Decision fusion method and device, electronic equipment and storage medium

Abstract

The invention provides a decision fusion method and device, electronic equipment and a storage medium, and relates to the technical field of data processing, and the decision fusion method comprises the steps: obtaining basic data of a to-be-decided event and a plurality of candidate decisions corresponding to the basic data; scoring the plurality of candidate decisions based on the basic data by using the plurality of basic reward models to generate an initial scoring matrix; generating an initial decision weight based on the basic data; based on the basic data, the initial scoring matrix and the initial decision weight, fusing to obtain a meta-state matrix; calibrating the initial decision weight based on the meta-state matrix to generate a target decision weight; and based on the initial scoring matrix and the target decision weight, determining a fusion decision result. According to the invention, the accuracy of a multi-decision fusion result can be improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a decision fusion method, apparatus, electronic device, and storage medium. Background Technology

[0002] In complex decision-making scenarios involving multi-source information fusion, it is often necessary to integrate opinions from multiple different assessment subjects or with different professional perspectives to form a more comprehensive and reliable final decision. This multi-decision fusion scenario is widespread across various fields. For example, in the financial sector, investment committees need to synthesize the opinions of market analysts, technical experts, and legal advisors to assess the feasibility of a merger or acquisition; in the industrial sector, fault diagnosis systems need to integrate information from multiple sensors such as temperature, vibration, and pressure to determine the root cause of equipment failure; in corporate management, strategic decisions require the combined judgments of multiple departments such as sales, R&D, and production; and in the medical field, the diagnosis and treatment of complex diseases requires multidisciplinary consultation to determine the final treatment plan. In these scenarios, the key to improving the quality of decision fusion lies in how to quickly and effectively aggregate the wisdom of different experts, identify and handle conflicts between opinions, and ultimately arrive at an optimized decision that surpasses the capabilities of any single expert.

[0003] Traditional decision fusion methods typically employ simple fixed-weighting, score averaging, or rule-based aggregation. These methods treat each evaluation dimension as a static, independent voting device, failing to delve into the intrinsic connections, contradictions, and synergies between different decision opinions. They also lack quantitative assessments and dynamic calibration mechanisms for the credibility of the decision-making process itself, leading to a disconnect between weight allocation and actual decision-making needs, and consequently, poor accuracy in the final multi-decision fusion results. Taking multidisciplinary medical diagnosis as an example, driven by precision medicine and hierarchical medical policies, multidisciplinary teams (MDTs) have become the core model for the diagnosis and treatment of complex diseases (such as rare diseases, malignant tumors, and cross-system chronic diseases). Through the collaborative judgment of experts from different disciplines such as cardiology, neurology, and immunology, the limitations of a single disciplinary perspective can be effectively avoided, reducing the risk of misdiagnosis and missed diagnosis. However, in practice, traditional MDTs often use fixed weighting when integrating the collaborative judgments of different experts. For example, cardiology has a weight of 0.6, neurology has a weight of 0.3, and immunology has a weight of 0.1. Alternatively, they may simply count the number of experts who support a certain diagnosis, completely ignoring the heterogeneity of cases and the correlation between expert opinions. This results in a disconnect between weighting and actual treatment needs.

[0004] Therefore, improving the accuracy of multi-decision fusion results is a technical challenge that urgently needs to be addressed in this field. Summary of the Invention

[0005] This invention provides a decision fusion method, apparatus, electronic device, and storage medium that can improve the accuracy of multi-decision fusion results.

[0006] This invention provides a decision fusion method, comprising: Acquire the basic data of the event to be decided and its corresponding multiple candidate decisions; Using multiple basic reward models, the multiple candidate decisions are scored based on the basic data to generate an initial score matrix; Based on the aforementioned basic data, initial decision weights are generated; Based on the aforementioned basic data, the initial scoring matrix, and the initial decision weights, a meta-state matrix is ​​obtained through fusion. Based on the meta-state matrix, the initial decision weights are calibrated to generate target decision weights; Based on the initial scoring matrix and the target decision weights, the fusion decision result is determined.

[0007] According to a decision fusion method provided by the present invention, before determining the fusion decision result based on the initial scoring matrix and the target decision weights, the method further includes: Based on the aforementioned meta-state matrix, decision divergence detection is performed, and divergence detection results are generated. Based on the meta-state matrix, a decision calibration vector is generated for the plurality of candidate decisions; The process of determining the fusion decision result based on the initial scoring matrix and the target decision weights includes: Based on the initial scoring matrix, the target decision weights, the divergence detection results, and the decision calibration vector, the fusion decision result is determined.

[0008] According to a decision fusion method provided by the present invention, the step of performing decision divergence detection based on the meta-state matrix and generating divergence detection results includes: Calculate the similarity between each row vector in the meta-state matrix, and determine the consensus score based on the similarity; The meta-state matrix is ​​mapped to a mapping score matrix, and the score variance corresponding to the multiple candidate decisions is calculated based on the mapping score matrix. The score variance is compared with a preset threshold, and the focus of decision disagreement is determined based on the comparison results; The disagreement detection result is generated based on the decision disagreement focus and the consensus score.

[0009] According to a decision fusion method provided by the present invention, generating a decision calibration vector for the plurality of candidate decisions based on the meta-state matrix includes: Sequence encoding features are obtained by using a bidirectional long short-term memory network to perform sequence modeling on the meta-state matrix. The sequence encoding features are processed through a first multilayer perceptron network to generate the decision calibration vector.

[0010] According to a decision fusion method provided by the present invention, determining the fusion decision result based on the initial scoring matrix, the target decision weight, the divergence detection result, and the decision calibration vector includes: The initial fusion score corresponding to the multiple candidate decisions is obtained by weighting the target decision weights and the initial score matrix. Based on the divergence detection results, a consensus confidence vector is determined; The initial fusion score, the consensus confidence vector, and the decision calibration vector are summed to obtain the target fusion score corresponding to each candidate decision; The fusion decision result includes the multiple candidate decisions and their corresponding target fusion scores.

[0011] According to a decision fusion method provided by the present invention, the step of fusing the basic data, the initial scoring matrix, and the initial decision weights to obtain a meta-state matrix includes: The scoring vectors of each basic reward model in the initial scoring matrix are mapped to feature vectors of a preset dimension to obtain the model scoring state matrix; After expanding the initial decision weights, they are fused with the model scoring state matrix to obtain a fused feature matrix; The fused feature matrix is ​​globally modulated with the context feature vector of the basic data, and the modulated feature matrix is ​​encoded to obtain the meta-state matrix.

[0012] According to a decision fusion method provided by the present invention, the step of calibrating the initial decision weights based on the meta-state matrix to generate target decision weights includes: Global average pooling is performed on the meta-state matrix along the dimension of the basic reward model to obtain the global meta-state vector; The global meta-state vector is mapped to the weight scores corresponding to each basic reward model through a second multilayer perceptron network. The weight scores are normalized to generate the target decision weights.

