A knowledge data double-drive industrial hazard identification method, system, device and medium

Abstract

This invention discloses a knowledge-data dual-driven method, system, equipment, and medium for identifying industrial hazards, relating to the field of intelligent identification technology for high-risk industrial hazards. The method includes mapping knowledge embedding vectors and multimodal feature vectors from the industrial site to a unified semantic space to form unified semantic features; performing multimodal attention fusion based on knowledge guidance, combining modal baseline weights and real-time data quality to form a fused feature vector; determining the combination and weights of reasoning logic based on scene characteristics and performing collaborative reasoning; fusing the reasoning results and calibrating them with knowledge confidence probabilities to obtain the identification result of high-risk industrial hazards. The method of this invention achieves knowledge data representation through unified semantic mapping, dynamically fuses modal weights and data quality, determines reasoning logic and weights based on scene characteristics, and calibrates with knowledge confidence probabilities, thereby improving the accuracy, adaptability, and reliability of hazard identification.

Description

Technical Field

[0001] This invention relates to the field of intelligent identification technology for high-risk industrial hazards, specifically to a knowledge- and data-driven method, system, equipment, and medium for identifying industrial hazards. Background Technology

[0002] As high-risk industrial production systems continue to evolve towards large-scale, complex, and continuous operations, the coupling between production equipment and process links is constantly increasing. Equipment hazards exhibit typical characteristics of strong concealment, multiple transmission paths, and rapid evolution. Traditional safety control models relying on manual inspections and single-point threshold alarms not only suffer from incomplete inspection coverage and delayed hazard detection, but also struggle to cope with complex scenarios involving multiple coupled hazards and gradual evolution. They can no longer meet the core needs of precise supervision and proactive prevention in modern high-risk industries.

[0003] In existing technologies, AI-based hazard identification solutions mainly fall into two technical routes: knowledge-driven and data-driven. The knowledge-driven route, centered on expert rules and traditional knowledge graphs, can output compliant and interpretable judgment results. However, it is limited by the semantic heterogeneity and logical implicitness of hazard standard texts, making it difficult to directly convert a large amount of unstructured specifications into computable structured knowledge. Meanwhile, traditional knowledge graphs generally adopt a deterministic "entity-relationship-entity" triplet structure, lacking probabilistic representation of uncertain knowledge and failing to achieve semantic-level coupling with multimodal perception data. This results in a long-standing disconnect between knowledge reasoning and data recognition, and the guiding and constraining role of the knowledge base on the identification model has not been fully realized.

[0004] While data-driven approaches can achieve high-precision hazard identification, existing multimodal fusion methods typically rely on fixed weights, making it difficult to adaptively adjust to changes in operating conditions and data quality fluctuations. This leads to feature distortion and identification bias in complex environments with strong noise and missing data. Furthermore, existing solutions often employ single-logic reasoning, limiting the adaptability of rule-based reasoning and lacking mechanistic constraints in data reasoning. This makes them prone to misjudgments and omissions in scenarios involving multiple coupled hazards and gradual evolution. Moreover, the reasoning process lacks effective connections with causal relationships and regulatory basis, failing to meet the requirements of traceability, verifiability, and interpretability for high-risk industrial safety supervision. Existing knowledge-driven and data-driven solutions typically only fuse results at the output stage. Knowledge graphs primarily serve as post-verification tools for reasoning results and have not effectively participated in multimodal feature fusion and reasoning scheduling. Attribute information in tuple knowledge graphs also fails to fully participate in algorithm computation, hindering deep collaboration between prior knowledge and on-site perception data. This restricts further improvements in the accuracy and reliability of hazard identification in complex high-risk industrial scenarios. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is that existing intelligent identification methods for high-risk industrial hazards have difficulties in deeply coordinating knowledge mechanisms and multimodal perception data, the fusion process is difficult to adapt to changes in working conditions and data quality, the reasoning strategy lacks scenario adaptability, and there are problems in how to achieve multimodal dynamic fusion and multi-logic collaborative reasoning under knowledge constraints.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a knowledge-data dual-drive method for identifying industrial hazards, comprising mapping the knowledge embedding vector corresponding to a knowledge graph of high-risk industrial hazards and the multimodal feature vector of the industrial site to a unified semantic space, forming a unified semantic feature corresponding to the knowledge semantics and the multimodal data semantics; performing knowledge-guided multimodal attention fusion based on the unified semantic feature, and determining the fusion weight of each modal data feature by combining the modal baseline weight in the knowledge graph of high-risk industrial hazards and the real-time data quality, forming a fusion feature vector under knowledge constraints; determining the corresponding inference logic combination and the weight of each inference logic according to the scene features corresponding to the fusion feature vector, and performing collaborative inference on the fusion feature vector according to the inference logic combination; forming a fusion inference result based on the inference result of the inference logic combination and the weight of each inference logic, and calibrating the fusion inference result based on the knowledge confidence probability in the knowledge graph of high-risk industrial hazards to obtain the identification result of high-risk industrial hazards.

[0008] As a preferred embodiment of the knowledge-data dual-drive industrial hazard identification method of the present invention, the step of mapping the knowledge embedding vector corresponding to the knowledge graph of high-risk industrial hazards and the multimodal feature vector of the industrial site to a unified semantic space includes projecting the knowledge embedding vector and the multimodal feature vector of the industrial site through a knowledge-side projection matrix and a data-side projection matrix, respectively, constructing a positive sample pair between the knowledge embedding vector and the multimodal feature vector of the industrial site corresponding to the same hazard type, and constructing a negative sample pair between the knowledge embedding vector and the multimodal feature vector of the industrial site corresponding to different hazard types.

[0009] The similarity of the positive and negative sample pairs in the unified semantic space is constrained by the bilateral contrast loss, and the knowledge-side projection matrix and data-side projection matrix are adjusted according to the constraint results so that the knowledge semantics of the same hidden danger type corresponds with the multimodal data semantics in the unified semantic space.

[0010] As a preferred embodiment of the knowledge-data dual-drive industrial hazard identification method of the present invention, the knowledge-guided multimodal attention fusion includes determining the initial parameters of the projection matrix corresponding to the multi-head attention based on the modal benchmark weights in the high-risk industrial hazard knowledge graph, and forming a feature sequence by combining the knowledge features in the unified semantic space with the data features of each modality.

[0011] Joint self-attention calculation is performed on the feature sequence, and the information contribution of each modal data feature is adjusted according to the degree of correlation between the knowledge feature and the modal data feature to form each modal enhanced feature after knowledge-guided interaction.

[0012] As a preferred embodiment of the knowledge-data dual-drive industrial hazard identification method described in this invention, the determination of the fusion weights of each modal data feature includes forming a corresponding data quality score based on the clarity, signal-to-noise ratio, and data completeness of each modal data.

[0013] Based on the candidate results of the hidden danger category, the modal baseline weights in the knowledge graph of high-risk industrial hidden dangers are corrected. The corrected modal baseline weights are then normalized and calculated with the data quality score to obtain the fusion weights of each modal data feature.

[0014] A lower limit constraint is set for the fusion weight of each modality data feature, and the enhanced features of each modality are weighted and aggregated according to the fusion weight to form the fusion feature vector.

[0015] As a preferred embodiment of the knowledge-data dual-drive industrial hazard identification method of the present invention, the step of determining the corresponding combination of reasoning logic and the weight of each reasoning logic includes taking the scene features as the state of reinforcement learning, and taking the combination and corresponding weights of symbolic rule reasoning, statistical law reasoning, causal relationship reasoning and equipment mechanism reasoning as the action of reinforcement learning.

[0016] The reasoning logic combination and the weight of each reasoning logic are determined according to the reasoning scheduling strategy corresponding to the current state. The hidden danger reasoning is executed according to the determined reasoning logic combination, and the reward value is generated based on the accuracy of hidden danger identification and reasoning time.

[0017] The inference scheduling strategy is iteratively updated based on the reward value, so that subsequent scene features determine the corresponding inference logic combinations and the weights of each inference logic according to the updated inference scheduling strategy.

[0018] As a preferred embodiment of the knowledge data dual-drive industrial hazard identification method of the present invention, the step of forming a fusion reasoning result based on the reasoning result of the combination of reasoning logics and the weight of each reasoning logic includes weighting each reasoning result according to the weight corresponding to each reasoning logic to form an initial fusion reasoning result.

[0019] The knowledge confidence probability of the corresponding causal rules in the knowledge graph of high-risk industrial hazards is used as the calibration coefficient to perform knowledge confidence calibration on the initial fusion reasoning results.

[0020] Cross-validation is performed on the judgment results of different reasoning logics. When both symbolic rule reasoning and equipment mechanism reasoning are judged to be normal and statistical regularity reasoning is judged to be abnormal, the corresponding abnormal results are filtered out. The cross-validation results are used to determine whether the hidden danger is established. When the hidden danger is determined to be established, a high-risk industrial hidden danger identification result is formed.

[0021] As a preferred embodiment of the knowledge data dual-drive industrial hazard identification method described in this invention, the method involves: interpolating and completing missing features in multimodal data, and identifying and removing abnormal outlier features.

[0022] When a single modality fails, the fusion weight of the corresponding modality is adjusted to zero. The fusion feature vector is re-determined based on the enhanced features corresponding to the remaining effective modalities and the prior knowledge provided by the knowledge graph of high-risk industrial hazards. Hazard inference is then performed based on the re-determined fusion feature vector.

[0023] Another objective of this invention is to provide a knowledge-data dual-drive industrial hazard identification system, which can achieve collaborative processing of knowledge information and multimodal data through knowledge data semantic alignment, knowledge-guided multimodal fusion, scenario-adaptive collaborative reasoning, and reasoning fusion calibration. This solves the problems of insufficient knowledge and data fusion and difficulty in adapting reasoning methods to different scenarios in current industrial hazard identification technologies.

[0024] As a preferred embodiment of the knowledge-data dual-drive industrial hazard identification system described in this invention, it includes: a knowledge-data semantic alignment module, a knowledge-guided multimodal fusion module, a scenario-adaptive collaborative reasoning module, and a reasoning fusion calibration module.

[0025] The knowledge data semantic alignment module is used to map the knowledge embedding vector and the industrial field multimodal feature vector to a unified semantic space to form corresponding unified semantic features.

[0026] The knowledge-guided multimodal fusion module is used for knowledge-guided multimodal fusion based on unified semantic features, and combines modality benchmark weights and real-time data quality to form a fusion feature vector.

