 64-dimensional behavioral feature vectors, and 256-dimensional visual feature vectors into a 128-dimensional Euclidean space feature vector.

4. The method for team knowledge multimodal perception and gap cause prediction based on reinforcement learning according to claim 3, characterized in that, In S1, a manifold learning algorithm is used to reduce the dimensionality of each Euclidean space eigenvector in the set of Euclidean space eigenvectors, resulting in multiple low-dimensional manifold vectors, as follows: The manifold learning algorithm is used to reduce the dimensionality of each Euclidean space feature vector in the set of Euclidean space feature vectors to 8, 12 or 16 dimensions, resulting in multiple low-dimensional feature vectors.

5. The team knowledge multimodal perception and gap cause prediction method based on reinforcement learning according to claim 4, characterized in that, In S1, the team knowledge manifold is constructed based on multiple low-dimensional manifold features, as follows: The geodesic distance between each low-dimensional manifold feature is obtained through a manifold learning algorithm. The reciprocal of the geodesic distance is used as the similarity quantification value. Two low-dimensional manifold features with similarity quantification values ​​greater than a preset similarity threshold are regarded as neighborhood feature pairs. All neighborhood feature pairs and their corresponding similarity quantification values ​​are integrated into a structured association relationship to generate neighborhood similarity associations between features. All low-dimensional manifold features and neighborhood similarity associations between features are integrated into an overall topological structure, which is used as the team knowledge manifold. Each low-dimensional manifold feature corresponds to a point on the team knowledge manifold, which serves as a knowledge state point. Each knowledge state point uniquely corresponds to the team's knowledge state in a single collaboration phase.

6. The method for team knowledge multimodal perception and gap cause prediction based on reinforcement learning according to claim 5, characterized in that, In S2, a reinforcement learning policy network is generated based on the team's knowledge manifold, as follows: Obtain the historical collaboration strategy actions corresponding to each knowledge state point in the team's knowledge manifold; The task completion time, task output quality index, and collaboration cost index in the historical collaboration result data are normalized. Based on the preset weights of task completion time, task output quality data, and collaboration cost data, a linear weighted scoring algorithm is used to obtain the collaboration benefit value of the normalized task completion time, task output quality index, and collaboration cost index, which is used as the collaboration benefit value corresponding to each knowledge state point. The training sample set is obtained by using each knowledge state point on the team knowledge manifold, the historical collaboration strategy action corresponding to each knowledge state point, and the collaboration benefit value corresponding to each knowledge state point as training samples. Based on the training sample set, a neural network algorithm is used to generate a neural policy network. The neural policy network takes knowledge state points as input and outputs the probability distribution of the cooperative policy action corresponding to the knowledge state point. Based on the training sample set and the preset team collaboration goal, with the objective of maximizing the cumulative collaboration benefit value under the preset team collaboration goal, the network parameters of the neural policy network are iteratively updated using a reinforcement learning algorithm to obtain the reinforcement learning policy network. The reinforcement learning policy network takes knowledge state points as input and outputs the probability distribution of the collaboration policy action corresponding to the knowledge state point.

7. The method for team knowledge multimodal perception and gap cause prediction based on reinforcement learning according to claim 6, characterized in that, In S3, a manifold feature image is generated based on the team knowledge manifold and the reinforcement learning policy network, as follows: S3.1 Project the low-dimensional manifold features in the team knowledge manifold to a two-dimensional Euclidean space using dimensionality reduction techniques to obtain the corresponding two-dimensional point cloud; in the two-dimensional Euclidean space, divide the region where the two-dimensional point cloud is located into an M×N regular grid, with each grid serving as a pixel grid. M and N are both positive integers greater than 1, and M and N may be equal or unequal; S3.2 For each pixel grid, count all the knowledge state points corresponding to the two-dimensional point cloud that fall within it, and use them as pixel knowledge state points. The probability distribution of the cooperative policy action corresponding to the pixel knowledge state point is obtained by using a reinforcement learning policy network, which is then used as the probability distribution of the pixel cooperative policy action. Obtain the Shannon entropy of the probability distribution of pixel cooperative strategy actions, and use it as the Shannon entropy corresponding to the pixel knowledge state point. Aggregate the Shannon entropy corresponding to all prime knowledge state points within the pixel grid to obtain the Shannon entropy value corresponding to the pixel grid. S3.3 Obtain the negative value of Shannon entropy based on the Shannon entropy value corresponding to the pixel grid; The negative values ​​of the Shannon entropy corresponding to the pixel grid are normalized and mapped to the gray value range of 0-255. The normalized gray value is used as the gray intensity value of the corresponding pixel grid. After traversing all pixel grids, an M×N grayscale image is obtained, which is used as the manifold policy consistency map.

8. The team knowledge multimodal perception and gap cause prediction method based on reinforcement learning according to claim 7, characterized in that, In S3.1, dimensionality reduction techniques include autoencoders; autoencoders are used to map the features of each low-dimensional manifold in the team knowledge manifold to a two-dimensional Euclidean space.

9. The team knowledge multimodal perception and gap cause prediction method based on reinforcement learning according to claim 7, characterized in that, In step S3.3, the normalization process uses the minimum-maximum normalization method to linearly map the negative value of Shannon entropy to the gray value range of 0-255.

10. The team knowledge multimodal perception and gap cause prediction method based on reinforcement learning according to claim 7, characterized in that, In S4, the manifold strategy consistency map is segmented to identify regions with grayscale intensity values ​​below a preset threshold as topological defect regions, as follows: Based on a preset threshold, pixel grids with grayscale intensity values ​​lower than the preset threshold are marked as abnormal pixel grids; By forming a connected region from adjacent abnormal pixel grids, one or more connected regions can be obtained; Connected regions are treated as topological defect regions.

Images