[0013] The present invention also provides a decision fusion device, comprising: The data acquisition module is used to acquire the basic data of the event to be decided and its corresponding multiple candidate decisions; The first generation module is used to score the multiple candidate decisions based on the basic data using multiple basic reward models to generate an initial scoring matrix. The second generation module is used to generate initial decision weights based on the basic data; The data fusion module is used to fuse the basic data, the initial scoring matrix, and the initial decision weights to obtain a meta-state matrix; The weight calibration module is used to calibrate the initial decision weights based on the meta-state matrix to generate target decision weights. The result determination module is used to determine the fusion decision result based on the initial scoring matrix and the target decision weights.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the decision fusion method as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the decision fusion method as described in any of the preceding claims.

[0016] The decision fusion method, apparatus, electronic device, and storage medium provided by this invention first acquire basic data of the event to be decided and its corresponding multiple candidate decisions. Then, using multiple basic reward models, the multiple candidate decisions are scored based on the basic data to generate an initial scoring matrix. By decomposing the complex decision problem into multiple professional sub-perspective evaluation tasks, each basic reward model performs refined scoring of each evaluation task in its area of ​​expertise to form the initial scoring matrix. This multi-perspective evaluation avoids the bias of a single model. Next, based on the basic data, initial decision weights are generated. These initial decision weights are used to evaluate the importance of each basic reward model in the current event to be decided. By establishing a dynamic mapping between basic data features and model importance, the fundamental problem of traditional decision fusion methods using fixed weights, leading to a disconnect between weight allocation and actual decision requirements, is solved. Finally, based on the basic data, the initial scoring matrix, and the initial decision weights, a meta-state matrix is ​​obtained through fusion. This meta-state matrix breaks through the limitations of traditional single data input, retaining the professional judgments of each basic reward model and capturing the synergistic and divergent relationships among expert opinions, as well as the correlation characteristics between basic data and candidate decisions. It provides a global and comprehensive information foundation for subsequent weight calibration, solving the problem of decision-making bias caused by relying solely on basic data or model scores. Finally, based on the meta-state matrix, the initial decision weights are dynamically calibrated to generate target decision weights. Based on the initial scoring matrix and target decision weights, the fusion decision result is determined. By dynamically calibrating the initial decision weights using global information from the meta-state matrix, the target decision weights can both align with the current decision event and respond to the synergy of expert opinions, completely resolving the adaptability defects of fixed weights and significantly improving the accuracy of the fusion decision result. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts illustrating the decision fusion method provided by the present invention; Figure 2 This is the second flowchart of the decision fusion method provided by the present invention; Figure 3 This is the third flowchart of the decision fusion method provided by the present invention; Figure 4 This is the fourth flowchart of the decision fusion method provided by the present invention; Figure 5This is the fifth flowchart of the decision fusion method provided by the present invention; Figure 6 This is the sixth flowchart of the decision fusion method provided by the present invention; Figure 7 This is a schematic diagram of the decision fusion device provided by the present invention; Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] This invention proposes a decision fusion method, device, electronic device, and storage medium, which are described below in conjunction with... Figures 1-8 Describe it.

[0021] Figure 1 This is one of the flowcharts illustrating the decision fusion method provided by the present invention, such as... Figure 1 As shown, the decision fusion method includes steps S110, S120, S130, S140, S150 and S160.

[0022] Step S110: Obtain the basic data of the event to be decided and its corresponding multiple candidate decisions.

[0023] In this embodiment of the invention, generalized applications are supported for various decision-making scenarios, such as medical diagnosis scenarios, industrial fault diagnosis scenarios, investment decision-making scenarios in the financial field, and decision-making scenarios in enterprise management.

[0024] The basic data of the event to be decided refers to the raw information used for decision analysis.

[0025] Candidate decisions refer to a series of possible decision outcomes for a given decision event.

[0026] In medical diagnostic scenarios, the basic data for events requiring decision-making specifically refers to the patient's structured medical record text, including but not limited to symptoms, signs, laboratory test results, and imaging reports. Candidate decisions specifically refer to the preliminary diagnosed disease types that require further fusion, such as pneumonia, acute bronchitis, upper respiratory tract infection, and mycoplasma pneumonia.

[0027] In industrial fault diagnosis scenarios, the basic data for the event to be decided specifically refers to the operational data of the machine or equipment used for fault diagnosis, including but not limited to textual descriptions from maintenance personnel, sensor data, and image data. Candidate decisions specifically refer to the initially determined fault types that require further integration. For example, for fault diagnosis of a wind turbine generator set, the basic data may include vibration spectrum diagrams, nacelle temperature time-series data, and fault information described by maintenance personnel such as "recent intermittent abnormal noises and power output fluctuations." Candidate decisions include, but are not limited to: bearing wear, blade imbalance, gearbox failure, electrical control system failure, and tower resonance.

[0028] For other scenarios such as investment decision-making in the financial sector and decision-making in corporate management, corresponding basic data and candidate decisions can be set according to the actual situation. No specific limitations are made here.

[0029] In this embodiment of the invention, the medical diagnosis scenario is used as an example for illustration.

[0030] Step S120: Using multiple basic reward models, score the multiple candidate decisions based on the basic data to generate an initial score matrix.

[0031] A foundational reward model refers to a deep learning model trained with specific domain knowledge, capable of professionally evaluating and scoring input information, thus ensuring its expertise in that domain. The initial model for a foundational reward model primarily employs a pre-trained language model (PLM), especially a text semantic matching model based on the Transformer architecture.

[0032] In medical diagnostic scenarios, the basic reward model specifically refers to multiple AI (Artificial Intelligence) models from different specialties, such as cardiology models, respiratory models, gastroenterology models, neurology models, and immunology models. Each specialty model is pre-trained and fine-tuned using datasets specific to its specialty. For example, a cardiology model can be pre-trained and fine-tuned using a myocardial infarction case database from cardiology, while a neurology model can be pre-trained and fine-tuned using a stroke case database from neurology.

[0033] In industrial fault diagnosis scenarios, the basic reward model specifically refers to an AI model with multiple different evaluation dimensions. For example, in the fault diagnosis scenario of a wind turbine generator mentioned above, the basic reward model could be: a mechanical vibration analysis model, a thermodynamic analysis model, an electrical signal analysis model, an acoustic analysis model, an operation and maintenance experience model, etc.

[0034] After obtaining the basic data of the event to be decided and its corresponding multiple candidate decisions, several basic reward models are used to score the candidate decisions based on the basic data, generating an initial score matrix. The scores given by each basic reward model to the multiple candidate decisions can be viewed as scores given by different experts to each candidate decision.

[0035] Specifically, the basic data is first concatenated with multiple candidate decisions to obtain multiple matching pairs to be scored. Each matching pair includes basic data and a candidate decision. If the basic data is P, and the candidate decisions are {D1, D2, ..., D...}, then... m The system will generate m matching pairs to be scored: "P + candidate decision D1", "P + candidate decision D2", ..., "P + candidate decision D1". m ".