[0027] The scene-adaptive collaborative reasoning module is used to determine the combination of reasoning logic and the weight of each reasoning logic based on the scene features corresponding to the fused feature vector, and to perform collaborative reasoning.

[0028] The reasoning fusion calibration module is used to form a fusion reasoning result based on the results of each reasoning logic and the weight of each reasoning logic, and to perform calibration in combination with the knowledge confidence probability to obtain the identification result of high-risk industrial hazards.

[0029] Another object of the present invention is to provide a knowledge-data dual-drive industrial hazard identification device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program as a step to implement the knowledge-data dual-drive industrial hazard identification method.

[0030] Another object of the present invention is to provide a knowledge data dual-drive industrial hazard identification storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the knowledge data dual-drive industrial hazard identification method.

[0031] The beneficial effects of this invention are as follows: The knowledge-data dual-drive industrial hazard identification method provided by this invention maps the knowledge embedding vector of the knowledge graph of high-risk industrial hazards to the multimodal features of the industrial site into a unified semantic space, thereby achieving collaborative representation of knowledge semantics and multimodal data semantics; through knowledge-guided multimodal attention fusion, and by dynamically determining the fusion weight in combination with modal baseline weights and real-time data quality, the fusion result can adapt to different working conditions and data quality changes; by adaptively determining the combination of reasoning logics and their weights based on scene characteristics, different reasoning logics can collaboratively participate in hazard judgment for the current industrial scene; by fusing the reasoning results and calibrating them in combination with knowledge confidence probability, the impact of misjudgment caused by a single reasoning logic is reduced, thereby improving the accuracy, adaptability, and reliability of hazard identification in complex high-risk industrial scenes. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is an overall flowchart of the knowledge data-driven industrial hazard identification method of the present invention.

[0034] Figure 2 This is a processing sequence diagram of a knowledge data-driven industrial hazard identification method according to the present invention.

[0035] Figure 3 This is a schematic diagram illustrating the use of knowledge graph attributes in the calculation of a dual-drive knowledge data method for identifying industrial hazards according to the present invention.

[0036] Figure 4 This is a schematic diagram illustrating the scenario-adaptive collaborative reasoning and result fusion of the knowledge-data dual-drive industrial hazard identification method of the present invention.

[0037] In the figure: 101, knowledge embedding vector; 102, multimodal feature vector of industrial site; 103, unified semantic feature; 201, fusion feature vector; 301, combination of reasoning logic; 302, weight of each reasoning logic; 401, fusion reasoning result; 402, high-risk industrial hazard identification result. Detailed Implementation

[0038] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0039] To facilitate the explanation of the specific implementation process of the knowledge-data dual-drive industrial hazard identification method of this invention, the following embodiments use a reaction vessel in a chemical workshop as a typical industrial object for illustration. During the operation of the reaction vessel, on-site information such as temperature, pressure, vibration, acoustic signature, and equipment appearance status is involved, and industrial hazards such as overheating and sealing leaks exist. Therefore, it can be used to illustrate the semantic alignment, multimodal feature fusion, scene-adaptive collaborative reasoning, and reasoning result calibration process between the knowledge graph of high-risk industrial hazards and multimodal data from the industrial site. The reaction vessel operating status, hazard types, and related data mentioned below are all used to illustrate the specific processing process of this invention and do not constitute a limitation on the application objects of this invention.

[0040] Example 1

[0041] Reference Figures 1-2 This first embodiment of the invention provides a knowledge-data dual-driven method for identifying industrial hazards. This embodiment converts knowledge of high-risk industrial hazards and multimodal data from the industrial site into corresponding semantic expressions. Multimodal feature fusion is achieved based on knowledge information constraints. The inference logic combination 301 and the contribution ratio of different inference logics are adjusted according to the current industrial scenario. The fused inference result 401 is then calibrated using the knowledge confidence probability in the knowledge graph. This addresses the problems of the disconnect between knowledge mechanisms and on-site multimodal perception data, the difficulty in adapting the fusion process to changes in data state, and the difficulty in adjusting the inference method according to changes in the industrial scenario.

[0042] like Figure 1As shown in the figure, the knowledge embedding vector 101 and the industrial site multimodal feature vector 102 are used as two types of inputs. The knowledge embedding vector 101 represents the industrial mechanism knowledge in the knowledge graph of high-risk industrial hazards, while the industrial site multimodal feature vector 102 represents on-site operating information such as reactor temperature, pressure, vibration, sound signature, and equipment appearance. The two are mapped to form a unified semantic feature 103, establishing a correspondence between knowledge semantics and on-site data semantics. Based on this, the information contribution of different modes is adjusted according to the modal baseline weights in the knowledge graph and the real-time data quality of each mode, forming a fusion feature vector 201. The fusion feature vector 201 is further used to represent the current industrial scenario, and based on this, the reasoning logic combination 301 and the weights 302 of each reasoning logic are determined, enabling symbolic rule reasoning, statistical law reasoning, causal relationship reasoning, and equipment mechanism reasoning to be analyzed collaboratively according to the current working conditions. The reasoning results are fused according to the weights 302 of each reasoning logic to obtain the fusion reasoning result 401, which is then calibrated using the knowledge confidence probability in the knowledge graph, ultimately forming the high-risk industrial hazard identification result 402. thus, Figure 1 This demonstrates that knowledge and information are not only verified after inference is completed, but also continuously participate in the complete processing chain of semantic alignment, multimodal fusion, inference scheduling, and result calibration.

[0043] The knowledge embedding vector 101 corresponding to the knowledge graph of high-risk industrial hazards and the multimodal feature vector 102 of the industrial site are mapped to a unified semantic space to form a unified semantic feature 103 corresponding to the knowledge semantics and the multimodal data semantics.

[0044] Specifically, the high-risk industrial hazard knowledge graph is used to represent the relationships between industrial equipment, operating status, hazard types, and various types of industrial safety knowledge. After vectorization, the knowledge graph transforms the knowledge originally stored in the form of entities and relationships into knowledge embedding vectors 101 that can participate in computation.

[0045] The industrial site multimodal feature vector 102 is used to characterize the on-site operating status reflected by different sensing sources. Different modes can describe the current industrial object from aspects such as equipment appearance, operating parameters, sound status, vibration status and spatial status, so that the on-site operating status can be characterized by different data features.

[0046] The knowledge embedding vector 101 originates from the industrial knowledge system, while the industrial site multimodal feature vector 102 originates from industrial site perception data. Since their original representations differ, directly combining them makes it difficult to determine the semantic correspondence between knowledge information and site data. Therefore, mapping the knowledge embedding vector 101 and the industrial site multimodal feature vector 102 to the same semantic space provides a common semantic foundation for both knowledge information and site data.

[0047] After mapping is completed, the knowledge semantics corresponding to the same industrial hazard can be correlated with the data semantics reflecting the state of the hazard in the multimodal data of the field. For example, the knowledge semantics reflecting the abnormal state of equipment in the knowledge graph can be associated with the multimodal features reflecting the corresponding abnormal state in the industrial field, thereby forming a unified semantic feature 103.

[0048] The unified semantic feature 103 retains the semantic information corresponding to industrial hazard knowledge and also includes the actual operating status reflected by multimodal data from the industrial site. This provides a common semantic basis for subsequent knowledge information to directly participate in multimodal feature fusion, avoiding the fact that knowledge information and data features are always in mutually independent processing links.

[0049] Multimodal attention fusion guided by knowledge is performed based on unified semantic features 103, and the fusion weights of each modality data feature are determined by combining the modality benchmark weights in the knowledge graph and the real-time data quality, forming a fusion feature vector 201 under knowledge constraints.

[0050] Specifically, after forming a unified semantic feature 103, knowledge semantics is used as a priori guide in the multimodal feature fusion process, so that the information interaction between different modal data is not only based on the feature relationship of the data itself, but also constrained by industrial hazard knowledge.

[0051] Knowledge-guided multimodal attention fusion is used to analyze the correlation between knowledge semantics and the semantics of data from different modalities, and to adjust the information contribution of different modal features in the fusion process according to the degree of correlation. Modal features with a high degree of correlation with the current hidden danger mechanism receive greater attention, while modal features with a low degree of correlation with the current hidden danger receive a correspondingly reduced contribution, thereby making the fused features more concentrated in reflecting the current state of industrial hidden dangers.

[0052] Modal baseline weights in a knowledge graph are used to reflect the prior importance of different modalities for corresponding industrial hazards. For different types of industrial hazards, the importance of various perceptual information is not entirely the same; therefore, modal baseline weights in a knowledge graph are used to establish prior contribution relationships between different modalities.

[0053] Real-time data quality reflects the actual validity of data from various modalities in the current industrial environment. Industrial environments can be affected by factors such as noise, changes in equipment operation, and changes in sensor status; therefore, the validity of data from the same modality may vary at different times. Using real-time data quality and modal baseline weights together to determine fusion weights allows for adjusting the contribution ratios of different modalities based on the actual data status in the field while preserving prior knowledge.

[0054] When a certain type of data is strongly correlated with the current potential hazard and has high data quality, the contribution of the corresponding modality in the fusion result increases; when the quality of a certain type of data decreases, its fusion contribution is adjusted accordingly. This avoids using fixed modality weights for different industrial scenarios.

[0055] After knowledge-guided attention interaction and fusion weight adjustment, the data from each modality form a fusion feature vector 201 under knowledge constraints. The fusion feature vector 201 comprehensively includes prior information provided by industrial knowledge and real-time status information provided by different modal data on site, serving as a unified feature basis for subsequent collaborative reasoning about potential hazards.

[0056] Based on the scene features corresponding to the fused feature vector 201, the corresponding reasoning logic combination 301 and the weight 302 of each reasoning logic are determined, and collaborative reasoning is performed on the fused feature vector 201 according to the reasoning logic combination 301.

[0057] Specifically, the fusion feature vector 201 is used to characterize the current industrial operating state after knowledge constraints and multimodal fusion, and on this basis, scene features corresponding to the current industrial operating state are formed.

[0058] Different industrial scenarios present different hidden dangers and require different judgments. Applying a fixed reasoning method to all scenarios would lead to unnecessary complex reasoning processes in simple scenarios, or a lack of sufficient reasoning basis in complex scenarios. Therefore, a reasoning logic combination 301 is determined based on the characteristics of the current scenario and matches the current industrial state.

