Intelligent management and traceability method and system for disinfection and sterilization of biological safety cabinet

Abstract

The present application relates to the technical field of biological safety cabinet disinfection and sterilization, and particularly relates to a biological safety cabinet disinfection and sterilization intelligent management and tracing method and system, which comprises collecting multi-modal sterilization data flow and performing causal decoupling modeling, constructing and verifying a causal knowledge graph to form a trusted knowledge subgraph; extracting the implicit operation characteristics of an operator to identify different operation modes; for a task to be inherited, reverse tracing necessary causal paths and generating an inheritance scheme in combination with the operation characteristics; and in execution, correcting the knowledge graph according to feedback deviation and re-verifying. The present application realizes intelligent management of the sterilization process and trusted inheritance of operation experience, and improves the reliability and consistency of the sterilization effect.

Description

Technical Field

[0001] This invention relates to the field of biosafety cabinet disinfection and sterilization technology, and in particular to a method and system for intelligent management and traceability of biosafety cabinet disinfection and sterilization. Background Technology

[0002] As a crucial piece of biosafety equipment, the effectiveness of biosafety cabinets in sterilization directly impacts the biosafety level of a laboratory. Currently, the conventional practice in this field relies primarily on a combination of pre-set sterilization programs and manual operation. The sterilization process is typically executed by the equipment's built-in controller according to fixed parameters such as time, temperature, and concentration, while the operator is responsible for pre-treatment preparation, monitoring during the process, and recording and confirming the results afterward. For evaluating sterilization effectiveness, biological or chemical indicators are commonly used for qualitative or semi-quantitative verification, and the verification results are archived along with the operation records, forming paper-based or simple electronic sterilization records.

[0003] This conventional management and traceability model, centered on the execution of fixed procedures, has gradually revealed some inherent flaws in practice. On the one hand, the success of the sterilization process highly depends on the universality of preset parameters and the strict adherence of operators to procedures. However, actual application scenarios are complex and varied. For example, fluctuations in factors such as the loading status of biosafety cabinets, initial contamination levels, and environmental temperature and humidity can all affect the final sterilization effect under fixed parameters. Existing methods lack the ability to model and analyze the deep causal relationship between these dynamic factors and sterilization effects, resulting in the inability to predict or explain the risk of sterilization failure under non-standard operating conditions, creating a "black box" blind spot in the management process.

[0004] On the other hand, the impact of differences in operator experience and skills on sterilization results is often overlooked. Experienced operators can handle unexpected situations on-site through subtle, unwritten adjustments, thus ensuring sterilization effectiveness. These implicit, experience-based operational skills are difficult to effectively capture and pass on through standard operating procedures. During personnel changes or training, novices can only learn the fixed procedural steps but cannot grasp the implicit knowledge that substantially affects the results. This increases the risk of sterilization failure due to improper operation, making the reliability and transferability of the entire process insufficient. Summary of the Invention

[0005] This invention provides a method and system for intelligent management and traceability of biosafety cabinet disinfection and sterilization, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for intelligent management and traceability of biosafety cabinet disinfection and sterilization, comprising: Multimodal sterilization data streams are collected, and causal decoupling modeling is performed on the multimodal sterilization data streams. A causal knowledge graph is constructed by identifying the causal dependency between changes in sterilization parameters and sterilization effect response. The causal paths in the causal knowledge graph are validated. The failure boundaries of the causal paths under parameter offset or environmental change are tested by introducing adversarial scenario perturbation. Stable causal paths that pass the validation are retained to form a reliable knowledge subgraph. The implicit operation features corresponding to the operator interaction behavior data are extracted from the trusted knowledge subgraph. By comparing the differences in implicit operation features of different operators under the same sterilization task, the distinguishing features of different operation modes are identified. For the sterilization task to be inherited, the necessary causal path set that satisfies the target sterilization effect is traced backward from the causal path in the trusted knowledge subgraph, and the feasibility of the necessary causal path set is screened by combining the distinguishing features to generate an inheritance scheme that includes causal guarantee constraints and implicit operation guidance. During the execution of the inheritance plan, actual sterilization effect feedback data is collected. Deviation analysis is performed between the actual sterilization effect feedback data and the expected effect of the inheritance plan. When the deviation exceeds the tolerance, the causal dependency relationship of the corresponding causal path in the causal knowledge graph is corrected, and adversarial scenario perturbation is triggered to re-verify the corrected causal path.

[0007] The validity of causal paths in the causal knowledge graph is verified by introducing adversarial scenario perturbations to test the failure boundaries of causal paths under parameter offsets or environmental changes. Stable causal paths that pass verification are retained to form a reliable knowledge subgraph, including: An adversarial detection mechanism is constructed for each causal path in the causal knowledge graph. The adversarial detection mechanism backpropagates the deviation gradient from the sterilization effect response node along the causal dependency to the parameter node and the environmental state node, calculates the gradient sensitivity of the parameter node and the environmental state node to the sterilization effect response node, and generates the perturbation direction and perturbation intensity that maximizes the prediction deviation based on the gradient sensitivity. The worst-case perturbation scenario is constructed by injecting directional perturbations into the parameter node and the environmental state node using the perturbation direction and the perturbation intensity. Under the worst-case perturbation scenario, the predicted sterilization effect value is obtained by forward deduction along the causal path and the corresponding actual sterilization effect measurement value is collected. The failure deviation between the two is calculated. When the failure deviation exceeds the preset failure threshold, the gradient sensitivity of the adversarial detection mechanism is used to trace and locate the weak node that caused the failure, record the failure mode formed by the weak node and the perturbation features of the worst perturbation scenario, and mark the causal path as a vulnerable path and remove it. When the failure deviation is lower than the preset failure threshold, the causal path is marked as a stable causal path, and all stable causal paths and their associated nodes are combined to form a trusted knowledge subgraph.