[0036] For example, in a medical diagnosis scenario, the basic data is a medical record text: "Headache for 3 weeks, accompanied by orbital pain and limb weakness, symptoms recurred after hormone treatment." The corresponding multiple candidate decisions are multiple candidate diagnoses: multiple sclerosis, neuromyelitis optica, and antimyelin antibody-related diseases. In this case, using a "medical record text + diagnosis" concatenation strategy, three matching pairs can be obtained for scoring: ① "Headache for 3 weeks, accompanied by orbital pain and limb weakness, symptoms recurred after hormone treatment" + multiple sclerosis; ② "Headache for 3 weeks, accompanied by orbital pain and limb weakness, symptoms recurred after hormone treatment" + neuromyelitis optica; ③ "Headache for 3 weeks, accompanied by orbital pain and limb weakness, symptoms recurred after hormone treatment" + antimyelin antibody-related diseases. This concatenation method allows the specialty model to focus on assessing the match between a specific diagnosis and the medical record, avoiding the ambiguity of generalized judgments.

[0037] Then, each matching pair to be scored is input into multiple basic reward models, so that multiple basic reward models can be used to score multiple matching pairs to be scored, and the confidence score of the matching degree output of each basic reward model for each matching pair to be scored is obtained. The scores are stacked to obtain a scoring matrix, which is denoted as the initial scoring matrix.

[0038] Assume there are n basic reward models and m matching pairs to be rated. Correspondingly, the initial rating matrix S is an n x m (n x m) matrix. Here, S[i][j] represents the matching score (0-1) of the i-th basic reward model for the j-th candidate decision. The higher S[i][j] is, the greater the probability that the i-th basic reward model considers the j-th candidate decision to be.

[0039] The above method transforms the evaluation judgments of each basic reward model into calculable quantitative scores, and provides a foundation for subsequent decision fusion through a standardized initial score matrix.

[0040] Step S130: Generate initial decision weights based on the basic data.

[0041] The initial decision weight is an initial importance assessment or trust level given for n basic reward models (i.e. decision dimensions) based on the basic data. It is an n-dimensional vector.

[0042] In one implementation, keywords are extracted from basic data, and matching degrees are calculated based on the extracted keywords and preset feature templates corresponding to each basic reward model. Initial decision weights are determined based on the matching degree calculation results. The preset feature templates can be a set of feature words. During the matching degree calculation, the number of common words between the keywords and the preset feature templates can be counted to determine the initial decision weights. These initial decision weights are then calculated through normalization.

[0043] For example, in a medical diagnosis scenario, the basic data is: "A 65-year-old male patient presented with sudden onset of severe, tearing chest pain radiating to his back for 2 hours, accompanied by profuse sweating and dyspnea. He has a history of hypertension. Physical examination revealed asymmetrical blood pressure in both upper limbs." The extracted keywords are {sudden onset, chest pain (tearing), radiating pain (back), profuse sweating, dyspnea, hypertension, asymmetrical blood pressure}. The basic reward models include cardiology, respiratory medicine, gastroenterology, neurology, and immunology models. The pre-defined feature templates for the cardiology model include: {chest pain, chest tightness, palpitations, electrocardiogram, myocardial enzymes, heart failure...}. The shared word between the extracted keywords and the pre-defined feature templates is {chest pain}, with a quantity of 1. Assuming the shared word quantities with the other four basic reward models are 2, 1, 0, and 0 respectively, after normalization, the initial decision weights are [0.25, 0.5, 0.25, 0, 0].

[0044] In another implementation, the basic data is encoded to obtain a context feature vector. A Multilayer Perceptron (MLP) network is then used to map the context feature vector to weight scores for each decision dimension. These weight scores are then normalized to generate initial decision weights. Specifically, a BERT (Bidirectional Encoder Representation from Transformers) encoder is first used to deeply encode the basic data (such as medical record text), generating a 512-dimensional context feature vector that captures key information from the medical record text. Subsequently, the context feature vector is input into a three-layer MLP network with hidden layer dimensions of 512, 256, and n (the number of basic reward models), and activated by the Softmax (normalization exponent) function to obtain the initial decision weights.

[0045] It should be noted that the second implementation method described above, compared to the first implementation method, can capture complex semantic relationships and contextual information, thereby improving the accuracy of the initial decision weights, but its computational cost is relatively higher.

[0046] It should also be noted that by determining the initial decision weights, the importance of each decision dimension (corresponding to the basic reward model dimensions) in the current decision-making event can be preliminarily assessed. Taking a medical diagnosis scenario as an example, the core symptoms in medical records are often directly associated with specific specialties. For instance, a medical record with "headache accompanied by blurred vision" would be given a higher initial weight for neurology, while "dyspnea accompanied by lower extremity edema" would increase the initial weight for cardiology. This prediction based on medical record characteristics enables the weight allocation to initially adapt to the scenario, laying the foundation for subsequent fine-tuning of the weights and solving the adaptability problem of using the same specialty weights for all cases in existing technologies.

[0047] Step S140: Based on the basic data, the initial scoring matrix, and the initial decision weights, a meta-state matrix is ​​obtained by fusing them.

[0048] The basic data, initial scoring matrix, and initial decision weights are fused to obtain a meta-state matrix, denoted as M. By constructing a meta-state matrix containing multi-dimensional information, the decision bias problem caused by the one-sidedness of a single information source in existing technologies is solved, providing a comprehensive information foundation for subsequent weight optimization and disagreement arbitration.

[0049] Specifically, the scoring vectors of each basic reward model in the initial scoring matrix are mapped to feature vectors of a preset dimension to obtain the model scoring state matrix. Then, the initial decision weights are expanded and fused with the model scoring state matrix to obtain a fused feature matrix. Finally, the fused feature matrix is ​​globally modulated with the context feature vectors of the basic data, and the modulated feature matrix is ​​encoded to obtain the meta-state matrix. The specific execution process can be found in the following embodiment, which will not be elaborated here.

[0050] Step S150: Based on the meta-state matrix, calibrate the initial decision weights to generate target decision weights.

[0051] Based on the group consensus reflected in the meta-state matrix, the initial decision weights are dynamically calibrated to generate the final target decision weights for decision fusion.

[0052] Specifically, the global average pooling of the multilayer perceptron's meta-state matrix along the dimension of the basic reward model is performed to obtain the global meta-state vector. Then, through the second multilayer perceptron network, the global meta-state vector of the multilayer perceptron is mapped to the weight scores corresponding to each basic reward model. Finally, the weight scores of the multilayer perceptron are normalized to generate the target decision weights of the multilayer perceptron.

[0053] Step S160: Determine the fusion decision result based on the initial scoring matrix and the target decision weight.

[0054] The initial scoring matrix and the target decision weights are weighted and averaged to obtain the fusion score of each candidate decision. The fusion decision result includes each candidate decision and its fusion score.