[0059] While determining the reasoning logic combination 301, the contribution of each reasoning logic in the combination is configured according to the current scenario, forming the weight corresponding to each reasoning logic. The reasoning logic weight reflects the contribution ratio of the corresponding reasoning result in the comprehensive judgment of the current scenario, enabling different reasoning methods to form differentiated collaborative relationships for the current industrial scenario.

[0060] After determining the reasoning logic combination 301 and its corresponding weights, each reasoning logic conducts a hidden danger analysis based on the fusion feature vector 201, and generates corresponding reasoning results.

[0061] This approach allows the reasoning process to adjust the logic involved in the reasoning and its contribution ratio according to changes in the industrial scenario, avoiding the use of a single reasoning logic or a fixed combination of reasoning results. This ensures that the knowledge information and on-site data contained in the fused features can continue to play a role in subsequent hazard assessment.

[0062] The fusion reasoning result 401 is formed based on the reasoning result of the reasoning logic combination 301 and the weights 302 of each reasoning logic. The fusion reasoning result 401 is then calibrated based on the knowledge confidence probability in the knowledge graph of high-risk industrial hazards to obtain the identification result 402 of high-risk industrial hazards.

[0063] Specifically, each reasoning logic generates a corresponding reasoning result after completing the hazard analysis. Since different reasoning logics have varying degrees of applicability to the current industrial scenario, no single result is directly selected as the final judgment. Instead, the reasoning results are merged based on the corresponding weights determined in the previous step.

[0064] The integrated reasoning result 401 comprehensively reflects the judgments made by different reasoning logics regarding the current state of industrial hazards, so that the final identification result does not depend on a single reasoning logic and retains the collaborative information between different reasoning methods.

[0065] The knowledge confidence probability in the knowledge graph of high-risk industrial hazards is used to characterize the credibility of the corresponding industrial knowledge relationships. After the fusion reasoning result 401 is formed, it is further calibrated according to the knowledge confidence probability corresponding to the current hazard, so that the reasoning result based on field data is further constrained by the credibility of industrial knowledge.

[0066] After calibration using knowledge confidence probability, the knowledge graph is no longer just a source of information retrieval after inference, but continuously participates in the identification of industrial hidden dangers in the process of front-end semantic mapping, intermediate feature fusion and final inference calibration.

[0067] The final high-risk industrial hazard identification result 402 integrates industrial knowledge, on-site multimodal perception features, reasoning logic corresponding to the current industrial scenario, and the credibility of knowledge, to achieve continuous collaboration of knowledge and data in the main processing link of hazard identification.

[0068] like Figure 2As shown in the figure, the temporal relationship between the technical steps of this invention is represented according to the actual data processing order. First, the knowledge embedding vector 101 and the industrial site multimodal feature vector 102 are semantically aligned to form a unified semantic feature 103, enabling knowledge information from different sources and on-site perception information to participate in subsequent calculations on the same semantic basis. Then, the modal benchmark weights provided by the knowledge graph and the real-time data quality are used to guide the fusion of each modal feature to form a fused feature vector 201. After the fused feature vector 201 is formed, the reasoning logic combination 301 and the weights 302 of each reasoning logic are determined according to the candidate hazard category, data quality, anomaly degree, and working condition complexity reflected therein, and collaborative reasoning is performed according to the corresponding combination. After each reasoning logic outputs the hazard judgment result, the results are fused according to the weights 302 of each reasoning logic to form a fused reasoning result 401, and the result is calibrated by knowledge confidence probability and cross-validation between different reasoning logics to obtain the high-risk industrial hazard identification result 402.

[0069] Example 2

[0070] This is the second embodiment of the present invention. Unlike the previous embodiment, this embodiment establishes projection relationships between the knowledge side and the data side respectively, and uses the sample relationships between similar and different hidden dangers to constrain the projection results, thereby solving the problem that the knowledge embedding vector 101 and the industrial site multimodal feature vector 102 are difficult to directly correspond in the original semantic space.

[0071] Mapping the knowledge embedding vector 101 corresponding to the knowledge graph of high-risk industrial hazards and the multimodal feature vector 102 of the industrial site to a unified semantic space includes projecting the knowledge embedding vector 101 and the multimodal feature vector 102 of the industrial site using a knowledge-side projection matrix and a data-side projection matrix, respectively. This constructs positive sample pairs between the knowledge embedding vector 101 and the multimodal feature vector 102 of the industrial site corresponding to the same hazard type, and negative sample pairs between the knowledge embedding vector 101 and the multimodal feature vector 102 of the industrial site corresponding to different hazard types. The similarity between the positive and negative sample pairs in the unified semantic space is constrained based on bilateral inter-side contrast loss, and the knowledge-side projection matrix and the data-side projection matrix are adjusted according to the constraint results to ensure that the knowledge semantics and multimodal data semantics of the same hazard type correspond within the unified semantic space.

[0072] Specifically, bidirectional projection semantic alignment is used to map knowledge vectors from different sources and industrial field data features to the same high-dimensional semantic space, so that knowledge representations and data representations of the same type of hidden danger are close to each other, while knowledge representations and data representations of different types of hidden dangers are far apart.

[0073] The knowledge graph of high-risk industrial hazards is transformed into knowledge vectors through knowledge vectorization. In the reactor implementation scenario, three algorithms—TransE, RotatE, and RGCN—were used in tandem for knowledge graph embedding, ultimately generating a 128-dimensional knowledge embedding vector 101.

[0074] Industrial field multimodal data are processed through spatiotemporal alignment, standardization, and feature extraction to form data features. Before entering the unified semantic space, knowledge vectors and data features are transformed through corresponding projection relationships.

[0075] The knowledge side sets the projection matrix as follows: , in, For the knowledge-side projection matrix, To unify the dimensions of the semantic space, This represents the original feature dimension.

[0076] Set the projection matrix on the data side: , in, This is the data-side projection matrix.

[0077] The aforementioned knowledge-side projection relationship and data-side projection relationship are established because the knowledge embedding vector and the multimodal data features from the industrial site originate from knowledge graphs and field perception data, respectively. These differ in their original feature dimensions and semantic distribution, making it impossible to directly determine the semantic correspondence for the same industrial hazard based solely on vector distance. Therefore, the two types of features are transformed using knowledge-side projection matrices and data-side projection matrices, respectively, to bring them into a unified semantic space with consistent dimensions. The calculated projected knowledge representation and projected data representation are not directly used as the high-risk industrial hazard identification result 402, but rather as inputs for subsequent similarity calculations of positive and negative sample pairs. Furthermore, the two projection matrices are trained using bilateral inter-bilateral contrastive loss to establish the correspondence between knowledge semantics and field data semantics.

[0078] Knowledge Vector With data characteristics The representation after projection onto a unified space is: , in, For the projected knowledge representation, For knowledge vectors, This is the projected data representation. For data characteristics.

[0079] Both projection matrices are initialized using a He normal distribution to ensure stable gradient propagation during the initial training phase. In the reactor implementation scenario, the 128-dimensional knowledge embedding vector 101 is mapped to a 256-dimensional unified semantic space through projection processing, and the data features formed by the on-site perception data are synchronously entered into the same 256-dimensional unified semantic space.

[0080] After completing the initial projection, a positive and negative sample relationship of knowledge data is constructed based on the type of hidden danger.

[0081] Positive sample pairs are constructed for knowledge vectors and data features corresponding to the same type of hazard. For example, the knowledge vector corresponding to the over-temperature hazard and the data features formed by the actual over-temperature state of the reactor constitute a positive sample pair.

[0082] Negative sample pairs are constructed by combining knowledge data corresponding to different types of potential hazards. For example, the knowledge vector corresponding to the overheating hazard and the data features formed by the sealing leakage status constitute a negative sample pair.

[0083] By using the relationship between positive and negative samples to train two projection matrices, we can make similar knowledge data have high similarity while suppressing the similarity between dissimilar knowledge data.

[0084] The projection matrix parameters are optimized using the bilateral contrast loss function, as follows: , in, This represents the bilateral comparative loss value. This refers to the number of samples within a batch. The positive sample margin threshold, The negative sample marginal threshold, No. The knowledge representation formed by mapping a knowledge vector through a knowledge-side projection matrix. For the first The data representation formed by mapping data features through a data-side projection matrix For the first The data representation is formed by mapping data features through a data-side projection matrix.

[0085] The semantic alignment requirement—that knowledge representations of the same type of hazard should be similar to each other, and that knowledge representations and data representations of different types of hazard should remain distinct—is transformed into numerical constraints that can be used for training the projection matrix. Positive sample terms are used to generate loss when the similarity between knowledge representations and data representations of the same type of hazard is insufficient, while negative sample terms are used to generate loss when the similarity between different types of hazard is too high. This avoids aliasing that occurs when only dimensionality transformation is performed but different categories remain within a unified semantic space.

[0086] The calculated bilateral contrast loss value is used as the basis for training and optimizing the knowledge-side projection matrix and the data-side projection matrix. The parameters of the two projection matrices are adjusted in reverse according to the loss value. As the loss gradually converges, the knowledge representation and data representation of the same type of hidden danger form a stable aggregation in a unified semantic space, and semantic intervals are maintained between different types of hidden dangers.

[0087] The positive sample marginal threshold is set to 0.7, the negative sample marginal threshold is set to 0.3, and the batch size is set to 32. The first term is used to ensure that the similarity between the knowledge representation and the data representation of the same hazard type is not less than 0.7, and the second term is used to ensure that the similarity between the knowledge representation and the data representation of different hazard types is not higher than 0.3.

[0088] Specifically, the positive sample margin threshold is set to 0.7 and the negative sample margin threshold is set to 0.3. This creates a constraint in the unified semantic space that aggregates similar hazards and separates dissimilar hazards, ensuring that the knowledge representation and data representation of the same hazard type maintain a high similarity, while preserving sufficient semantic spacing between different hazard types to reduce the possibility of feature overlap between different hazards. Under the parameter configuration of this embodiment, the loss function tends to stabilize after 50 iterations, and the cosine similarity between the knowledge representation and data representation of similar hazards stabilizes above 0.75, indicating that the above margin thresholds can meet the requirements of semantic alignment between knowledge and data.

[0089] It should also be noted that by setting separate knowledge-side projection matrices and data-side projection matrices, and by using bilateral inter-comparative loss of positive and negative sample pairs to constrain the projection relationship, the knowledge representation and data representation of the same type of hazard can be aggregated in a unified semantic space, while maintaining the distinction between knowledge representation and data representation of different types of hazard. This establishes a stable correspondence between knowledge semantics and on-site multimodal data semantics, providing a unified semantic foundation for subsequent knowledge-guided multimodal feature fusion. After iterative training, the cosine similarity between knowledge representation and data representation of the same type of hazard is stable above 0.75.