[0008] The gradient sensitivity of the adversarial detection mechanism is used to locate the weak nodes that cause the failure. The weak nodes and the perturbation features of the worst-case perturbation scenario constitute the failure mode. The causal path is marked as a vulnerable path and removed, including: The contribution weights of each node in the causal path to the failure deviation are extracted from the gradient sensitivity calculated by the adversarial detection mechanism. The contribution weights are sorted in descending order and the set of sensitive nodes with the highest contribution weights is identified. For each sensitive node in the set of sensitive nodes, a reverse compensation perturbation is applied under the worst perturbation scenario and the causal path is re-deduced to obtain the predicted sterilization effect value after compensation. Calculate the failure mitigation degree between the compensated predicted sterilization effect value and the original predicted sterilization effect value. When the failure mitigation degree exceeds a preset mitigation threshold, the sensitive node is determined to be a weak node that causes failure. Extract the topological location and node type of the weak node in the causal path, and extract the perturbation direction and perturbation amplitude acting on the weak node from the worst perturbation scenario as perturbation features. Combine the topological location, the node type and the perturbation features to form a failure mode. The failure mode is bound and labeled with the causal path, and the causal path is marked as a vulnerable path and removed from the causal knowledge graph.

[0009] Implicit operational features corresponding to operator interaction behavior data are extracted from the trusted knowledge subgraph. By comparing the differences in implicit operational features among different operators under the same sterilization task, distinguishing features for different operation modes are identified, including: Operator interaction behavior data is extracted from the stable causal path of the trusted knowledge subgraph. Behavioral pattern mining is performed on the operator interaction behavior data to identify the operation timing selection pattern and parameter adjustment response pattern. The operation timing selection pattern and the parameter adjustment response pattern are encoded into implicit operation feature vectors. For the same sterilization task, implicit operation feature vectors of multiple operators are collected to construct a feature comparison space, and feature difference measure between implicit operation feature vectors of different operators is calculated in the feature comparison space. Based on the feature difference measure, the implicit operation feature vectors are clustered to form operation pattern clusters. Inter-cluster separation analysis is performed on the operation pattern clusters to identify the feature dimensions that contribute the most to the clustering results. The implicit operation features corresponding to the feature dimensions are used as distinguishing features to differentiate different operation patterns.

[0010] The set of necessary causal paths that satisfy the target sterilization effect are traced backward along the causal path from the trusted knowledge subgraph, and the feasibility of the necessary causal path set is screened by combining the distinguishing features to generate a succession scheme that includes causal guarantee constraints and implicit operational guidance, including: The target sterilization effect of the sterilization task to be inherited is located as the target sterilization effect node in the trusted knowledge subgraph. The causal dependency relationship along the stable causal path is traced backward from the target sterilization effect node. During the reverse tracing process, the causal necessity is verified for each intermediate node. The intermediate nodes that pass the verification are retained to form a set of causal backbone nodes. The causal dependency relationship connecting each node in the set of causal backbone nodes is connected to form a set of necessary causal paths. For each necessary causal path in the set of necessary causal paths, the operation node is matched with the operation mode recognition boundary corresponding to the operation node and the implicit operation feature carried by the operation node. When the matching degree reaches the preset matching threshold, the necessary causal path is determined to be executable at the operation level. Necessary causal paths with operational feasibility are selected to form an executable causal path set. The causal dependencies between the causal backbone nodes are parsed from the executable causal path set, and the causal dependencies are condensed into causal guarantee constraints. The implicit operation characteristics of the operation nodes in the executable causal path set are parsed and transformed into implicit operation guidelines. The causal guarantee constraints and the implicit operation guidelines are combined to generate a succession scheme.

[0011] The causal dependencies of the target sterilization effect node are traced backward along a stable causal path. During the backward tracing process, the causal necessity of each intermediate node is verified. The intermediate nodes that pass the verification constitute the causal backbone node set, including: The process involves tracing back from the target sterilization effect node along a stable causal path towards the operation initiation direction, identifying the preceding intermediate nodes that have a direct causal dependency with the target sterilization effect node during the tracing back process. For each identified precursor intermediate node, a causal verification scenario is constructed. The causal verification scenario obtains the predicted sterilization effect value after shielding the causal transmission of the precursor intermediate node in the stable causal path and extrapolating to the target sterilization effect node. Calculate the effect deviation between the predicted sterilization effect value after shielding and the predicted sterilization effect value without shielding. When the effect deviation exceeds a preset deviation threshold, it is determined that the causal contribution of the precursor intermediate node is indispensable and passes the causal necessity verification. The predecessor intermediate node that passes the causal necessity verification is added to the causal backbone node set. The predecessor intermediate node in the causal backbone node set is used as the new tracing starting point to continue tracing back to the operation starting direction and repeating the causal necessity verification process until the tracing back to the operation starting node forms a complete causal backbone node set.

[0012] The actual sterilization effect feedback data is compared with the expected effect of the inheritance plan. When the deviation exceeds the tolerance, the causal dependency relationship of the corresponding causal path in the causal knowledge graph is corrected, and an adversarial scenario perturbation is triggered to re-verify the corrected causal path, including: The actual sterilization effect feedback data is compared with the expected effect of the inheritance plan to obtain the effect deviation amount. When the effect deviation amount exceeds the preset tolerance, the deviation causal path corresponding to the effect deviation amount is located in the causal knowledge graph, and the deviation causal dependency relationship that leads to the effect deviation amount in the deviation causal path is identified. The biased causal dependencies are corrected to reduce the effect bias. The corrected causal dependencies replace the biased causal dependencies in the biased causal path to form a corrected causal path. The causal dependencies of the corresponding causal path in the causal knowledge graph are updated based on the corrected causal path. Adversarial scenario perturbations are triggered for the modified causal path in the causal knowledge graph. The adversarial scenario perturbations verify the causal stability of the modified causal path under perturbation conditions by introducing abnormal operating conditions into the modified causal path and extrapolating along the modified causal path to the target sterilization effect.