[0055] Furthermore, existing decision fusion processes are mostly black-box operations, only outputting the final fusion result of multiple decisions and failing to provide relevant decision-making basis. However, for many decision-making scenarios, especially medical diagnosis scenarios, traceability is extremely important. Users need to clearly understand the logical support of AI decisions to determine whether to accept them, and the black-box nature directly limits the trustworthiness and scope of application of AI decisions.

[0056] Therefore, the fusion decision results can also include an arbitration report generated based on the initial scoring matrix and the target decision weights, so as to make the core parameters in the decision fusion process transparent, allow users to clearly understand the logic of result generation, and greatly improve the traceability and trustworthiness of AI decision-making.

[0057] When displaying the fusion decision results, candidate decisions can be sorted and displayed in descending order of fusion score, and the arbitration report can be displayed at the same time.

[0058] The decision fusion method provided in this invention first acquires basic data of the event to be decided and its corresponding multiple candidate decisions. Then, using multiple basic reward models, it scores the multiple candidate decisions based on the basic data to generate an initial scoring matrix. By decomposing the complex decision problem into multiple professional sub-perspective evaluation tasks, each basic reward model performs refined scoring of each evaluation task in its area of ​​expertise to form the initial scoring matrix. This multi-perspective evaluation avoids the bias of a single model. Next, based on the basic data, initial decision weights are generated. These initial decision weights are used to evaluate the importance of each basic reward model in the current event to be decided. By establishing a dynamic mapping between basic data features and model importance, the fundamental problem of traditional decision fusion methods using fixed weights, leading to a disconnect between weight allocation and actual decision requirements, is solved. Finally, based on the basic data, the initial scoring matrix, and the initial decision weights, a meta-state matrix is ​​obtained through fusion. This meta-state matrix breaks through the limitations of traditional single data input, retaining the professional judgments of each basic reward model and capturing the synergistic and divergent relationships among expert opinions, as well as the correlation characteristics between basic data and candidate decisions. It provides a global and comprehensive information foundation for subsequent weight calibration, solving the problem of decision-making bias caused by relying solely on basic data or model scores. Finally, based on the meta-state matrix, the initial decision weights are dynamically calibrated to generate target decision weights. Based on the initial scoring matrix and target decision weights, the fusion decision result is determined. By dynamically calibrating the initial decision weights using global information from the meta-state matrix, the target decision weights can both align with the current decision event and respond to the synergy of expert opinions, completely resolving the adaptability defects of fixed weights and significantly improving the accuracy of the fusion result of multiple decisions (i.e., the fusion decision result).

[0059] Based on any of the above embodiments Figure 2 This is the second flowchart of the decision fusion method provided by the present invention, as shown below. Figure 2 As shown, before step S160, steps S170 and S180 are also included.

[0060] Step S170: Based on the meta-state matrix, perform decision divergence detection and generate divergence detection results.

[0061] Existing multi-decision fusion methods typically lack the ability to handle disagreements. When significant conflicts of opinion arise, existing methods cannot accurately identify the focus of the disagreement, nor can they arbitrate based on high risk, only outputting vague suggestions. Therefore, in this embodiment of the invention, decision disagreement detection is performed based on a meta-state matrix to generate disagreement detection results, while decision calibration vectors for each candidate decision are generated to accurately identify and arbitrate disagreements.

[0062] The process of generating the divergence detection result is as follows: Calculate the similarity between the row vectors in the meta-state matrix, and determine the consensus score based on the similarity; map the meta-state matrix to a mapping score matrix, and calculate the score variance corresponding to multiple candidate decisions based on the mapping score matrix; compare the score variance with a preset threshold, and determine the decision divergence focus based on the comparison result; generate the divergence detection result based on the decision divergence focus and the consensus score. The divergence detection result includes the decision divergence focus, the consensus score, and a global divergence identifier. The specific execution process can be found in the following embodiment, and will not be elaborated here.

[0063] Step S180: Based on the meta-state matrix, generate a decision calibration vector for the plurality of candidate decisions.

[0064] The process of generating decision calibration vectors for multiple candidate decisions is as follows: Sequence modeling of the meta-state matrix is ​​performed using a bidirectional long short-term memory network to obtain sequence-encoded features; these features are then processed using a first multilayer perceptron network to generate decision calibration vectors. The specific execution process can be found in the following embodiment, and will not be elaborated upon here.

[0065] It should be noted that steps S170 and S180 are executed in parallel, and the execution order is not important.

[0066] At this point, step S160 includes step S161.

[0067] Step S161: Determine the fusion decision result based on the initial scoring matrix, the target decision weight, the divergence detection result, and the decision calibration vector.

[0068] The results of the disagreement detection include the focus of the decision disagreement, the consensus score, and the global disagreement indicator.

[0069] In one embodiment, the target decision weights and the initial scoring matrix are weighted and averaged to obtain the initial fusion scores corresponding to multiple candidate decisions; the consensus confidence vector is determined based on the divergence detection results; the initial fusion scores, the consensus confidence vector, and the decision calibration vector are summed to obtain the target fusion scores corresponding to each candidate decision; wherein, the fusion decision results include multiple candidate decisions and their corresponding target fusion scores.

[0070] In another embodiment, when the consensus score is detected to be greater than the preset activation score, the target decision weight and the initial scoring matrix are weighted and averaged to obtain the initial fusion scores corresponding to multiple candidate decisions; the consensus confidence vector is determined based on the divergence detection results; the initial fusion scores, the consensus confidence vector, and the decision calibration vector are summed to obtain the target fusion scores corresponding to each candidate decision; wherein, the fusion decision results include multiple candidate decisions and their corresponding target fusion scores.

[0071] It should be noted that when the global divergence flag is 1, indicating that there is a significant divergence, the decision calibration vector Δ will be adjusted specifically for the global score: a positive offset will be applied to candidate decisions that are high-risk but have low scores, and a negative offset will be applied to candidate decisions that appear to have high scores but have logical contradictions.

[0072] Furthermore, the fusion decision results can also include an arbitration report based on the initial scoring matrix, target decision weight generation, disagreement detection results, and decision calibration vector, to further make the core parameters in the decision fusion process transparent, allowing users to clearly understand the logic of result generation and significantly improving the traceability and trustworthiness of AI decision-making.

[0073] The decision fusion method provided in this invention accurately identifies expert disagreements, obtains disagreement detection results, and generates a decision calibration vector. This solves the problem of decision failure in the prior art when weighted fusion is used in the event of conflicting opinions, and in particular, avoids the risk of high-risk decisions being underestimated due to disagreements.

[0074] Based on any of the above embodiments Figure 3 This is the third flowchart of the decision fusion method provided by the present invention, as shown below. Figure 3 As shown, step S170 includes: step S171, step S172, step S173 and step S174.

[0075] Step S171: Calculate the similarity between each row vector in the meta-state matrix, and determine the consensus score based on the similarity.