[0090] Example 3

[0091] Reference Figure 3 This is the third embodiment of the present invention. Unlike the previous embodiment, this embodiment establishes knowledge priors for multi-head attention by using modal benchmark weights in the knowledge graph, and adds knowledge features and data features of each modality to the feature sequence for joint self-attention calculation. This enables knowledge information to directly participate in feature interaction between multimodal data, solving the problem of relying solely on data correlation for fusion and lacking industrial mechanism constraints.

[0092] The knowledge-guided multimodal attention fusion includes determining the initial parameters of the projection matrix corresponding to multi-head attention based on the modal baseline weights in the knowledge graph of high-risk industrial hazards, forming a feature sequence by combining knowledge features in the unified semantic space with data features of each modality, performing joint self-attention calculation on the feature sequence, adjusting the information contribution of each modality data feature according to the degree of correlation between the knowledge features and data features of each modality, and forming enhanced features of each modality after knowledge-guided interaction.

[0093] Specifically, after semantic alignment, the projected knowledge features and data features of each modality are first subjected to L2 normalization to map different features to a unified spatial benchmark, thereby reducing the impact of feature vector magnitude differences on subsequent attention calculations.

[0094] The projection matrix corresponding to multi-head attention is initialized using modal baseline weights from the quintuple attribute set of the high-risk industrial hazard knowledge graph. For the first... Each attention head queries the initial value of the projection matrix, which is represented as follows: , in, For the first Initial values ​​for the query projection matrix of each attention head. For the first Baseline weights for class modalities For the first The attention head corresponds to the first Learnable basis matrices for modalities For activation function, For modalities, the bond projection matrix is... The initial parameters are determined in the same way.

[0095] For example, in the five-tuple knowledge corresponding to the risk of overheating in a reactor, the visual modality baseline weight is 0.2 and the temporal modality baseline weight is 0.8, which makes the temporal features have a relatively higher prior attention level in the initial stage of attention calculation.

[0096] The projected knowledge representation serves as the knowledge token in the feature sequence, forming the input sequence together with the data representations of different modalities. Knowledge features are not independently considered as data features for the query end to focus on unidirectionally; rather, they participate in self-attention computation together with data features, thereby simultaneously establishing connections between knowledge and data, as well as between different data modalities.

[0097] No. Each attention head calculates the query, key, and value respectively, represented as follows: , in, For the first The query feature matrix of each attention head. For the first The key feature matrix of each attention head, For the first The value feature matrix of each attention head, The input feature sequence is computed using multi-head self-attention. For the first The query projection matrix of each attention head. For the first The key projection matrix of each attention head For the first The value projection matrix of each attention head.

[0098] Further results were obtained: , in, For the first The output features of each attention head The dimension of the key vector.

[0099] The reactor scenario employs eight self-attention heads, with different attention heads analyzing the correlations between knowledge features, visual features, temporal features, voiceprint features, and other on-site data features from different relational dimensions.

[0100] The outputs of each attention head are concatenated and then linearly transformed. Simultaneously, a fused feature representation is formed through residual connections and layer normalization, as follows: , in, This refers to the fusion feature formed after multi-head attention fusion. This is for outputting the projection matrix.

[0101] The calculation of the query feature matrix, key feature matrix, and value feature matrix transforms knowledge features and data features from different modalities in a unified semantic space into information representations capable of calculating the degree of association. The matching results between query features and key features reflect the degree of association between knowledge features, visual features, temporal features, voiceprint features, and other on-site features. After normalization, corresponding attention allocation relationships are formed, and these attention relationships are then used to adjust the information contribution of the value features.

[0102] The calculation results of each attention head reflect the knowledge data interaction relationships under different correlation dimensions. After concatenation, linear transformation, residual connection, and layer normalization, a feature representation after knowledge-guided interaction is formed. The purpose of this calculation result is to obtain the effective features of each modality after knowledge semantic correlation adjustment, so that the data features that match the current hidden danger mechanism are strengthened and the features with low correlation are suppressed. The enhanced features formed continue to enter the subsequent dynamic weight allocation process, and their actual fusion contribution is further determined by real-time data quality and modality benchmark weights, finally forming a fusion feature vector 201.

[0103] Under the over-temperature condition of the reactor, the temperature time series change and the infrared temperature field anomaly can be mutually verified through self-attention formation. The knowledge features further enhance the information contribution of the features related to the temperature anomaly based on the over-temperature hazard mechanism, and suppress the data features with a low degree of correlation with the current hazard.

[0104] like Figure 3 As shown, the high-risk industrial hazard knowledge graph uses head entity E1, relation R, tail entity E2, attribute set A, and confidence probability P to jointly represent industrial hazard knowledge. Head entity E1 represents objects such as industrial scenarios, equipment, hazards, and operating parameters; relation R represents knowledge associations such as failure initiation, environmental triggering, constraint satisfaction, and synergistic effects; and tail entity E2 represents faults and their potential consequences. Head entity E1 and relation R are processed through knowledge embedding to form a knowledge embedding vector 101, enabling entity relations in the knowledge graph to participate in subsequent unified semantic mapping. Attribute set A further stores knowledge attributes such as modality baseline weights, temporal evolution features, hazard risk levels, fusion weight initialization, inference scheduling priority, and evolution prediction benchmarks. Modality baseline weights constrain the prior contributions of each modality in the knowledge-guided fusion process. Combined with real-time data quality, the enhanced features of each modality are dynamically weighted to form a fusion feature vector 201. Inference scheduling-related attributes are used to form the attention bias and scheduling code required for scene adaptive scheduling, and further participate in the determination of inference logic combination 301 and the weights 302 of each inference logic. The confidence probability P is formed based on expert calibration and historical data. It is used to quantify the uncertainty of the corresponding knowledge relationship and serves as the basis for knowledge confidence calibration in the subsequent reasoning and fusion stage.

[0105] It should also be noted that by using the modal baseline weights in the knowledge graph of high-risk industrial hazards to determine the initial parameters of the multi-head attention projection matrix, and by having knowledge features and data features of each modality participate in the joint self-attention calculation, prior knowledge can directly participate in the adjustment of the information contribution of different modal features. At the same time, it can establish the correlation between knowledge features and data features as well as between different data modalities, thereby strengthening the effective features related to the current hazard mechanism and suppressing the interference features with low correlation, thus improving the degree of integration of knowledge information with multimodal data from the industrial site.

[0106] Example 4

[0107] This is the fourth embodiment of the present invention. Unlike the previous embodiment, this embodiment evaluates the data quality of different modalities in real time and dynamically adjusts the contribution ratio of different modalities by combining the modal baseline weights and hazard category candidate results in the knowledge graph. This enables the fusion weights to change with industrial operating conditions and on-site data status, solving the problem that fixed fusion weights are difficult to adapt to complex on-site environments.

[0108] Determining the fusion weights for each modal data feature includes: forming corresponding data quality scores based on the clarity, signal-to-noise ratio, and data completeness of each modal data; correcting the modal baseline weights in the high-risk industrial hazard knowledge graph based on the hazard category candidate results; normalizing the corrected modal baseline weights with the data quality scores to obtain the fusion weights for each modal data feature; setting lower limit constraints on the fusion weights of each modal data feature; and weighting and aggregating each modal enhancement feature according to the fusion weights to form a fusion feature vector 201.

[0109] Specifically, after multi-head attention fusion is completed, the real-time quality of data from different modalities is evaluated. Data quality is jointly determined by clarity or data stationarity, signal-to-noise ratio, and data completeness, and is expressed as follows: , in, For the first Modal data quality scoring For the first Image sharpness score for visual modalities (corresponding to data stationarity score for temporal modalities). The signal-to-noise ratio is scored. Score the data completeness. , , The weights are the corresponding evaluation indicators.

[0110] In the reactor scenario, the weights for clarity or data stability, signal-to-noise ratio, and data integrity are set to 0.3, 0.4, and 0.3, respectively.

[0111] Specifically, the weights for clarity or data stability, signal-to-noise ratio (SNR), and data completeness are set to 0.3, 0.4, and 0.3, respectively. The optimal approach is that industrial field data is susceptible to electromagnetic interference, mechanical vibration, and environmental noise, and the SNR has a more direct impact on feature effectiveness, hence the higher weight of 0.4. Simultaneously, a weight of 0.3 retains the evaluation role of clarity or data stability and data completeness, preventing data quality scoring from over-reliance on a single indicator. This weighting configuration of 0.3, 0.4, and 0.3 effectively balances noise resistance with feature identifiability and data continuity, allowing the data quality score to more comprehensively reflect the current actual effectiveness of each modality and providing a more stable evaluation basis for subsequent dynamic adjustments to the fusion weights of each modality.

[0112] While generating data quality scores, lightweight coarse-grained pre-classification is used to pre-judge potential hazards and obtain corresponding hazard category candidate results. Coarse-grained classification is used to distinguish between major hazard categories such as temperature, pressure, leakage, and vibration, without directly making the final determination of specific hazard subcategories.

[0113] The modal baseline weights in the knowledge graph are adjusted based on the candidate hazard probabilities. Modalities with a high degree of correlation with the current candidate hazard have their corresponding baseline contributions increased according to their candidate probabilities, while other modalities are adjusted accordingly. The total baseline weights of all modalities remain unchanged before and after the adjustment.

[0114] The modal baseline weights and real-time data quality, after normalization, yield the final fusion weights, expressed as follows: , in, For the final fusion weight, For the first Modal baseline weights for class modes To correspond to the real-time data quality score, The number of modes participating in the fusion.

[0115] This fusion weighting formula addresses the issues that prior knowledge alone cannot reflect the real-time data status, and that data quality alone cannot reflect the degree of modality dependence of different hazard mechanisms. The modality baseline weight, corrected by the hazard category candidate results, reflects the prior importance of the corresponding modality for the current candidate hazard, while the data quality score reflects the actual effectiveness of the modality at the current moment. Both are used together in the calculation and then normalized to allow the final contributions of different modalities to be compared on a unified scale.