[0013] A second aspect of this invention provides an intelligent management and traceability system for the disinfection and sterilization of biosafety cabinets, comprising: The causal modeling unit is used to collect multimodal sterilization data streams, perform causal decoupling modeling on the multimodal sterilization data streams, construct a causal knowledge graph by identifying the causal dependency between sterilization parameter changes and sterilization effect response, and verify the effectiveness of causal paths in the causal knowledge graph. By introducing adversarial scenario perturbations, the failure boundaries of causal paths under parameter offsets or environmental changes are tested, and the verified stable causal paths are retained to form a reliable knowledge subgraph. The feature recognition unit is used to extract implicit operation features corresponding to operator interaction behavior data from the trusted knowledge subgraph, and to identify distinguishing features of different operation modes by comparing the differences in implicit operation features of different operators under the same sterilization task. The scheme generation unit is used to trace the necessary causal path set that satisfies the target sterilization effect from the causal path in reverse from the trusted knowledge subgraph for the sterilization task to be inherited, and to perform operation feasibility screening on the necessary causal path set in combination with the distinguishing features to generate an inheritance scheme that includes causal guarantee constraints and implicit operation guidance. The feedback correction unit is used to collect actual sterilization effect feedback data during the execution of the inheritance scheme, analyze the deviation between the actual sterilization effect feedback data and the expected effect of the inheritance scheme, and correct the causal dependency relationship of the corresponding causal path in the causal knowledge graph when the deviation exceeds the tolerance, and trigger adversarial scenario perturbation to re-verify the corrected causal path.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] This method enables intelligent management and precise traceability of the biosafety cabinet sterilization process. By constructing a causal knowledge graph through causal decoupling modeling and introducing adversarial scenario perturbations for effectiveness verification, the scientific validity and robustness of the extracted causal relationships between sterilization parameters and effects are ensured. This process effectively eliminates spurious associations in the data, identifies core causal paths that remain stable under different environmental perturbations, and thus forms a reliable core of sterilization knowledge, providing a solid theoretical basis for subsequent decision-making and knowledge transfer.

[0017] By extracting and comparing implicit operational features from trusted knowledge subgraphs, key differences in human operational patterns affecting sterilization effectiveness can be accurately identified. This feature recognition based on causal relationships goes beyond traditional recording of explicit operational steps. It can capture implicit knowledge such as operational techniques and habits that are difficult to quantify, providing data support for the standardization and optimization of operational skills and helping to reduce fluctuations in sterilization effectiveness caused by differences in personnel operations.

[0018] For specific sterilization tasks, by tracing the necessary causal paths backward and combining this with operational feasibility screening, a scientifically sound and practical transfer program can be generated. This program not only clarifies the causal logic chain that must be followed to achieve the target sterilization effect, but also incorporates verified implicit operational guidelines. This transforms experience transfer from relying on personal oral traditions into structured, verifiable, and executable operational procedures, significantly improving the accuracy and efficiency of key sterilization technology transfer. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the intelligent management and traceability method for disinfection and sterilization of biosafety cabinets according to an embodiment of the present invention. Figure 2 This is a flowchart of the causal path correction and verification based on deviation analysis in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] Figure 1 This is a flowchart illustrating the intelligent management and traceability method for disinfection and sterilization of biosafety cabinets according to an embodiment of the present invention. Figure 1 As shown, the intelligent management and traceability method for biosafety cabinet disinfection and sterilization includes: Multimodal sterilization data streams are collected, and causal decoupling modeling is performed on the multimodal sterilization data streams. A causal knowledge graph is constructed by identifying the causal dependency between changes in sterilization parameters and sterilization effect response. The causal paths in the causal knowledge graph are validated. The failure boundaries of the causal paths under parameter offset or environmental change are tested by introducing adversarial scenario perturbation. Stable causal paths that pass the validation are retained to form a reliable knowledge subgraph. The implicit operation features corresponding to the operator interaction behavior data are extracted from the trusted knowledge subgraph. By comparing the differences in implicit operation features of different operators under the same sterilization task, the distinguishing features of different operation modes are identified. For the sterilization task to be inherited, the necessary causal path set that satisfies the target sterilization effect is traced backward from the causal path in the trusted knowledge subgraph, and the feasibility of the necessary causal path set is screened by combining the distinguishing features to generate an inheritance scheme that includes causal guarantee constraints and implicit operation guidance. During the execution of the inheritance plan, actual sterilization effect feedback data is collected. Deviation analysis is performed between the actual sterilization effect feedback data and the expected effect of the inheritance plan. When the deviation exceeds the tolerance, the causal dependency relationship of the corresponding causal path in the causal knowledge graph is corrected, and adversarial scenario perturbation is triggered to re-verify the corrected causal path.

[0023] In one optional implementation, the causal paths in the causal knowledge graph are validated by introducing adversarial scenario perturbations to test the failure boundaries of the causal paths under parameter offsets or environmental changes. The validated stable causal paths are retained to form a reliable knowledge subgraph, including: An adversarial detection mechanism is constructed for each causal path in the causal knowledge graph. The adversarial detection mechanism backpropagates the deviation gradient from the sterilization effect response node along the causal dependency to the parameter node and the environmental state node, calculates the gradient sensitivity of the parameter node and the environmental state node to the sterilization effect response node, and generates the perturbation direction and perturbation intensity that maximizes the prediction deviation based on the gradient sensitivity. The worst-case perturbation scenario is constructed by injecting directional perturbations into the parameter node and the environmental state node using the perturbation direction and the perturbation intensity. Under the worst-case perturbation scenario, the predicted sterilization effect value is obtained by forward deduction along the causal path and the corresponding actual sterilization effect measurement value is collected. The failure deviation between the two is calculated. When the failure deviation exceeds the preset failure threshold, the gradient sensitivity of the adversarial detection mechanism is used to trace and locate the weak node that caused the failure, record the failure mode formed by the weak node and the perturbation features of the worst perturbation scenario, and mark the causal path as a vulnerable path and remove it. When the failure deviation is lower than the preset failure threshold, the causal path is marked as a stable causal path, and all stable causal paths and their associated nodes are combined to form a trusted knowledge subgraph.