[0076] The similarity between the row vectors in the meta-state matrix M is calculated, and a consensus score is determined based on the similarity. The row vectors of the meta-state matrix M represent the opinions of different experts. Since the meta-state matrix M is an n-row matrix, an n×n similarity matrix is ​​formed when calculating the similarity. The mean of the similarity matrix is ​​calculated to obtain the consensus score, denoted as Consensus. The consensus score ranges from 0 to 1; a higher value indicates a greater agreement among the different experts.

[0077] Step S172: Map the meta-state matrix to a mapping score matrix, and calculate the score variance corresponding to the multiple candidate decisions based on the mapping score matrix.

[0078] Step S173: Compare the scoring variance with a preset threshold, and determine the focus of decision disagreement based on the comparison results.

[0079] Simultaneously, the meta-state matrix M (n×512) is mapped back to an n×m dimension using an MLP to obtain the mapped score matrix. For each column of the mapped score matrix (i.e., each candidate decision j), the variance of the n scores is calculated, denoted as the score variance σ_j. 2 When σ_j2 When the threshold is exceeded, it means that multiple experts have given the candidate decision scores that are fluctuating, making it the focus of the debate. Therefore, the candidate decision is marked as the focus of the decision disagreement.

[0080] The preset threshold is determined based on historical baseline data, and is determined as follows: the variance distribution of each candidate decision (e.g., candidate diagnosis) in historical baseline data (e.g., historical cases) is calculated, and a certain percentile (e.g., 90%) is taken as the threshold. Furthermore, some data can be extracted to verify and calibrate the specific percentile to ensure the accuracy of the preset threshold setting.

[0081] Step S174: Generate the disagreement detection result based on the decision disagreement focus and the consensus score.

[0082] Determine whether the number of decision-making disagreements exceeds a preset number and whether the consensus score is less than a preset score. Based on these determinations, a global disagreement flag is set. If the number of decision-making disagreements exceeds the preset number and the consensus score is less than the preset score, the global disagreement flag is set to 1, indicating a significant disagreement. If the number of decision-making disagreements is less than or equal to the preset number, or the consensus score is greater than or equal to the preset score, the global disagreement flag is set to 0, indicating no significant disagreement.

[0083] The preset quantity and preset score can be set according to the actual situation, and no specific limitation is made here. For example, the preset quantity can be set to 1, and the preset score can be set to 0.6. If the number of decision disagreement points is ≥1 and Consensus<0.6, then the global disagreement flag is determined to be 1, indicating that there is a significant disagreement.

[0084] The results of the disagreement detection include the focus of the decision disagreement, the consensus score, and the global disagreement indicator.

[0085] The decision fusion method provided in this invention achieves the quantification and precise positioning of disagreements through the above-mentioned means, providing a basis for the calibration of subsequent fusion decision results and solving the problem of decision failure in traditional weighted fusion when opinions conflict.

[0086] Based on any of the above embodiments, step S180 includes: step S181 and step S182.

[0087] Step S181: Sequence modeling is performed on the meta-state matrix using a bidirectional long short-term memory network to obtain sequence coding features.

[0088] By using a bidirectional long short-term memory (Bi-LSTM) network, the meta-state matrix is ​​sequence-modeled to extract the temporal dependencies between different row vectors and obtain sequence coding features.

[0089] Specifically, the n rows of the meta-state matrix M are treated as a sequence and input into a Bi-LSTM (256-dimensional hidden layer) to perform sequence modeling on the meta-state matrix and capture the temporal dependence of expert opinions. For example, whether a high score from specialist A for diagnosis X is related to a low score from specialist B for diagnosis Y.

[0090] Step S182: The sequence coding features are processed through a first multilayer perceptron network to generate the decision calibration vector.

[0091] The sequence-encoded features are input into a 3-layer multilayer perceptron network (denoted as the first multilayer perceptron network), and the m-dimensional offset vector output by the first multilayer perceptron network is obtained by mapping, denoted as the decision calibration vector Δ.

[0092] It should be noted that when a significant discrepancy is detected, and there may be some low-scoring candidates in the underlying data that are ignored by the basic reward model but are captured as high-risk by the Bi-LSTM, the decision calibration vector will apply a positive offset to the candidate decision, for example, 0.3. For candidate decisions that appear to have high scores but have logical contradictions, a negative offset will be applied, for example, -0.2.

[0093] The decision fusion method provided in this invention uses Bi-LSTM to perform sequence modeling of the meta-state matrix, and then obtains a decision calibration vector through MPL mapping. This vector is used to apply positive and negative offsets to each candidate decision, thereby improving the accuracy and security of the fused decision results. In particular, for medical diagnostic scenarios, it can reduce the risk of missed diagnoses of high-risk diseases, increase vigilance against high-risk diseases, and greatly improve the security and reliability of diagnostic fusion results.

[0094] Based on any of the above embodiments Figure 4 This is the fourth flowchart of the decision fusion method provided by the present invention, as shown below. Figure 4 As shown, step S161 includes: step S1611, step S1612 and step S1613.

[0095] Step S1611: Perform a weighted average of the target decision weights and the initial scoring matrix to obtain the initial fusion score corresponding to the multiple candidate decisions.

[0096] The weighted average of the target decision weights and the initial score matrix is ​​used to obtain the fusion score corresponding to multiple candidate decisions, which is denoted as the initial fusion score.

[0097] Step S1612: Determine the consensus confidence vector based on the divergence detection results.

[0098] The results of the disagreement detection include the focus of the decision disagreement, the consensus score, and the global disagreement indicator.

[0099] The consensus score is expanded to the same dimension m as the candidate decision. At the same time, a mask vector is determined based on the decision disagreement focus. In this mask vector, the mask corresponding to the decision disagreement focus is 0, and the others are 1. The expanded consensus score and the mask vector are multiplied element by element to obtain the consensus confidence vector.

[0100] For example, if the consensus score (Consensus) is 0.9 and there are 3 candidate decisions, then the Consensus is expanded to 3 dimensions, i.e., [0.9, 0.9, 0.9]. If the decision divergence focus indicates that the second candidate decision is the divergence focus, then the mask vector is [1, 0, 1]. The final consensus confidence vector (Consensus_list = [0.9, 0.9, 0.9] ⊙ [1, 0, 1] = [0.9, 0, 0.9].

[0101] This consensus confidence vector is used to improve the credibility of results when experts agree, and to avoid interfering with offset calibration when experts disagree.

[0102] Furthermore, before executing the above implementation method, it is first detected that the consensus score is greater than a preset activation score (e.g., 0.8). If the consensus score is greater than the preset activation score, the consensus confidence vector is determined according to the above implementation method. If the consensus score is less than or equal to the preset activation score, all dimensions of the consensus confidence vector are determined to be 0. In this way, the consensus confidence vector is activated only when opinions are highly consistent, which strengthens the confidence in reliable decision-making and avoids interference with the crucial safety calibration process when there is disagreement, thereby improving the credibility representation of the consensus confidence vector.

[0103] Step S1613: Sum the initial fusion score, the consensus confidence vector, and the decision calibration vector to obtain the target fusion score corresponding to each candidate decision.