[0116] The calculated final fusion weights are directly used to control the contribution ratio of each modality's enhanced features in the fusion feature vector 201. That is, the enhanced features after knowledge-guided interaction are weighted and aggregated according to the final fusion weights corresponding to each modality. Therefore, modalities with high correlation to the current hidden danger mechanism and good real-time data quality contribute more to the fusion feature vector 201, while modalities with declining data quality contribute less. The resulting fusion feature vector 201 continues to serve as input for scene feature analysis and subsequent multi-logic collaborative reasoning. Therefore, this fusion weight formula constitutes a direct connection between real-time data quality and subsequent reasoning results.

[0117] In the reactor over-temperature scenario, the temperature sensing data completeness reached 98% and the signal-to-noise ratio reached 45dB. Meanwhile, the temporal mode corresponding to the over-temperature hazard has a high modal reference weight. After dynamic adjustment, the temporal mode fusion weight is increased to about 0.68, making the current high-quality temporal information with a strong relationship with the over-temperature mechanism the main fusion basis.

[0118] To prevent candidate type corrections and changes in data quality from completely suppressing effective modes, a lower limit is set for the modality baseline weights before normalization. The baseline weight of any normal effective mode is no less than half of the corresponding weight value of that mode under uniform distribution.

[0119] The enhanced features of each modality are weighted and aggregated according to the final fusion weight to form the final fused feature. When the visual data quality is high, the contribution of visual features is increased; when the temporal data is complete and stable, the contribution of temporal features is increased, so that the fused feature can change with the actual data state on site.

[0120] To address data anomalies in industrial settings, missing modal features are processed using a combination of linear interpolation and mean completion, while outliers are identified and removed using the 3σ criterion.

[0121] When a single mode completely fails, the fusion weight of the corresponding mode is adjusted to 0, and the enhanced features and prior knowledge of the remaining effective modes are used to complete the fusion and subsequent hazard identification. For example, when the flange acoustic sensor experiences a data interruption lasting more than 10 seconds due to electromagnetic interference, the acoustic mode no longer participates in the current fusion calculation, and the hazard identification process is maintained by relying on effective modes such as infrared images, temperature, and vibration.

[0122] It should also be noted that by evaluating the real-time data quality of each modality based on clarity, signal-to-noise ratio, and data completeness, and by adjusting the modality baseline weights in conjunction with the candidate results of hazard categories, and then dynamically determining the actual contribution of each modality through weight lower limit constraints and weighted aggregation, the fusion weights can be adjusted according to changes in hazard type and on-site data quality. This reduces the interference of low-quality modalities on the fusion results and avoids excessive suppression of effective modalities due to weight adjustments, thereby improving the adaptability and stability of multimodal feature fusion under complex working conditions. After removing the knowledge-guided dynamic weight processing, the average recognition accuracy decreased from 94.7% to 87.2%, and the false alarm rate increased from 3.1% to 7.9%, indicating that dynamic weight processing can improve the effectiveness of multimodal fusion results.

[0123] Example 5

[0124] This is the fifth embodiment of the present invention. Unlike the previous embodiment, this embodiment uses the current industrial scenario as the reinforcement learning state, takes the combination of different inference logics and their corresponding weights as selectable actions, and forms reward feedback based on the accuracy of inference and the computation time, so that different industrial scenarios can be matched with the corresponding combination of inference logic 301, thus solving the problem that fixed inference methods are difficult to balance the processing efficiency of simple scenarios and the recognition accuracy of complex scenarios.

[0125] Determining the corresponding inference logic combination 301 and the weights 302 of each inference logic includes: taking scene features as the state of reinforcement learning; taking the combination and corresponding weights of symbolic rule inference, statistical law inference, causal relationship inference, and equipment mechanism inference as the actions of reinforcement learning; determining the inference logic combination 301 and corresponding weights according to the inference scheduling strategy corresponding to the current state; executing hazard inference according to the determined inference logic combination 301; and generating a reward value based on the hazard identification accuracy and inference time; iteratively updating the inference scheduling strategy based on the reward value, so that subsequent scene features determine the corresponding inference logic combination 301 and the weights 302 of each inference logic according to the updated inference scheduling strategy.

[0126] Specifically, the reasoning stage includes symbolic rule reasoning, statistical law reasoning, causal relationship reasoning, and equipment mechanism reasoning. Different reasoning methods are used to conduct complementary analysis of the hidden danger status reflected by the fusion characteristics.

[0127] Symbolic rule reasoning is based on industrial safety standards and safety red lines. It matches structured on-site operating parameters with corresponding rule conditions to achieve compliance judgment.

[0128] Statistical regularity inference utilizes random forest analysis to identify anomalous patterns in historical data. This addresses potential risks. The confidence level of the output of statistical regularity inference is expressed as: , in, To use statistical regularity reasoning to target hidden dangers The confidence level of the output. The number of decision trees in the random forest. For the first Decision trees based on fusion features Output of hazard categories, This is an indicator function.

[0129] Causal association reasoning analyzes the causal relationship between abnormal features and hidden dangers based on the causal association chain in the knowledge graph of high-risk industrial hidden dangers, and forms the posterior probability of hidden dangers and the contribution ranking of different causal factors.

[0130] The equipment mechanism reasoning is based on the dynamics and thermodynamics of the equipment to verify whether the abnormal characteristics on site meet the real physical operating laws, and to form fault judgment, characteristic matching degree and hidden danger evolution time window.

[0131] Based on the four types of reasoning methods, Q-learning is used to establish a scenario-adaptive scheduling strategy, and the reasoning scheduling process is modeled as a Markov decision process.

[0132] The scene features serve as the state space, including the candidate distribution code of the hidden danger category, the comprehensive score of the current data quality, the degree of feature anomaly, and the complexity of the working condition. Each dimension is normalized and then used as the state input for inference scheduling.

[0133] The action space consists of 301 inference logic combinations and their corresponding weights. During typical engineering operations, eight sets of actions are pre-set to cover different processing needs, including high-speed response for single inference, routine detection for two types of inference, enhanced analysis for three types of inference, and high-precision calculation for four types of inference.

[0134] Simple explicit anomaly scenarios can employ a low-overhead combination of symbolic rule reasoning; regular anomaly scenarios can employ a combination of symbolic rule reasoning and statistical regularity reasoning; multi-parameter coupled anomaly scenarios can incorporate causal relationship reasoning; and high-risk complex scenarios employ all four types of reasoning.

[0135] When the temperature of the medium in the reactor continues to rise and is accompanied by uneven distribution of the infrared temperature field in the reactor body, the working condition complexity is 0.65. A high-precision combination involving four types of reasoning is adopted, with the weights of symbolic rule reasoning, statistical law reasoning, causal relationship reasoning and equipment mechanism reasoning being 0.2, 0.3, 0.3 and 0.2, respectively.

[0136] Setting the operating condition complexity to 0.65 is advantageous because this value can characterize the significant multi-feature coupling when the medium temperature continues to rise and is accompanied by uneven distribution of the infrared temperature field. This is suitable for triggering a high-precision combination involving four types of reasoning, avoiding insufficient judgment basis caused by using only a single or a small number of reasoning logics under complex operating conditions. At the same time, the weights of symbolic rule reasoning, statistical law reasoning, causal relationship reasoning, and equipment mechanism reasoning are set to 0.2, 0.3, 0.3, and 0.2, respectively. This allows statistical law reasoning and causal relationship reasoning to make higher contributions, focusing on analyzing continuous changing trends and the correlation between multiple parameters, while retaining the verification role of symbolic rule reasoning and equipment mechanism reasoning in safety rules and physical operating laws. The sum of the weights is 1, which can highlight the analysis of complex abnormal features while avoiding an excessive proportion of a single reasoning logic, improving the balance and reliability of multi-logic collaborative judgment.

[0137] Reinforcement learning rewards based on the accuracy of hazard identification and the time taken for reasoning, represented as follows: , in, For state Next action The reward value obtained afterwards The accuracy rate of hazard identification after applying the corresponding reasoning combination to the current scenario. The reasoning time after normalization processing For precision weights, Efficiency weight.

[0138] Taking a single inference scheduling under complex operating conditions of a reactor as an example, Q-learning selects a dynamic scheduling action involving four types of inference based on the current scene characteristics. The hazard identification accuracy corresponding to this action is 0.92, and the normalized inference time is 0.4. Substituting the accuracy weight of 0.7 and the efficiency weight of 0.3 into the reward function, we get: , In contrast, under the same working conditions, when four types of reasoning are used with fixed equal weights, the corresponding hazard identification accuracy rate is 0.88, and the normalized reasoning time is 0.5. Therefore, the corresponding reward value is: , Therefore, the overall reward value obtained by using scenario-adaptive scheduling is 12.4% higher than that obtained by the fixed equal weight method.

[0139] The explanation states that dynamically determining the combination of inference logic 301 and the weights of each inference logic 302 based on the current scenario can improve the accuracy of hazard identification while controlling the computational overhead of inference. This enables reinforcement learning scheduling to form a corresponding scheduling relationship between identification accuracy and inference efficiency for industrial scenarios of different complexities.

[0140] In typical industrial scenarios, the accuracy weight is set to 0.7 and the efficiency weight is set to 0.3, so that the inference strategy prioritizes the accuracy of hazard identification and simultaneously constrains unnecessary computational overhead.

[0141] Q-learning uses a temporal difference approach for iterative updates, which can be represented as follows: , in, For state Next action State-action value For learning rate, The reward value obtained after performing the current inference scheduling action. As a discount factor, The scene state corresponding to the next sampling window after the current inference action is performed. The discount factor is set to 0.9 for the candidate inference scheduling actions that can be selected in the next scenario state.

[0142] Training phase adopts - A greedy strategy balances exploration and exploitation. In the initial training phase, the exploration probability of new reasoning combinations is increased; as training progresses, the exploration probability is gradually decreased, allowing the reasoning scheduling strategy to gradually stabilize.

[0143] The offline training phase utilizes labeled historical industrial data to calculate accuracy based on the matching degree between inference results and actual hazard labels, and completes the initial inference scheduling strategy training. During the online operation phase, auxiliary feedback is generated based on the degree of discrepancy between different inference results to dynamically adjust the scheduling strategy, and further calibration is performed using manually reviewed samples every 15 days.

[0144] During online operation, the dispersion of confidence scores between different inference logic outputs can be used to evaluate the discrepancy of the current inference results. Taking a typical edge case sample as an example, the confidence scores of symbolic rule inference, statistical regularity inference, causal correlation inference, and equipment mechanism inference are 0.15, 0.88, 0.40, and 0.92, respectively. The average confidence scores of the four types of inference are: , Calculate the degree of divergence of inference based on the standard deviation of the confidence levels for the four types of inference: , The calculation yielded: When the preset divergence judgment value is 0.3, the inference divergence degree of this sample is 0.325, which is higher than 0.3. This indicates that there are significant differences in the judgments formed by different inference logics regarding the current industrial state. Therefore, this sample is marked as a high divergence sample and is used for subsequent periodic calibration of the scheduling strategy after manual review, so that the inference scheduling strategy can further adapt to changes in the field working conditions.