[0024] After the causal knowledge graph is constructed, the causal paths within it need to be validated to select reliable causal dependencies. The validation process first constructs an adversarial detection mechanism for each causal path in the graph. The core of this mechanism is to proactively generate the most unfavorable perturbation conditions to test the robustness boundaries of the causal paths. Specifically, a given causal path contains multiple nodes, including sterilization parameter nodes (such as hydrogen peroxide concentration nodes, vaporization temperature nodes), environmental state nodes (such as environmental humidity nodes, cabinet airtightness nodes), and sterilization effect response nodes (such as microbial logarithmic reduction value nodes). The adversarial detection mechanism starts from the sterilization effect response node, assuming a unit deviation between the actual measured value and the predicted value, and propagates this deviation to upstream nodes along the reverse dependency of the causal path. During backpropagation, each upstream node receives the deviation signal from downstream nodes and allocates the deviation according to the strength of the causal dependency between that node and the downstream node. For parameter nodes or environmental state nodes, the cumulative deviation received reflects the degree of influence of that node on the final sterilization effect response, i.e., gradient sensitivity. For example, if the cumulative deviation received by the vaporization temperature node is 0.35 units, while the cumulative deviation received by the ambient humidity node is 0.12 units, it indicates that the vaporization temperature has a significantly greater impact on the sterilization effect than the ambient humidity.

[0025] After acquiring the gradient sensitivities of all parameter nodes and environmental state nodes, the adversarial detection mechanism generates a perturbation strategy based on these sensitivity values. The principle for determining the perturbation direction is to maximize the prediction bias. Specifically, if a parameter node is positively correlated with the sterilization effect response node, a negative perturbation is injected into that parameter node; if they are negatively correlated, a positive perturbation is injected. The perturbation intensity is proportional to the gradient sensitivity; nodes with higher gradient sensitivity are subject to stronger perturbation amplitudes. For example, when the normal operating range of the vaporization temperature node is 60℃ to 80℃ and its gradient sensitivity is 0.35 units, the perturbation intensity can be calculated as the product of this sensitivity value and a preset perturbation coefficient. Assuming the perturbation coefficient is set to 15, the perturbation amplitude is 5.25℃. If this node is positively correlated with the sterilization effect, a negative perturbation of -5.25℃ is applied based on its current operating point. For environmental state nodes, the perturbation amplitude is also calculated based on their gradient sensitivity and perturbation coefficient. It should be noted that perturbation injection must be performed within the physical feasible range of each node. If the calculated perturbation value exceeds the limit boundary of the node, the perturbation value should be limited to the boundary value.

[0026] After generating the perturbation strategies, these perturbations are simultaneously injected into the corresponding parameter nodes and environmental state nodes to construct the worst-case perturbation scenario. Under this scenario, forward inference is performed along the causal path to obtain the predicted sterilization effect value. The forward inference process follows causal dependencies, starting from the perturbed initial node and calculating the state values ​​of subsequent nodes layer by layer, ultimately obtaining the predicted value of the sterilization effect response node. For example, when the vaporization temperature is perturbed to 74.75℃ and the environmental humidity is perturbed to 58%, based on the dependency relationship between temperature and humidity on the uniformity of hydrogen peroxide distribution in the causal path, the uniformity of distribution can be calculated to decrease to 86% of its original value, thus deriving a predicted final microbial logarithmic reduction value of 5.2. Simultaneously, the sterilization process is actually run under the same perturbation scenario, and the actual sterilization effect measurement value under this scenario is collected through a biological indicator test, assuming the measured microbial logarithmic reduction value is 4.7. The failure deviation between the predicted value and the actual measured value is calculated; here, the deviation is 0.5 log units.

[0027] The calculated failure deviation is compared with a preset failure threshold. The failure threshold is set by comprehensively considering the safety boundary of sterilization effect and the measurement error range, and is usually set to 0.3 to 0.5 log units. If the failure deviation exceeds this threshold, it indicates that the causal path has failed significantly under adversarial perturbation, and its causal dependency is not stable enough. At this time, the gradient sensitivity calculated in the adversarial detection mechanism is used to trace and locate the weak node. The specific method is to distribute the failure deviation to each parameter node and environmental state node according to the gradient sensitivity ratio, and the node that bears the largest distribution deviation is the weak node. For example, if the failure deviation of 0.5 is distributed to 0.32 at the vaporization temperature node and 0.18 at the environmental humidity node, then the vaporization temperature node is identified as the main weak node. The identifier of the weak node, its position in the causal path, the corresponding perturbation direction is negative, the perturbation intensity is 5.25℃, and the combination of environmental state parameters when the failure is triggered are recorded. This information together constitutes the failure mode of the causal path. The failure mode is stored in the failure knowledge base, and the causal path containing the weak node is marked as a vulnerable path and removed from the causal knowledge graph, so that it will no longer be used as the basis for generating subsequent inheritance schemes.

[0028] If the failure deviation is below a preset failure threshold, for example, a deviation of 0.25 log units, it indicates that the causal path can still maintain acceptable prediction accuracy under the worst-case perturbation scenario, and the causal dependency has strong robustness. This path is then marked as a stable causal path. The above verification process is repeated for all causal paths in the causal knowledge graph, with each path undergoing adversarial perturbation testing and failure deviation evaluation. After verification, all marked stable causal paths and their associated nodes are extracted to form a trusted knowledge subgraph. This subgraph retains reliable causal dependencies that have undergone adversarial verification and eliminates vulnerable paths that are prone to failure under parameter shifts or environmental changes, ensuring that subsequent inheritance schemes generated based on this subgraph have higher execution reliability. Each stable causal path in the trusted knowledge subgraph includes perturbation boundary information when it passed the adversarial test, including the maximum perturbation amplitude that each node can withstand and the corresponding failure deviation value. This boundary information provides a quantitative basis for security margin assessment in subsequent practical applications.

[0029] In constructing a trusted knowledge subgraph, the gradient sensitivity calculation for adversarial detection mechanisms needs to consider the nonlinear interaction effects between nodes in the causal path. When a sterilization effect response is jointly influenced by multiple parameter nodes and environmental state nodes, and these influences are coupled, the backpropagation of gradient sensitivity needs to be decomposed using the chain rule. For example, hydrogen peroxide concentration and vaporization temperature jointly affect the vaporization rate, which in turn affects the uniformity of concentration distribution within the cabinet, ultimately impacting the sterilization effect. In this scenario, backpropagation first calculates the gradient of sterilization effect with respect to uniformity of concentration distribution, then calculates the gradient of uniformity of concentration distribution with respect to vaporization rate, and finally calculates the gradients of vaporization rate with respect to hydrogen peroxide concentration and vaporization temperature, obtaining the combined gradient sensitivity of these two parameter nodes through chain multiplication. This approach ensures that in complex causal networks, the gradient sensitivity of each node accurately reflects its true impact on the final effect, making the generated perturbation strategy more targeted and effectively exposing weaknesses in the causal path.