[0104] The fusion decision result includes the multiple candidate decisions and their corresponding target fusion scores.

[0105] The initial fusion score, consensus confidence vector, and decision calibration vector are summed to obtain the target fusion score for each candidate decision.

[0106] For example, if the initial fusion score is [0.71, 0.63, 0.26], the consensus confidence vector is [0, 0, 0], and the decision calibration vector is [+0.2, +0.15, -0.3], then the target fusion score is [0.91, 0.78, -0.04]. That is, the target fusion scores corresponding to each candidate decision are 0.91, 0.78, and -0.04, respectively.

[0107] The decision fusion method provided in this invention organically integrates three types of information—weighted consensus (corresponding to the initial fusion score), confidence reward (corresponding to the consensus confidence vector), and risk calibration (corresponding to the decision calibration vector)—in the above manner. This makes the final output target fusion score contain a comprehensive consideration of the system's divergence, consensus, and risk. Its information content and decision value far exceed those of traditional methods, significantly improving the accuracy and interpretability of the fusion decision results.

[0108] Based on any of the above embodiments Figure 5 This is the fifth flowchart of the decision fusion method provided by the present invention, as shown below. Figure 5 As shown, step S140 includes: step S141, step S142 and step S143.

[0109] Step S141: Map the score vectors of each basic reward model in the initial score matrix to feature vectors of a preset dimension to obtain the model score state matrix.

[0110] The rating vectors of each basic reward model in the initial rating matrix S (n×m) (i.e., each row of the initial rating matrix S with dimension m) are mapped to feature vectors of a preset dimension to obtain a model rating state matrix. This model rating state matrix transforms simple ratings into abstract opinion states containing rating patterns, possessing richer evaluation semantics than the initial ratings. The preset dimension can be set to 512 dimensions, consistent with the dimension of the context feature vectors of the basic data.

[0111] Specifically, for each row of the initial rating matrix S (i.e., the m-dimensional rating vector corresponding to each basic reward model), it is mapped to a 512-dimensional feature vector through a 2-layer MLP (m-dimensional input, 512-dimensional output). Then, the feature vectors corresponding to each row are concatenated to form an n×512 model rating state matrix H, so as to realize the dimensionality increase and abstract representation of the rating information.

[0112] Step S142: After expanding the initial decision weights, they are fused with the model scoring state matrix to obtain a fused feature matrix.

[0113] After expanding the initial decision weights, they are fused with the model scoring state matrix to obtain the fused feature matrix.

[0114] Specifically, the initial decision weights G_initial are expanded into an n×512 matrix through a Repeat operation, and then horizontally concatenated with the model scoring state matrix H to obtain a fused feature matrix.

[0115] Step S143: Globally modulate the fused feature matrix with the context feature vector of the basic data, and encode the modulated feature matrix to obtain the meta-state matrix.

[0116] The context feature vector of the basic data is obtained by encoding the basic data through an encoder. Specifically, a BERT encoder can be used to generate a 512-dimensional context feature vector.

[0117] The fused feature matrix and the context feature vector of the basic data are globally modulated, and the modulated feature matrix is ​​encoded to obtain the meta-state matrix.

[0118] Specifically, through a broadcast mechanism, the fused feature matrix is ​​added to the 512-dimensional context feature vector to achieve global adjustment, thereby enabling the basic data (such as medical record information) to globally modulate the fused features of the model score and initial decision weights. Next, a four-layer encoder encodes the modulated feature matrix, outputting an n×512 matrix, denoted as the meta-state matrix M. This meta-state matrix M fully preserves the three pieces of information: model score, basic data background, and initial decision weights.

[0119] The decision fusion method provided in this invention integrates model scores, initial decision weights, and basic data to construct a meta-state matrix containing multi-dimensional information. This addresses the decision bias caused by the one-sidedness of a single information source and provides a comprehensive information foundation for subsequent weight optimization and disagreement arbitration. Specifically, in medical diagnostic scenarios, the meta-state matrix not only retains the professional judgments of each specialty model but also captures the synergistic and divergent relationships between expert opinions and the correlation features between medical records and diagnoses. Furthermore, it integrates initial decision weights, thus providing a global and three-dimensional information foundation for subsequent weight optimization and disagreement arbitration, solving the decision bias problem caused by relying solely on medical records or expert scores.

[0120] Based on any of the above embodiments Figure 6 This is the sixth flowchart of the decision fusion method provided by the present invention, as shown below. Figure 6 As shown, step S150 includes: step S151, step S152 and step S153.

[0121] Step S151: Perform global average pooling on the meta-state matrix along the dimension of the basic reward model to obtain the global meta-state vector.

[0122] Global average pooling is performed on the meta-state matrix along the dimension of the basic reward model (i.e., the decision dimension) to obtain the global meta-state vector.

[0123] Specifically, the n×512 meta-state matrix M is subjected to global average pooling along the decision dimension (i.e., across the n rows). That is, the n 512-dimensional row vectors are averaged and pooled into a single 512-dimensional global meta-state vector, denoted as m_global. This global meta-state vector represents the consensus feature formed after discussion of all basic reward models.

[0124] The above operation aggregates the opinion features of all basic reward models, filters out noise interference from a single basic reward model, and highlights the trend of group opinions.

[0125] Step S152: The global meta-state vector is mapped to the weight scores corresponding to each basic reward model through the second multilayer perceptron network.

[0126] Step S153: Normalize the weight scores to generate the target decision weights.

[0127] The global meta-state vector is input into a multilayer perceptron network (denoted as the second multilayer perceptron network) to map and obtain the weight scores corresponding to each basic reward model. The second multilayer perceptron network can have two layers, and its output dimension is the number of basic reward models, n.

[0128] Then, the weight scores corresponding to each basic reward model are normalized by the Softmax function, and a weight distribution is output, which is the target decision weight.

[0129] For example, in a medical diagnostic scenario, when multiple specialty models reach a consensus on a diagnosis, the meta-state matrix exhibits synergistic characteristics. The target decision weights m_global, after global average pooling, will strengthen the weights of these specialty models. Conversely, if a specialty model's score contradicts that of the majority of specialty models, its weight will be suppressed. For instance, in the initial decision weights, the neurology model's weight is 0.5. The meta-state matrix shows a high degree of synergy between the neurology and immunology models' scores, and the neurology model's final weight may increase to 0.55.

[0130] The decision fusion method provided in this invention solves the problem that the initial decision weights only rely on basic data and do not consider the synergy of the basic reward models by adjusting the weights of each basic reward model according to the global information in the meta-state matrix. This makes the weight allocation more in line with the actual decision-making needs and thus improves the accuracy of the fusion decision results.