[0145] It should also be noted that by using scene features as reinforcement learning states, inference logic combination 301 and corresponding weights as reinforcement learning actions, and iteratively updating the inference scheduling strategy based on reward feedback generated according to the accuracy of hazard identification and inference time, the appropriate inference logic combination 301 and weights 302 of each inference logic can be adaptively selected according to the complexity of the industrial scene. This reduces unnecessary inference overhead in simple scenarios and achieves more comprehensive collaborative judgment in complex scenarios, thus balancing the accuracy of hazard identification and the efficiency of on-site inference. Compared with fixed equal-weighted voting, the average identification accuracy has increased from 91.3% to 94.7%, the root cause localization accuracy has increased from 82.5% to 89.2%, and the single-sample inference latency at the edge has been stably controlled within 12ms.

[0146] Example 6

[0147] Reference Figure 4 This is the sixth embodiment of the present invention. Unlike the previous embodiment, after obtaining different reasoning results through scene adaptive reasoning, this embodiment forms a high-risk industrial hazard identification result 402 by weighting the reasoning results, calibrating knowledge confidence, and cross-validating. Furthermore, it reorganizes the effective modal features for cases of missing multimodal data, outliers, and single modal failures, thereby solving the problems that single reasoning anomalies can easily affect the final identification result and that abnormal perception data can easily cause fusion and reasoning distortion.

[0148] The fusion reasoning result 401 formed based on the reasoning results of the reasoning logic combination 301 and the weights 302 of each reasoning logic includes: weighting each reasoning result according to the weights corresponding to each reasoning logic to form an initial fusion reasoning result 401; using the knowledge confidence probability of the corresponding causal rules in the high-risk industrial hazard knowledge graph as a calibration coefficient to perform knowledge confidence calibration on the initial fusion reasoning result 401; cross-validating the judgment results of different reasoning logics; filtering the corresponding abnormal results when both symbolic rule reasoning and equipment mechanism reasoning are judged as normal and statistical regularity reasoning is judged as abnormal; and determining whether the hazard is established based on the cross-validation results. When the hazard is determined to be established, a high-risk industrial hazard identification result 402 is formed.

[0149] like Figure 4As shown, the scenario-adaptive collaborative reasoning uses the fused feature vector 201 and the original working condition structured parameters as input information for the current industrial state. The fused feature vector 201 represents the field state after knowledge constraints and multimodal dynamic fusion, while the original working condition structured parameters provide corresponding operational parameters for symbolic rule reasoning and equipment mechanism reasoning. Scene state features are extracted based on the fused feature vector 201, and combined with basic priority triggering rules, the scenario-adaptive scheduling determines the reasoning logic combination 301 and the weights 302 of each reasoning logic corresponding to the current scenario. The scenario state includes candidate hazard categories, data quality, feature anomaly degree, and working condition complexity; the scheduling action corresponds to a preset combination of reasoning modules and their weights. After determining the scheduling results, symbolic rule reasoning makes compliance judgments based on safety regulations and production rules; statistical regularity reasoning analyzes abnormal patterns in historical data; causal relationship reasoning locates the root cause based on causal relationships in the high-risk industrial hazard knowledge graph; and equipment mechanism reasoning physically verifies abnormal states based on equipment dynamics and thermodynamic relationships. This allows different reasoning logics to form complementary judgments from the perspectives of rules, data, causality, and mechanism. Each reasoning result is weighted and fused according to the weight 302 of each reasoning logic to form a fused reasoning result 401. Then, the confidence probability in the knowledge graph of high-risk industrial hazards is used for confidence calibration, and the consistency of abnormal results is verified by cross-validation between different reasoning logics. When different reasoning results have sufficient rule basis, statistical basis, causal basis or mechanism basis, a high-risk industrial hazard identification result 402 is formed, and the hazard type, risk level, comprehensive confidence level, core root cause and disposal suggestions are output.

[0150] After the inference results are aggregated, they are weighted according to the weights 302 of each inference logic to form a fused inference result 401, which is then calibrated using the knowledge confidence probability of the corresponding causal rules in the high-risk industrial hazard knowledge graph. Subsequently, cross-validation between different inference logics is used to determine whether the current anomaly has sufficient rule-based and physical mechanism basis. For example, if both symbolic rule inference and equipment mechanism inference determine that the anomaly is normal, but only statistical regularity inference outputs an anomaly, the statistical anomaly is filtered out. After the above calibration and cross-validation, the high-risk industrial hazard identification result 402 is formed, and the hazard type, risk level, comprehensive confidence level, core root cause, and handling recommendations are output.

[0151] Four types of reasoning logics respectively generate hazard judgment results from the perspectives of rule compliance, statistical regularity, causal relationship, and equipment mechanism. After summarizing the reasoning results, they are weighted and fused according to the weights 302 of each reasoning logic to form a fused reasoning result 401. The knowledge confidence probability of the corresponding causal rules in the high-risk industrial hazard knowledge graph is further used to calibrate the knowledge confidence of the fused reasoning result 401, and the anomaly judgment is verified through cross-validation between different reasoning logics, finally forming the high-risk industrial hazard identification result 402.

[0152] Specifically, after the scenario-adaptive reasoning is completed, symbolic rule reasoning, statistical regularity reasoning, causal relationship reasoning, and equipment mechanism reasoning each form a corresponding hazard judgment result. Each reasoning result uniformly includes the hazard type, confidence level, and characteristic basis, so that the outputs of different reasoning logics can be subsequently fused in a unified form.

[0153] Symbolic rule-based reasoning primarily relies on safety regulations and red lines to form compliance judgments; statistical regularity reasoning uses anomaly patterns in historical data to form hazard judgments; causal association reasoning uses causal relationships in a knowledge graph to determine the correspondence between hazards and their causes; and equipment mechanism reasoning uses the operating mechanisms of industrial equipment to verify the physical rationality of abnormal conditions on site. Since different reasoning methods analyze the same industrial state from different perspectives, no single reasoning result is directly selected as the final identification result. Instead, a weighted fusion is performed based on the weights formed during the scenario-adaptive reasoning stage.

[0154] For hidden dangers The overall confidence level of potential hazards is calculated using the following formula: , in, Hidden danger The comprehensive hazard confidence level is obtained after fusing multiple logical reasoning results and calibrating knowledge confidence. The first allocation for the scheduling algorithm The weights of each inference module, Potential risks in the corresponding logical reasoning output Confidence level, The confidence calibration coefficient for knowledge.

[0155] The knowledge confidence calibration coefficient directly adopts the confidence probability of the corresponding causal rules in the knowledge graph of high-risk industrial hazards, so that the initial result formed by weighting different inference results is further constrained by the credibility of industrial mechanism knowledge.

[0156] Taking the overheating hazard of a reactor as an example, the adaptive reasoning scenario determines the weights of symbolic rule reasoning, statistical regularity reasoning, causal relationship reasoning, and equipment mechanism reasoning to be 0.2, 0.3, 0.3, and 0.2, respectively. The knowledge confidence probability corresponding to the overheating causal rule in the knowledge graph is 0.92. Therefore, the overall hazard confidence is weighted according to the confidence level formed by each reasoning logic for the overheating hazard, and then calibrated using 0.92.

[0157] Specifically, the weights of symbolic rule reasoning, statistical regularity reasoning, causal correlation reasoning, and equipment mechanism reasoning are set to 0.2, 0.3, 0.3, and 0.2, respectively. This is preferred because reactor overheating is a complex hidden danger with a continuous changing trend and multi-factor correlation characteristics. Therefore, appropriately increasing the weights of statistical regularity reasoning and causal correlation reasoning can strengthen the analysis of abnormal evolution trends and causal relationships, while retaining the verification role of symbolic rule reasoning and equipment mechanism reasoning in safety regulations and physical operating laws, avoiding an excessively high proportion of single reasoning logic. Setting the knowledge confidence probability to 0.92 indicates that the corresponding overheating causal rule has a high degree of knowledge credibility. Using it as a calibration coefficient can, while preserving the differences in multi-logic reasoning results, further constrain the final comprehensive hidden danger confidence by highly reliable industrial knowledge, thereby reducing the impact of accidental data anomalies or single reasoning biases on the final judgment.

[0158] In actual operation, symbolic rule reasoning determined that the over-temperature hazard was valid, with a corresponding rule confidence level of 0.92; statistical regularity reasoning formed an over-temperature hazard statistical confidence level of 0.79 based on historical over-temperature anomaly patterns; causal association reasoning further determined that the abnormal feed rate was the main cause, with a posterior probability of 0.76; equipment mechanism reasoning combined temperature, pressure, and feed rate for thermodynamic verification, obtaining a mechanism matching degree of 0.83, and judging that the time window for the current over-temperature state to evolve into the over-pressure state is approximately 12 minutes.

[0159] The 0.76 value corresponding to the abnormal feed rate is used to characterize the posterior probability of the root cause. It is not directly substituted into the calculation of the overall risk confidence score for the over-temperature hazard itself, thus avoiding the mixed use of probability quantities with different meanings. After each inference logic forms its own risk confidence score for the over-temperature hazard, they are merged according to the scenario scheduling weight and knowledge confidence calibration is performed, finally resulting in an overall over-temperature hazard confidence score of 0.86.

[0160] Specifically, a rule confidence score of 0.92 indicates a high degree of matching between the over-temperature state and the corresponding safety rules; a statistical confidence score of 0.79 reflects a strong consistency between the current characteristics and historical over-temperature anomaly patterns; a root cause posterior probability of 0.76 highlights the possibility that the abnormal feed rate is the main cause and distinguishes it from the hazard classification confidence score; a mechanism matching score of 0.83 verifies that the current anomaly conforms to the thermodynamic laws of the reactor, reducing the possibility of misjudgment caused by simple data anomalies; an evolution time window of approximately 12 minutes is used to quantify the urgency of the over-temperature developing into over-pressure, providing a time basis for subsequent risk management; and a final comprehensive confidence score of 0.86 is formed by weighting multiple inference results and calibrating them with knowledge confidence, avoiding reliance on a single inference result and ensuring that the final over-temperature hazard judgment has data, rule, causal, and mechanistic basis.