[0030] In one optional implementation, the gradient sensitivity of the adversarial detection mechanism is used to locate the weak node causing the failure, and the failure mode is recorded by combining the weak node with the perturbation features of the worst-case perturbation scenario. The causal path is then marked as a vulnerable path and removed. The contribution weights of each node in the causal path to the failure deviation are extracted from the gradient sensitivity calculated by the adversarial detection mechanism. The contribution weights are sorted in descending order and the set of sensitive nodes with the highest contribution weights is identified. For each sensitive node in the set of sensitive nodes, a reverse compensation perturbation is applied under the worst perturbation scenario and the causal path is re-deduced to obtain the predicted sterilization effect value after compensation. Calculate the failure mitigation degree between the compensated predicted sterilization effect value and the original predicted sterilization effect value. When the failure mitigation degree exceeds a preset mitigation threshold, the sensitive node is determined to be a weak node that causes failure. Extract the topological location and node type of the weak node in the causal path, and extract the perturbation direction and perturbation amplitude acting on the weak node from the worst perturbation scenario as perturbation features. Combine the topological location, the node type and the perturbation features to form a failure mode. The failure mode is bound and labeled with the causal path, and the causal path is marked as a vulnerable path and removed from the causal knowledge graph.

[0031] When performing adversarial verification on causal paths, in-depth analysis of failed paths is necessary to ensure the accuracy of subsequent screening. When a causal path deviates beyond the tolerance under the perturbation of the adversarial detection mechanism, the entire path cannot be simply eliminated; instead, the weak link in the path that truly caused the failure should be accurately identified. By analyzing the gradient sensitivity distribution calculated during the adversarial detection process, the influence of each node in the causal path on the failure deviation can be quantified.

[0032] Gradient sensitivity characterizes the fluctuation range of the overall sterilization effect prediction caused by a small change in the parameter of a certain node. For [aspects including...] The causal path of the nth node is defined. The gradient sensitivity of each node is Its calculation is based on the predicted effect value. Parameters of this node The partial derivatives of . The gradient sensitivity of all nodes. Nodes are sorted in descending order of their numerical values ​​to form a sensitivity sequence. Based on experience, nodes with the top 30% contribution weights are considered part of the sensitive node set. For example, in a causal path with 12 nodes, the top four nodes by contribution weight will be included in the sensitive node set for further analysis.

[0033] For each sensitive node in the sensitive node set, a reverse compensation perturbation is applied under the worst-case perturbation scenario to verify whether the node is a true weak node. The worst-case perturbation scenario is the combination of perturbations that causes the largest prediction deviation of sterilization effect in the adversarial detection mechanism, and the specific perturbation direction and magnitude of each node are recorded. The reverse compensation perturbation refers to applying a perturbation opposite to the worst-case perturbation direction to the sensitive node under test while keeping the perturbations of other nodes unchanged, and its magnitude is calibrated according to the gradient sensitivity of the node. For example, in the causal path of formaldehyde fumigation sterilization of biosafety cabinets, if the circulation wind speed node is subjected to a 15% increase in perturbation under the worst-case perturbation scenario, then the reverse compensation perturbation is to reduce the perturbation amount by the corresponding proportion based on the original wind speed.

[0034] After applying a reverse compensation perturbation, the predicted sterilization effect value is recalculated along the causal path. The deduction process follows the established causal dependencies in the causal graph, propagating parameter influences from the starting node to downstream nodes step by step, and finally calculating the predicted microbial inactivation rate of different areas within the cabinet. This predicted value reflects the overall response change of the system after compensating for specific sensitive nodes.

[0035] By comparing the original predicted sterilization effect value with the compensated predicted sterilization effect value, the failure mitigation degree is calculated to determine the actual contribution of the sensitive node to the failure. The failure mitigation degree is defined as the proportion of improvement in the compensated deviation relative to the original deviation, denoted as . The calculation formula is: ,in This represents the absolute value of the deviation between the original predicted result and the expected value. This represents the absolute value of the deviation between the compensated predicted result and the expected value. When If the mitigation threshold is exceeded, for example, 0.4, it indicates that reverse compensation for that node can significantly improve the failure situation, thus determining that the sensitive node is indeed the weak node causing the failure. If the failure mitigation is small, it means that although the node is highly sensitive, it is not the root cause of the failure, but only exhibits a large gradient change due to the influence of other nodes.

[0036] After identifying weak nodes, detailed feature information of these nodes needs to be extracted to construct failure modes. Topological location features describe the hierarchy and connectivity of the weak node within the causal path. For example, a weak node located at layer 3 of the causal path, directly downstream of the formaldehyde concentration node and upstream of the humidity distribution node, would be recorded as "L3 - downstream of P7 - upstream of P12". Node type features identify the category of physical quantity represented by the weak node, including environmental parameters, equipment status, and control commands. For formaldehyde fumigation sterilization, weak nodes belong to airflow organization parameters, which directly affect the uniformity of formaldehyde distribution.

[0037] The specific disturbance characteristics acting on the weak point are extracted from the worst-case disturbance scenario. The direction of the disturbance describes the trend of parameter change, such as temperature increase, wind speed decrease, or humidity fluctuation. The magnitude of the disturbance quantifies the degree to which the parameter deviates from the standard value, usually expressed as a percentage or absolute value. In a real-world case, if a weak point is the static pressure difference parameter inside the cabinet, the disturbance applied in the worst-case scenario is a decrease in static pressure difference of 8 Pa. This disturbance increases the risk of formaldehyde leakage to the outside, resulting in insufficient effective concentration inside the cabinet. The direction of the disturbance is recorded as "pressure difference decrease," and the magnitude of the disturbance is recorded as "8 Pa." This information together constitutes the disturbance characteristics.