[0131] Furthermore, based on any of the above embodiments, the decision fusion method provided in this invention is applied to a decision fusion system. This system adopts a layered, progressive architecture, consisting of four layers: an input layer, a multi-basic reward model layer, an evaluation routing network layer, a meta-reward model layer, and an output layer. The input layer receives basic data of the event to be decided and multiple candidate decisions. The multi-basic reward model layer generates an initial scoring matrix, the evaluation routing network layer provides initial decision weights, the meta-reward model layer achieves intelligent decision fusion, and finally, the output layer presents the fused decision result. This architecture design retains the professional judgment advantages of each basic reward model while solving the core challenge of multi-expert opinion collaboration and fusion through a meta-arbitration mechanism, achieving the dual value of "professional depth + decision wisdom."

[0132] The overall process is as follows: After receiving the basic data of the event to be decided and its corresponding multiple candidate decisions, the input layer synchronously transmits the data to the multi-basic reward model layer and the evaluation routing network layer. The multi-basic reward model layer generates an initial scoring matrix for each candidate decision, while the evaluation routing network layer generates initial decision weights based on the basic data. After these two types of information enter the meta-reward model layer, they are first fused into a meta-state matrix, then the expert weights are optimized by a gated weight generator, and the discrepancies are handled by a decision divergence detection and decision offset generator. Finally, the fused decision result is calculated and fed back to the user by the output layer.

[0133] It should be noted that the core meta-arbitration logic of this invention (including the fusion of meta-state matrices, dynamic weight optimization, divergence detection, and decision calibration) does not depend on the model type and data characteristics of a specific domain. For different application scenarios, only the input data format and the corresponding basic reward model need to be adjusted to quickly adapt to the application, which greatly reduces the development cost of cross-domain applications and expands the scope of application of the technology.

[0134] Furthermore, taking the diagnosis of rare diseases in a 12-year-old child as an example, the complete operation process of the decision fusion system is demonstrated, namely the decision fusion method of this invention.

[0135] (1) The input medical record is "headache for 3 weeks, accompanied by orbital pain and weakness in the limbs, symptoms recurred after hormone treatment", and the candidate diagnoses include 3, namely: multiple sclerosis, neuromyelitis optica, and antimyelin antibody-related diseases; (2) Generate the initial scoring matrix S: The neurology model scores the three candidate diagnoses as [0.6, 0.3, 0.1], the immunology model scores the three candidate diagnoses as [0.4, 0.4, 0.2], and the cardiology model scores the three candidate diagnoses as [0.2, 0.2, 0.6]. (3) Based on the characteristics of neurological symptoms and recurrence, the initial decision weights G_initial=[0.5,0.4,0.1] are generated; (4) Calculate the meta-state matrix M to capture the synergy between the neurological model and the immunology model in multiple sclerosis scoring. For example, in rare neurological disease cases, the meta-state matrix can simultaneously integrate the scoring differences between the neurological model and the immunology model, the case feature of "relapse after hormone therapy", and the low importance marker of the internal medicine model in the initial decision weight center, providing complete information support for accurate decision-making.

[0136] (4) Based on the meta-state matrix M, perform weight calibration and output the target decision weight G_final=[0.55,0.4,0.05], where the weight of the cardiology model is reduced due to opinion isolation; (5) Decision divergence detection was performed based on the meta-state matrix M, and the score variance σ of neuromyelitis optica as a candidate diagnosis was detected. 2 =0.06 (>threshold 0.05), consensus score Consensus=0.18, global divergence flag=1, and at the same time, a decision calibration vector Δ=[0.2,-0.1,0] is generated based on the meta-state matrix M, thus determining Consensus_list=[0.18,0,0.18]. (6) Calculate the target fusion score for each diagnose. The target fusion score for multiple sclerosis is (0.55×0.6+0.4×0.3+0.05×0.1)+0.2+0.18=0.835, the target fusion score for neuromyelitis optica is (0.55×0.4+0.4×0.4+0.05×0.2)-0.1+0=0.29, and the target fusion score for antimyelin antibody-related diseases is (0.55×0.2+0.4×0.2+0.05×0.6)+0+0.18=0.4. The final ranking is [multiple sclerosis, antimyelin antibody-related diseases, neuromyelitis optica], which is consistent with the clinical diagnosis. The arbitration report clearly explains the weight adjustment and offset.

[0137] The decision fusion apparatus provided in the embodiments of the present invention is described below. The decision fusion apparatus described below and the decision fusion method described above can be referred to and correspond to each other.

[0138] Figure 7 This is a schematic diagram of the decision fusion device provided by the present invention, as shown below. Figure 7 As shown, the device includes a data acquisition module 710, a first generation module 720, a second generation module 730, a data fusion module 740, a weight calibration module 750, and a result determination module 760; wherein: The data acquisition module 710 is used to acquire the basic data of the event to be decided and its corresponding multiple candidate decisions; The first generation module 720 is used to score the multiple candidate decisions based on the basic data using multiple basic reward models to generate an initial scoring matrix. The second generation module 730 is used to generate initial decision weights based on the basic data; The data fusion module 740 is used to fuse the basic data, the initial scoring matrix and the initial decision weights to obtain a meta-state matrix. The weight calibration module 750 is used to calibrate the initial decision weights based on the meta-state matrix to generate target decision weights. The result determination module 760 is used to determine the fusion decision result based on the initial scoring matrix and the target decision weight.

[0139] The decision fusion apparatus provided in this invention first acquires basic data of the event to be decided and its corresponding multiple candidate decisions. Then, it uses multiple basic reward models to score the multiple candidate decisions based on the basic data, generating an initial scoring matrix. By decomposing the complex decision problem into multiple professional sub-perspective evaluation tasks, each basic reward model performs refined scoring of each evaluation task in its area of ​​expertise to form the initial scoring matrix. This multi-perspective evaluation avoids the bias of a single model. Next, based on the basic data, initial decision weights are generated. These initial decision weights are used to evaluate the importance of each basic reward model in the current event to be decided. By establishing a dynamic mapping between basic data features and model importance, the fundamental problem of traditional decision fusion methods using fixed weights, leading to a disconnect between weight allocation and actual decision requirements, is solved. Finally, based on the basic data, the initial scoring matrix, and the initial decision weights, a meta-state matrix is ​​obtained through fusion. This meta-state matrix breaks through the limitations of traditional single data input, retaining the professional judgments of each basic reward model and capturing the synergistic and divergent relationships among expert opinions, as well as the correlation characteristics between basic data and candidate decisions. It provides a global and comprehensive information foundation for subsequent weight calibration, solving the problem of decision-making bias caused by relying solely on basic data or model scores. Finally, based on the meta-state matrix, the initial decision weights are dynamically calibrated to generate target decision weights. Based on the initial scoring matrix and target decision weights, the fusion decision result is determined. By dynamically calibrating the initial decision weights using global information from the meta-state matrix, the target decision weights can both align with the current decision event and respond to the synergy of expert opinions, completely resolving the adaptability defects of fixed weights and significantly improving the accuracy of the fusion decision result.