[0161] After completing the knowledge confidence calibration, the judgment results of different reasoning logics are further cross-validated. Cross-validation does not simply rely on the abnormal output of a single data model to form an early warning, but rather uses the complementary relationship between different reasoning logics to determine whether the current abnormal result has sufficient industrial mechanism basis.

[0162] When symbolic rule reasoning determines that the field parameters have not triggered the corresponding safety rules, equipment mechanism reasoning determines that the current operating state conforms to the physical laws of the equipment, and statistical rule reasoning outputs an anomaly alone, the anomaly generated by statistical rule reasoning is filtered as a false anomaly caused by data noise, and no formal high-risk industrial hazard identification result is formed 402.

[0163] For example, when subjected to strong electromagnetic interference in an industrial setting, some sensor data may exhibit transient distortions. Statistical reasoning may lead to anomaly judgments based on abnormal data patterns. In this case, symbolic rule reasoning is used to verify whether the actual operating parameters trigger safety thresholds, and then equipment mechanism reasoning is used to verify whether the abnormal characteristics conform to the actual operating rules of the equipment. When both types of verification indicate that the equipment is in a normal state, statistical anomalies are filtered out.

[0164] When the cross-validation results of different reasoning logics can jointly support the establishment of the current hidden danger, and the fusion result meets the corresponding judgment conditions after being calibrated by knowledge confidence probability, a high-risk industrial hidden danger identification result 402 is formed.

[0165] In field tests, the average false alarm rate of the complete solution was 3.1%, while the average false alarm rate of the fixed-weight multimodal fusion solution was 9.9%, a reduction of 6.8 percentage points. This indicates that the fusion of inference results, knowledge confidence calibration, and cross-validation can improve the reliability of the final hazard assessment.

[0166] For missing features in multimodal data, interpolation is performed to complete them, and abnormal outlier features are identified and removed. When a single modality fails, the fusion weight of the corresponding modality is adjusted to zero. The fusion feature vector 201 is re-determined based on the enhanced features corresponding to the remaining effective modalities and the prior knowledge provided by the knowledge graph of high-risk industrial hazards. Hazard inference is then performed based on the re-determined fusion feature vector 201.

[0167] Specifically, during long-term operation in industrial settings, multimodal sensing data may be affected by factors such as communication fluctuations, electromagnetic interference, sensor malfunctions, and changes in the on-site environment, leading to data loss, outliers, or complete mode failure in some modes. In response to these anomalies, the current hazard identification process is not immediately halted; instead, different modal data are processed accordingly based on the severity of the anomaly.

[0168] For locally missing features in multimodal data, a combination of linear interpolation and mean imputation is used. For short-term missing data in a continuous sampling sequence, linear interpolation is performed based on the valid data before and after the missing position. For missing features that cannot be directly recovered using adjacent valid data, imputation is completed by combining the statistical mean of the corresponding data, so that the current modality can still form a continuous feature representation.

[0169] For outlier features in multimodal data, the following methods are used: The criteria are used for identification. When a data feature deviates significantly from the normal distribution range of the corresponding data, the feature is identified as an outlier and removed, reducing the impact of sudden collection anomalies on subsequent attention interaction and fusion weight calculation.

[0170] Different handling methods are used for partial data loss and complete failure of a single modality. When partial data loss occurs, the data continuity of the corresponding modality is restored through interpolation and completion; when a certain modality can no longer form effective features, the completion result is no longer used to maintain the fusion contribution of that modality, but the current fusion weight of the corresponding modality is adjusted to 0.

[0171] After adjusting the fusion weight of the failed mode to 0, the enhanced features corresponding to the other still valid modes are used to participate in the fusion again, and the fusion feature vector 201 is re-determined by combining the prior knowledge provided by the high-risk industrial hazard knowledge graph. As a result, the failed mode no longer makes erroneous contributions to the current fusion result, while the remaining valid modes are still guided by the industrial hazard knowledge.

[0172] For example, in the scenario of identifying potential flange seal leaks, when the acoustic sensor is subjected to electromagnetic interference and experiences a data interruption lasting more than 10 seconds, the acoustic mode is determined as the current failed mode, and the acoustic mode fusion weight is adjusted to 0.

[0173] At this point, the still valid field modes such as infrared images, temperature, and vibration continue to form corresponding enhanced features. The knowledge graph of high-risk industrial hazards provides prior knowledge related to sealing leakage hazards. Based on the enhanced features of the remaining valid modes, a new fusion feature vector 201 is formed, and then the newly formed fusion feature vector 201 is used for subsequent hazard reasoning.

[0174] After the fusion feature vector 201 is redefined, the subsequent reasoning process is still executed according to the reasoning logic combination 301 corresponding to the current scenario, thereby avoiding the interruption of the overall hidden danger identification link due to the failure of a single sensing device.

[0175] For data distortion scenarios caused by strong noise, abnormal outliers are first identified and removed, and then corresponding features are formed based on the processed effective data. When noise only affects local data, it is restored through data completion. When noise causes the entire modality to fail to form effective information, the corresponding modality is excluded from the current fusion calculation.

[0176] By differentiating between data missing data handling, outlier handling, and single-modal failure handling, different levels of data anomalies can be handled with corresponding paths: local missing data does not directly cause modal failure, outlier data does not continue to participate in normal feature calculations, and complete failed modes do not continue to retain erroneous fusion contributions.

[0177] It should also be noted that under extreme operating conditions such as strong industrial noise and single-mode failure, the accuracy of hazard identification after the above processing can still be maintained above 90%, indicating that the fusion feature based on the enhanced features of the remaining effective modes and prior knowledge can improve the continuity and stability of the hazard identification process under abnormal perception conditions in industrial sites.

[0178] Example 7

[0179] This is the seventh embodiment of the present invention. This embodiment verifies the identification accuracy, false alarm rate, reasoning real-time performance, and technical effects of knowledge-guided dynamic fusion and scenario-adaptive collaborative reasoning of the knowledge-data dual-drive industrial hazard identification method using on-site measured data from the reactor industrial site.

[0180] This embodiment is based on 60 days of continuous on-site operation data from six batch reactors of the same model in a chemical production workshop. Multimodal data such as temperature, pressure, vibration, acoustic signature, infrared image, visible light image, and equipment operating status were collected. After safety engineering personnel completed sample verification and hazard labeling, a total of 12,860 valid samples were formed, including 9,240 samples under normal operating conditions and 3,620 samples with hazards. The hazard samples cover four typical industrial hazards: reactor over-temperature, over-pressure, sealing leakage, and personnel improper positioning.

[0181] The data was divided into training and testing sets in a 7:3 ratio. The testing set contained 3858 valid samples, including 2772 samples under normal operating conditions and 1086 samples with potential hazards. All comparison schemes were tested using the same dataset partitioning and the same edge computing hardware. Average recognition accuracy, average false alarm rate, hazard recall rate, precision, and single-sample inference latency were used as evaluation metrics.

[0182] During the testing process, the statistical results of the confusion matrix corresponding to the present invention solution were as follows: TP = 968 correctly detected hidden dangers, TN = 2686 correctly judged under normal operating conditions, FP = 86 false alarms under normal operating conditions, and FN = 118 missed hidden dangers.

[0183] Based on the above statistical results, the average recognition accuracy of the present invention is: , in, For average recognition accuracy, To ensure the correct number of potential hazards are detected, To ensure accurate quantity determination under normal operating conditions, The number of false alarms under normal operating conditions. This represents the number of unreported potential hazards.

[0184] Substituting the experimental data, we get: .

[0185] The average false alarm rate is calculated based on the proportion of normal samples that are incorrectly identified as potential hazards: , in, This represents the number of false alarms under normal operating conditions.

[0186] Substituting the experimental data, we get: .

[0187] The recall rate for potential hazards is: , in, This refers to the recall rate for potential hazards.

[0188] Substituting the experimental data, we get: .

[0189] The accuracy rate of hazard identification is: , in, To improve the accuracy of hazard identification.

[0190] Substituting the experimental data, we get: , Therefore, the present invention can reduce the proportion of false warnings under normal operating conditions while maintaining a high rate of hazard detection.

[0191] Further comparisons were made using a traditional pure rule-based scheme, a pure data-driven scheme, and a fixed-weight multimodal fusion scheme. The results of each scheme on the same test set are shown in Table 1.

[0192] Table 1. Performance Comparison Results of Different Industrial Hazard Identification Schemes

[0193] The present invention's solution improves the average recognition accuracy by 6.8 percentage points compared to the fixed-weight multimodal fusion solution, and reduces the average false alarm rate from 9.9% to 3.1%, a reduction of 6.8 percentage points. Although knowledge-guided multimodal fusion, scene-adaptive scheduling, and multi-logic collaborative reasoning increase some computational processes, the end-to-end single-sample reasoning latency is still 12ms, lower than the 20ms real-time response requirement in industrial settings.

[0194] To further verify the effect of knowledge-guided dynamic fusion weights on the identification results, the knowledge-guided dynamic weight processing was removed while keeping other processing procedures unchanged. This prevented different modalities from dynamically adjusting their contributions based on the modality baseline weights in the knowledge graph and the real-time data quality. Experimental results showed that the average recognition accuracy decreased from 94.7% in the complete scheme to 87.2%, while the average false alarm rate increased from 3.1% to 7.9%. This indicates that the modality baseline weights and real-time data quality jointly participate in the determination of fusion weights, which can reduce the interference of low-quality modalities and modalities with low correlation to the current hidden danger on the fusion feature vector 201.

[0195] Further verification of the role of multi-logic collaborative reasoning revealed that after removing multi-logic collaborative reasoning and retaining only single statistical regularity reasoning, the average identification accuracy decreased to 89.6%, the average false alarm rate increased to 6.3%, and the root cause localization accuracy decreased from 89.2% to 67.8%. This demonstrates that symbolic rule reasoning, statistical regularity reasoning, causal relationship reasoning, and equipment mechanism reasoning form complementary judgments from different perspectives, which can reduce the identification bias caused by insufficient judgment of complex industrial conditions by a single reasoning method and improve the ability to analyze the causes of hidden dangers.

[0196] We further maintained the four types of inference logic unchanged, only canceled the Q-learning scenario adaptive scheduling, and adjusted each inference logic to participate with fixed equal weights. The experimental results are shown in Table 2.