[0038] The topological location, node type, and disturbance characteristics are combined to form a complete failure mode description. Failure modes are stored in structured data format, including fields such as node identifier, level number, node type label, disturbance direction code, and disturbance amplitude value. For example, a failure mode might be described as "Node ID-N23, Level-L3, Type-Airflow Parameter, Disturbance Direction-Wind Speed ​​Decrease, Disturbance Amplitude-12%, Mitigation-0.53". This structured description facilitates subsequent statistical analysis and pattern recognition of failure modes, enabling the identification of common vulnerabilities across different sterilization tasks.

[0039] After constructing the failure modes, each failure mode is bound and labeled with its corresponding causal path. The labeling information is attached to the metadata of the causal path, recording the scenario conditions in which the failure occurred, the intensity of the failure, and the location of the weak nodes. Through the labeling mechanism, each causal path carries its own stability evaluation information, providing a basis for subsequent knowledge graph selection and application.

[0040] Causal paths with identified weak nodes are marked as fragile paths and removed from the causal knowledge graph. This removal is not a physical deletion of the path data, but rather a setting to "unavailable" for the path's credibility, preventing its exclusion during subsequent construction of trusted knowledge subgraphs. After removing fragile paths, the paths remaining in the causal knowledge graph are all trusted paths that can maintain stable responses even in adversarial scenarios; these paths form the basis of the trusted knowledge subgraphs.

[0041] In practical applications, the manifestations of weak points differ depending on the type of disinfection and sterilization method. For hydrogen peroxide vaporization sterilization, weak points are concentrated in equipment control parameters such as vaporizer heating power and spray angle; for ultraviolet sterilization, weak points are factors such as lamp aging and irradiation distance. Through cluster analysis of failure modes across multiple batches, typical vulnerability characteristics of specific sterilization methods can be extracted, providing guidance for equipment optimization and operational procedure improvement.

[0042] Meanwhile, the accumulation of the failure mode library provides data support for predictive maintenance. When planning a new sterilization task, historical failure cases similar to the current task conditions can be retrieved from the failure mode library to identify potential risks and take preventive measures in advance. This failure prediction mechanism based on causal analysis can more accurately locate the root cause of problems and avoid false alarms and missed alarms compared to traditional alarm methods based on statistical thresholds.

[0043] In one optional implementation, implicit operational features corresponding to operator interaction behavior data are extracted from the trusted knowledge subgraph. By comparing the differences in implicit operational features among different operators under the same sterilization task, distinguishing features for different operational modes are identified, including: Operator interaction behavior data is extracted from the stable causal path of the trusted knowledge subgraph. Behavioral pattern mining is performed on the operator interaction behavior data to identify the operation timing selection pattern and parameter adjustment response pattern. The operation timing selection pattern and the parameter adjustment response pattern are encoded into implicit operation feature vectors. For the same sterilization task, implicit operation feature vectors of multiple operators are collected to construct a feature comparison space, and feature difference measure between implicit operation feature vectors of different operators is calculated in the feature comparison space. Based on the feature difference measure, the implicit operation feature vectors are clustered to form operation pattern clusters. Inter-cluster separation analysis is performed on the operation pattern clusters to identify the feature dimensions that contribute the most to the clustering results. The implicit operation features corresponding to the feature dimensions are used as distinguishing features to differentiate different operation patterns.

[0044] In intelligent management systems for biosafety cabinet sterilization, the implicit operational characteristics of operators contain rich experiential knowledge, which are often difficult to fully describe using explicit rules. To effectively extract and identify these implicit operational characteristics, it is necessary to obtain operator interaction data from stable causal paths within a trusted knowledge subgraph. This interaction data includes multi-dimensional information such as records of every parameter adjustment made by the operator during the sterilization task, sequences of equipment control commands, timestamps of interface interactions, and operation response delays. For a complete sterilization task, the operator interaction data includes elements such as the operation time, operation type, parameter change magnitude, operation duration, and equipment status before and after the operation.

[0045] After extracting operator interaction data, behavioral pattern mining is performed to identify operation timing selection patterns and parameter adjustment response patterns. Operation timing selection patterns reflect the characteristics of when operators choose to intervene at different stages of the sterilization process. For example, experienced operators tend to adjust airflow parameters in advance when the temperature rises to 85% of the set value, while novice operators wait until the temperature fully reaches the target before making adjustments. Parameter adjustment response patterns describe the operator's response strategy to changes in equipment status, including the magnitude, frequency, and sequence of parameter adjustments. By aggregating and statistically analyzing operation timing using time series analysis methods, the operator's habitual intervention time windows can be identified; and by analyzing parameter change trajectories, typical parameter adjustment patterns can be extracted, such as linear gradual adjustment, step adjustment, or oscillatory fine-tuning, among other strategies.

[0046] Encoding the timing selection pattern and parameter adjustment response pattern into implicit operational feature vectors requires feature engineering. For the timing selection pattern, timing feature dimensions can be constructed, including statistical features such as average intervention time, intervention time variance, early intervention frequency, and late intervention frequency. For the parameter adjustment response pattern, response feature dimensions can be constructed, including average amplitude of a single adjustment, adjustment frequency, parameter stabilization time, and adjustment fluctuation coefficient. Assuming the timing selection pattern is extracted... 3D features, parameter adjustment response mode extraction If a feature is defined in dimension 1, then the latent operation feature vector can be represented as: dimensional vector subscript Indicates the first Each operator. In the actual coding process, it is necessary to normalize features of different dimensions to ensure that features of each dimension are comparable in the feature space.

[0047] For the same sterilization task, implicit operational feature vectors from multiple operators are collected to construct a feature comparison space. Assume that for a given sterilization task, a total of [number missing] implicit operational feature vectors are collected. The interaction behavior data of each operator can then be used to construct a system containing... A feature comparison space is defined for implicit operation feature vectors. In this space, the feature differences between the implicit operation feature vectors of different operators are measured. Commonly used metrics include Euclidean distance, cosine similarity, or Mahalanobis distance. Euclidean distance intuitively reflects the geometric distance between feature vectors in multidimensional space; cosine similarity focuses more on the directional differences of feature vectors than their magnitude differences; and Mahalanobis distance considers the correlation between feature dimensions. The appropriate metric should be selected based on the actual application requirements. For example, cosine similarity is chosen when it is necessary to emphasize the overall similarity of operation patterns, while Euclidean distance is chosen when precise quantification of operational differences is required.