[0140] According to a decision fusion device provided by the present invention, the decision fusion device further includes: The divergence detection module is used to perform decision divergence detection based on the meta-state matrix and generate divergence detection results. The calibration vector generation module is used to generate decision calibration vectors for the multiple candidate decisions based on the meta-state matrix. The result determination module 760 is specifically used for: Based on the initial scoring matrix, the target decision weights, the divergence detection results, and the decision calibration vector, the fusion decision result is determined.

[0141] According to a decision fusion device provided by the present invention, the divergence detection module is specifically used for: Calculate the similarity between each row vector in the meta-state matrix, and determine the consensus score based on the similarity; The meta-state matrix is ​​mapped to a mapping score matrix, and the score variance corresponding to the multiple candidate decisions is calculated based on the mapping score matrix. The score variance is compared with a preset threshold, and the focus of decision disagreement is determined based on the comparison results; The disagreement detection result is generated based on the decision disagreement focus and the consensus score.

[0142] According to a decision fusion apparatus provided by the present invention, the calibration vector generation module is specifically used for: Sequence encoding features are obtained by using a bidirectional long short-term memory network to perform sequence modeling on the meta-state matrix. The sequence encoding features are processed through a first multilayer perceptron network to generate the decision calibration vector.

[0143] According to a decision fusion apparatus provided by the present invention, the result determination module 760 is further specifically used for: The initial fusion score corresponding to the multiple candidate decisions is obtained by weighting the target decision weights and the initial score matrix. Based on the divergence detection results, a consensus confidence vector is determined; The initial fusion score, the consensus confidence vector, and the decision calibration vector are summed to obtain the target fusion score corresponding to each candidate decision; The fusion decision result includes the multiple candidate decisions and their corresponding target fusion scores.

[0144] According to a decision fusion apparatus provided by the present invention, the data fusion module 740 is specifically used for: The scoring vectors of each basic reward model in the initial scoring matrix are mapped to feature vectors of a preset dimension to obtain the model scoring state matrix; After expanding the initial decision weights, they are fused with the model scoring state matrix to obtain a fused feature matrix; The fused feature matrix is ​​globally modulated with the context feature vector of the basic data, and the modulated feature matrix is ​​encoded to obtain the meta-state matrix.

[0145] According to a decision fusion device provided by the present invention, the weight calibration module 750 is specifically used for: Global average pooling is performed on the meta-state matrix along the dimension of the basic reward model to obtain the global meta-state vector; The global meta-state vector is mapped to the weight scores corresponding to each basic reward model through a second multilayer perceptron network. The weight scores are normalized to generate the target decision weights.

[0146] It should be noted that the decision fusion device provided in this embodiment of the invention can implement all the method steps implemented in the above-mentioned decision fusion method embodiment and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0147] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the decision fusion method provided in the above embodiments.

[0148] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0149] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the decision fusion methods provided in the above embodiments.

[0150] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0151] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A decision fusion method, characterized in that, include: Acquire the basic data of the event to be decided and its corresponding multiple candidate decisions; Using multiple basic reward models, the multiple candidate decisions are scored based on the basic data to generate an initial score matrix; Based on the aforementioned basic data, initial decision weights are generated; Based on the aforementioned basic data, the initial scoring matrix, and the initial decision weights, a meta-state matrix is ​​obtained through fusion. Based on the meta-state matrix, the initial decision weights are calibrated to generate target decision weights; Based on the initial scoring matrix and the target decision weights, the fusion decision result is determined.

2. The decision fusion method according to claim 1, characterized in that, Before determining the fusion decision result based on the initial scoring matrix and the target decision weights, the process also includes: Based on the aforementioned meta-state matrix, decision divergence detection is performed, and divergence detection results are generated. Based on the meta-state matrix, a decision calibration vector is generated for the plurality of candidate decisions; The process of determining the fusion decision result based on the initial scoring matrix and the target decision weights includes: Based on the initial scoring matrix, the target decision weights, the divergence detection results, and the decision calibration vector, the fusion decision result is determined.

3. The decision fusion method according to claim 2, characterized in that, The step of performing decision divergence detection based on the meta-state matrix and generating divergence detection results includes: Calculate the similarity between each row vector in the meta-state matrix, and determine the consensus score based on the similarity; The meta-state matrix is ​​mapped to a mapping score matrix, and the score variance corresponding to the multiple candidate decisions is calculated based on the mapping score matrix. The score variance is compared with a preset threshold, and the focus of decision disagreement is determined based on the comparison results; The disagreement detection result is generated based on the decision disagreement focus and the consensus score.

4. The decision fusion method according to claim 2, characterized in that, The step of generating a decision calibration vector for the plurality of candidate decisions based on the meta-state matrix includes: Sequence encoding features are obtained by using a bidirectional long short-term memory network to perform sequence modeling on the meta-state matrix. The sequence encoding features are processed through a first multilayer perceptron network to generate the decision calibration vector.

5. The decision fusion method according to claim 2, characterized in that, The process of determining the fusion decision result based on the initial scoring matrix, the target decision weight, the divergence detection result, and the decision calibration vector includes: The initial fusion score corresponding to the multiple candidate decisions is obtained by weighting the target decision weights and the initial score matrix. Based on the divergence detection results, a consensus confidence vector is determined; The initial fusion score, the consensus confidence vector, and the decision calibration vector are summed to obtain the target fusion score corresponding to each candidate decision; The fusion decision result includes the multiple candidate decisions and their corresponding target fusion scores.

6. The decision fusion method according to any one of claims 1 to 5, characterized in that, The meta-state matrix, obtained by fusing the basic data, the initial scoring matrix, and the initial decision weights, includes: The scoring vectors of each basic reward model in the initial scoring matrix are mapped to feature vectors of a preset dimension to obtain the model scoring state matrix; After expanding the initial decision weights, they are fused with the model scoring state matrix to obtain a fused feature matrix; The fused feature matrix is ​​globally modulated with the context feature vector of the basic data, and the modulated feature matrix is ​​encoded to obtain the meta-state matrix.

7. The decision fusion method according to any one of claims 1 to 5, characterized in that, The step of calibrating the initial decision weights based on the meta-state matrix to generate target decision weights includes: Global average pooling is performed on the meta-state matrix along the dimension of the basic reward model to obtain the global meta-state vector; The global meta-state vector is mapped to the weight scores corresponding to each basic reward model through a second multilayer perceptron network. The weight scores are normalized to generate the target decision weights.

8. A decision fusion device, characterized in that, include: The data acquisition module is used to acquire the basic data of the event to be decided and its corresponding multiple candidate decisions; The first generation module is used to score the multiple candidate decisions based on the basic data using multiple basic reward models to generate an initial scoring matrix. The second generation module is used to generate initial decision weights based on the basic data; The data fusion module is used to fuse the basic data, the initial scoring matrix, and the initial decision weights to obtain a meta-state matrix; The weight calibration module is used to calibrate the initial decision weights based on the meta-state matrix to generate target decision weights. The result determination module is used to determine the fusion decision result based on the initial scoring matrix and the target decision weights.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the decision fusion method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the decision fusion method as described in any one of claims 1 to 7.

Images