[0197] Table 2 Ablation Experiment Results of Core Inference Module and Scheduling Mechanism

[0198] When using four types of reasoning with fixed equal weights, the average recognition accuracy is 91.3% and the root cause localization accuracy is 82.5%. After using the present invention to dynamically determine the reasoning logic combination 301 and the weights 302 of each reasoning logic based on scene characteristics, the average recognition accuracy is increased to 94.7% and the root cause localization accuracy is increased to 89.2%, representing increases of 3.4 percentage points and 6.7 percentage points, respectively. The average false alarm rate is reduced from 5.2% to 3.1%.

[0199] The above experiments demonstrate that knowledge-guided dynamic fusion and scenario-adaptive multi-logic collaborative reasoning are not simply independent superpositions. Specifically, the knowledge graph participates in multi-modal fusion through modal baseline weights, enabling the resulting fusion feature vector 201 to more centrally represent effective information related to the current hazard mechanism. Based on this, the reasoning logic combination 301 and the weights 302 of each reasoning logic are determined according to the scenario characteristics reflected in the fusion feature vector 201, allowing different reasoning logics to form differentiated collaborations for the current working conditions. Each reasoning result is further calibrated by knowledge confidence probability and cross-validated to form a high-risk industrial hazard identification result 402, thereby jointly improving the accuracy and reliability of hazard identification in complex industrial scenarios.

[0200] In continuous operation stability testing, the initial test accuracy of the proposed solution was 94.7%, and the retest accuracy after 90 days of continuous operation was 92.9%. Therefore, the accuracy decay over 90 days is: , This indicates that even under continuously changing on-site conditions, the knowledge constraints, multimodal dynamic fusion, and scene adaptive reasoning methods adopted in this invention can maintain relatively stable hazard identification performance.

[0201] This embodiment verifies the practical role of each technical step in identifying high-risk industrial hazards 402 through comparative experiments on a unified test set and comparative experiments on knowledge-guided dynamic weighting, multi-logic collaborative reasoning, and Q-learning scheduling. Specifically, the solution achieved an average identification accuracy of 94.7% and an average false alarm rate of 3.1% on 3858 test samples, with a hazard recall rate of 89.1%, a precision rate of 91.8%, and a single-sample inference latency of 12ms. This demonstrates a balance between accuracy, reliability, and real-time on-site processing requirements in complex industrial scenarios.

[0202] Example 8

[0203] The eighth embodiment of the present invention provides a knowledge data dual-drive industrial hazard identification system, including a knowledge data semantic alignment module, a knowledge-guided multimodal fusion module, a scene adaptive collaborative reasoning module, and a reasoning fusion calibration module.

[0204] The knowledge data semantic alignment module is used to map the knowledge embedding vector 101 and the industrial site multimodal feature vector 102 to a unified semantic space to form a corresponding unified semantic feature 103.

[0205] The knowledge-guided multimodal fusion module is used for knowledge-guided multimodal fusion based on unified semantic features 103, and combines modality benchmark weights and real-time data quality to form a fusion feature vector 201.

[0206] The scene-adaptive collaborative reasoning module is used to determine the reasoning logic combination 301 and the weight 302 of each reasoning logic based on the scene features corresponding to the fused feature vector 201, and to perform collaborative reasoning.

[0207] The reasoning fusion calibration module is used to form a fusion reasoning result 401 based on the results of each reasoning logic and the corresponding weights, and to perform calibration in combination with the knowledge confidence probability to obtain the high-risk industrial hazard identification result 402.

[0208] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements a knowledge and data dual-drive industrial hazard identification system as proposed in the above embodiment.

[0209] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a knowledge-data dual-drive industrial hazard identification system as proposed in the above embodiment.

[0210] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, 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 this 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.

[0211] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0212] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0213] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0214] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A knowledge data dual-driven industrial hazard identification method, characterized in that, include: The knowledge embedding vector (101) corresponding to the knowledge graph of high-risk industrial hazards and the multimodal feature vector (102) of industrial site are mapped to a unified semantic space to form a unified semantic feature (103) corresponding to the knowledge semantics and multimodal data semantics. Based on the unified semantic features (103), knowledge-guided multimodal attention fusion is performed, and the fusion weights of each modal data feature are determined by combining the modal baseline weights in the knowledge graph of high-risk industrial hazards and the real-time data quality, forming a fusion feature vector (201) under knowledge constraints. Based on the scene features corresponding to the fusion feature vector (201), the corresponding reasoning logic combination (301) and the weight (302) of each reasoning logic are determined, and collaborative reasoning is performed on the fusion feature vector (201) according to the reasoning logic combination (301); Based on the reasoning results of the reasoning logic combination (301) and the weights (302) of each reasoning logic, a fusion reasoning result (401) is formed, and the fusion reasoning result (401) is calibrated based on the knowledge confidence probability in the knowledge graph of high-risk industrial hazards to obtain the identification result of high-risk industrial hazards (402).

2. The knowledge-data dual-driven industrial hazard identification method of claim 1, wherein: The process of mapping the knowledge embedding vector (101) corresponding to the knowledge graph of high-risk industrial hazards and the multimodal feature vector (102) of the industrial site to a unified semantic space includes, The knowledge embedding vector (101) and the industrial site multimodal feature vector (102) are projected using the knowledge-side projection matrix and the data-side projection matrix, respectively. The knowledge embedding vector (101) corresponding to the same type of hidden danger and the industrial site multimodal feature vector (102) are constructed as positive sample pairs, and the knowledge embedding vector (101) corresponding to different types of hidden dangers and the industrial site multimodal feature vector (102) are constructed as negative sample pairs. The similarity of the positive and negative sample pairs in the unified semantic space is constrained by the bilateral contrast loss, and the knowledge-side projection matrix and data-side projection matrix are adjusted according to the constraint results so that the knowledge semantics of the same hidden danger type corresponds with the multimodal data semantics in the unified semantic space.

3. The knowledge-data dual-driven industrial hazard identification method of claim 2, wherein: The knowledge-guided multimodal attention fusion includes... The initial parameters of the projection matrix corresponding to multi-head attention are determined based on the modal baseline weights in the knowledge graph of high-risk industrial hazards, and the knowledge features in the unified semantic space and the data features of each modality are combined to form a feature sequence. Joint self-attention calculation is performed on the feature sequence, and the information contribution of each modal data feature is adjusted according to the degree of correlation between the knowledge features and the modal data features to form each modal enhanced feature after knowledge-guided interaction.

4. The knowledge-data dual-driven industrial hazard identification method of claim 3, wherein: The determination of the fusion weights for each modal data feature includes, A corresponding data quality score is generated based on the clarity, signal-to-noise ratio, and data completeness of each modality of data; Based on the candidate results of the hidden danger category, the modal baseline weights in the knowledge graph of high-risk industrial hidden dangers are corrected. The corrected modal baseline weights are then normalized and calculated with the data quality score to obtain the fusion weights of each modal data feature. A lower limit constraint is set for the fusion weight of each modality data feature, and the enhanced features of each modality are weighted and aggregated according to the fusion weight to form the fusion feature vector (201).

5. The knowledge-data dual-driven industrial hazard identification method of claim 1 or 4, wherein: The determination of the corresponding reasoning logic combination (301) and the weight of each reasoning logic (302) includes, The scene features are used as the state of reinforcement learning, and the combination and corresponding weights of symbolic rule reasoning, statistical law reasoning, causal relationship reasoning and device mechanism reasoning are used as the actions of reinforcement learning. The reasoning logic combination (301) and the weight (302) of each reasoning logic are determined according to the reasoning scheduling strategy corresponding to the current state. The hidden danger reasoning is executed according to the determined reasoning logic combination (301), and a reward value is formed based on the hidden danger identification accuracy and reasoning time. The inference scheduling strategy is iteratively updated based on the reward value, so that subsequent scene features determine the corresponding inference logic combination (301) and the weight of each inference logic (302) according to the updated inference scheduling strategy.

6. The knowledge-data dual-driven industrial hazard identification method of claim 5, wherein: The process of forming a fused reasoning result (401) based on the reasoning result of the combination of reasoning logics (301) and the weights (302) of each reasoning logic includes, The inference results are weighted according to the weights corresponding to each inference logic to form the initial fusion inference result (401). The knowledge confidence probability of the corresponding causal rule in the knowledge graph of high-risk industrial hazards is used as the calibration coefficient to perform knowledge confidence calibration on the initial fusion reasoning result (401); Cross-validate the judgment results of different reasoning logics. When both symbolic rule reasoning and equipment mechanism reasoning are judged to be normal and statistical regular reasoning is judged to be abnormal, filter the corresponding abnormal results. Determine whether the hidden danger is established based on the cross-validation results. When the hidden danger is determined to be established, a high-risk industrial hidden danger identification result (402) is formed.

7. The knowledge-data dual-driven industrial hazard identification method of any one of claims 1, 2, or 6, wherein: It also includes, Interpolation is used to complete missing features in multimodal data, and outlier features are identified and removed. When a single modality fails, the fusion weight of the corresponding modality is adjusted to zero. The fusion feature vector (201) is re-determined based on the enhanced features corresponding to the remaining effective modalities and the prior knowledge provided by the knowledge graph of high-risk industrial hazards. Hazard inference is then performed based on the re-determined fusion feature vector (201).

8. A knowledge-data dual-drive industrial hazard identification system, employing the knowledge-data dual-drive industrial hazard identification method as described in any one of claims 1 to 7, characterized in that: It includes a knowledge data semantic alignment module, a knowledge-guided multimodal fusion module, a scene-adaptive collaborative reasoning module, and a reasoning fusion calibration module; The knowledge data semantic alignment module is used to map the knowledge embedding vector (101) and the industrial site multimodal feature vector (102) to a unified semantic space to form corresponding unified semantic features (103). The knowledge-guided multimodal fusion module is used to perform knowledge-guided multimodal fusion based on unified semantic features (103), and to form a fusion feature vector (201) by combining modality benchmark weights and real-time data quality. The scene adaptive collaborative reasoning module is used to determine the reasoning logic combination (301) and the weight (302) of each reasoning logic according to the scene features corresponding to the fused feature vector (201), and to perform collaborative reasoning; The reasoning fusion calibration module is used to form a fusion reasoning result (401) based on the result of the reasoning logic combination (301) and the weight (302) of each reasoning logic, and to calibrate it in combination with the knowledge confidence probability to obtain the high-risk industrial hazard identification result (402).

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the knowledge data dual-drive industrial hazard identification method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the knowledge data dual-drive industrial hazard identification method as described in any one of claims 1 to 7.

Images