[0048] Latent operational feature vectors are clustered based on feature difference metrics to form operational pattern clusters. Clustering algorithms can employ hierarchical clustering, K-means clustering, or density clustering. Hierarchical clustering constructs a hierarchical structure of operational patterns, facilitating the identification of groups of operators with different experience levels; K-means clustering is suitable for scenarios where the approximate number of operational pattern categories is known; density clustering can discover operational pattern clusters of arbitrary shapes and automatically identify abnormal operational patterns. During the clustering process, the optimal number of clusters needs to be determined based on clustering quality evaluation indicators such as the silhouette coefficient and Davies-Bouldin index. The clustering results divide operators into several operational pattern clusters, each representing a typical operational pattern. For example, a cautious operational pattern tends to adjust parameters frequently with small amplitudes, an aggressive operational pattern tends to adjust parameters rapidly with large amplitudes, and an experienced operational pattern exhibits characteristics of early prediction and precise intervention.

[0049] Inter-cluster separation analysis is performed on the operational pattern clusters to identify the feature dimensions that contribute most to the clustering results. Inter-cluster separation reflects the degree of distinction between different clusters in the feature space; higher separation indicates better clustering. By calculating the center vector of each cluster, the distribution of differences among different clusters across various feature dimensions can be analyzed. For the _____ For each feature dimension, the variance ratio of all clusters along that dimension is calculated. Dimensions with larger variance ratios indicate that they play a crucial role in distinguishing different operational modes. Specifically, analysis of variance (ANOVA) can be used to calculate the ratio of inter-cluster variance to intra-cluster variance; a larger ratio indicates a greater contribution of the feature dimension to clustering. Furthermore, feature importance scoring methods, such as the feature importance score in random forests or the contribution rate in principal component analysis, can be used to verify the discriminative power of feature dimensions from multiple perspectives.

[0050] The implicit operational features corresponding to the feature dimensions that contribute the most to the clustering results are used as the distinguishing features to differentiate different operational modes. Assuming that inter-cluster separation analysis identifies... These key feature dimensions, along with their corresponding implicit operational features, constitute a set of discriminative features. Discriminative features possess the following characteristics: they exhibit significant differences between clusters of different operational modes, remain relatively stable within the same operational mode cluster, and can effectively predict the operational mode category to which an operator belongs. In practical applications, discriminative features can be used to quickly identify the operational mode affiliation of new operators, providing them with targeted operational guidance and training suggestions. For example, if an operator's discriminative features indicate a cautious operational mode, the tolerance range for parameter adjustments in the transfer program can be appropriately relaxed to avoid excessive intervention leading to a lack of operational confidence; conversely, if an operator belongs to an aggressive operational mode, the boundary constraints for parameter adjustments need to be strengthened in the transfer program to prevent excessive adjustments from causing fluctuations in sterilization effectiveness.

[0051] In practice, the extraction of implicit operational features and the identification of distinguishing features require continuous data accumulation and dynamic updates. As more operators participate in sterilization tasks, the feature comparison space expands, and operational pattern clusters become more refined or reorganized. By periodically re-performing cluster analysis and inter-cluster separation assessments, the set of distinguishing features can be updated in a timely manner, ensuring that it always reflects the typical operational pattern distribution of the current operator group. Furthermore, an operational pattern evolution tracking mechanism can be established to record changes in the implicit operational features of the same operator at different times, identifying the feature evolution trajectory of operators transitioning from novices to skilled operators, and providing a quantitative assessment basis for operational skills training.

[0052] The implicit operational features and distinguishing features extracted using the methods described above lay the foundation for generating subsequent transfer programs. When generating transfer programs for specific sterilization tasks, causal paths and operational guidance strategies that match the current operational patterns of the operators to be trained can be selected, improving the acceptability and success rate of the transfer program. Simultaneously, distinguishing features can also serve as reference indicators for evaluating transfer effectiveness. By comparing changes in the distinguishing features of operators before and after training, the actual effect of the transfer program on improving operational capabilities can be quantitatively assessed.

[0053] In one optional implementation, the necessary causal path set for satisfying the target sterilization effect is traced backward from the trusted knowledge subgraph along the causal path, and the operational feasibility of the necessary causal path set is screened in conjunction with the distinguishing features to generate a succession scheme that includes causal guarantee constraints and implicit operational guidance, including: The target sterilization effect of the sterilization task to be inherited is located as the target sterilization effect node in the trusted knowledge subgraph. The causal dependency relationship along the stable causal path is traced backward from the target sterilization effect node. During the reverse tracing process, the causal necessity is verified for each intermediate node. The intermediate nodes that pass the verification are retained to form a set of causal backbone nodes. The causal dependency relationship connecting each node in the set of causal backbone nodes is connected to form a set of necessary causal paths. For each necessary causal path in the set of necessary causal paths, the operation node is matched with the operation mode recognition boundary corresponding to the operation node and the implicit operation feature carried by the operation node. When the matching degree reaches the preset matching threshold, the necessary causal path is determined to be executable at the operation level. Necessary causal paths with operational feasibility are selected to form an executable causal path set. The causal dependencies between the causal backbone nodes are parsed from the executable causal path set, and the causal dependencies are condensed into causal guarantee constraints. The implicit operation characteristics of the operation nodes in the executable causal path set are parsed and transformed into implicit operation guidelines. The causal guarantee constraints and the implicit operation guidelines are combined to generate a succession scheme.

[0054] After obtaining a trusted knowledge subgraph that has undergone adversarial verification, for the specific sterilization task to be inherited, it is necessary to extract the core causal chain that guarantees the achievement of the target sterilization effect from this subgraph. The sterilization task to be inherited usually contains clear sterilization effect indicator requirements. For example, for the formaldehyde fumigation sterilization task of a Class A biosafety cabinet, the target sterilization effect is defined as a logarithmic kill value of more than 6 for Bacillus stearothermophilus spores on all surfaces inside the cabinet, while the residual formaldehyde concentration is less than 0.5 mg / m³ after ventilation. This composite target is mapped to the node space of the trusted knowledge subgraph, and the corresponding target sterilization effect node is located as the starting point for reverse tracing.

[0055] Starting from the target node, the process traces backward along a verified stable causal path. The causal dependencies in the trusted knowledge subgraph connect the nodes via directed edges, with the direction of the edges indicating the path from cause to effect. Reverse tracing involves moving in the opposite direction of the edges to find all predecessor nodes that point to the target effect node. This process encounters multiple intermediate nodes at different levels. For example, tracing back from the effect node "spore kill log value 6" leads to the intermediate effect node "formaldehyde concentration in the cabinet maintained above 8000 mg / m³ for 30 minutes," and further back to the parameter node "formaldehyde evaporation rate stabilized at 150 ml / min," and then to the operational parameter node "heating plate temperature controlled between 65 and 70 degrees Celsius."

[0056] However, not all intermediate nodes that can be traced back are causally necessary for the target effect. Some nodes are only sufficient conditions, not necessary conditions, or can be replaced by other causal paths under certain combinations of conditions. Therefore, causal necessity verification is performed on each intermediate node. The verification method adopts a counterfactual inference mechanism, simulating the removal of the intermediate node or setting its parameters to an unsatisfied state in the knowledge graph, and then evaluating whether the target effect node can still be achieved through other causal paths. If there is no alternative path to guarantee the achievement of the target effect after removing the node, the node is determined to be causally necessary and included in the causal backbone node set. Taking formaldehyde fumigation as an example, if the operation node "continuous operation of the internal circulating fan" is removed, even if other parameters meet the requirements, formaldehyde gas cannot be evenly distributed in the cabinet, resulting in some dead areas not reaching the effective sterilization concentration. In this case, the node is identified as a causal backbone node. Conversely, if the operation "pre-irradiation of ultraviolet lamp for 30 minutes" is removed in the formaldehyde fumigation process, as long as the formaldehyde concentration and action time are sufficient, the target sterilization effect can still be achieved, and the node is not included in the backbone set.

[0057] The nodes in the causal backbone node set are linked together according to the temporal and logical order of causal dependencies to form a complete causal chain from initial operating parameters to the target sterilization effect. A sterilization task has multiple such causal chains, which constitute a set of necessary causal paths. For example, for a hydrogen peroxide aerosol sterilization task, there are two different necessary causal paths: a "low concentration long-term action path" and a "high concentration short-time pulse path." Both can achieve the same spore-killing effect, but the combinations of parameter nodes and operating nodes involved are different.

[0058] After obtaining the set of necessary causal paths, it is necessary to evaluate the feasibility of these paths at the practical operational level. Each necessary causal path contains several operational nodes, which record specific manual or equipment operation behaviors. The implicit operational features carried by these operational nodes are extracted from the trusted knowledge subgraph. These features include operational details that are difficult to describe with explicit procedures, such as the gradual adjustment rate of the operator when adjusting the temperature of the heating plate, the interval between batches when adding formaldehyde, and the reading frequency when observing changes in humidity in the cabinet.

[0059] The extracted implicit operational features are compared and analyzed with the previously identified distinguishing features. The distinguishing features define the recognition boundaries of different operational modes. For example, the feature boundary of the "rapid target achievement" operational mode is a temperature adjustment rate greater than 3 degrees Celsius per minute and a total adjustment time of less than 5 minutes, while the boundary of the "stable control" operational mode is an adjustment rate less than 1.5 degrees Celsius per minute but a temperature fluctuation range controlled within ±0.5 degrees Celsius. The fit between the implicit features of the operational nodes and the boundaries of each operational mode is calculated. The fit can be calculated by projecting the feature vector onto the hyperplane defined by the mode boundary, or by using a fuzzy membership function to evaluate the degree to which the features fall within the boundary range.

[0060] When the implicit operational characteristics of an operation node match the boundary of a verified and effective operation mode to a preset matching threshold (e.g., a matching score exceeding 0.75), it indicates that the required operation method of the operation node can be achieved within the capabilities of the target operator. If the implicit characteristics required by the operation node match the boundary of all known operation modes below the threshold, it means that the operation places demands on the operator beyond their existing experience, posing a risk of failure in actual execution. A full-path operational feasibility check is performed on each path in the necessary causal path set. Only when the matching degree of all operation nodes on the path meets the threshold requirement is the path considered operationally feasible and included in the executable causal path set.

[0061] Taking chlorine dioxide gas sterilization as an example, a necessary causal path requires operators to record concentration readings every 2 minutes during the gas concentration rise phase and fine-tune the generator power based on these readings. The implicit characteristics of this operational node include a high-frequency monitoring rhythm and rapid decision-making response. If the target user is an operator with extensive chemical experimental background, and their operational mode feature library contains a "fine chemical reaction monitoring mode," whose boundary characteristics include a monitoring frequency of up to once per minute and a parameter adjustment response time of less than 30 seconds, then this operational node has a high degree of compatibility with this mode, and the path is feasible. Conversely, if the target user is an equipment maintenance engineer, whose operational mode tends to be "automatic operation after equipment parameter presets," lacking the experience characteristic of high-frequency manual intervention, then the compatibility is insufficient, and this path is excluded.

[0062] The causal dependencies between the backbone nodes are analyzed from the selected set of executable causal paths. These dependencies reveal the constraints that parameters must meet, such as the temporal constraint that "the formaldehyde evaporation rate can only be activated after the cabinet temperature reaches 60 degrees Celsius," or the quantitative constraint that "the peak hydrogen peroxide concentration is inversely proportional to the action time, and the product must be greater than 3600 mg / min per cubic meter." These constraints are condensed into causal guarantee constraints, presented in the form of explicit logical expressions or conditional statements. Causal guarantee constraints ensure that the causal validity between key parameters is maintained when the inheritance scheme is executed, preventing the causal chain from breaking due to improper parameter configuration.

[0063] Simultaneously, implicit operational features of operational nodes are extracted from the set of executable causal paths, and these features are transformed into implicit operational guidelines that can be understood and executed by operators. The transformation of implicit features requires mapping numerical feature descriptions to operational behavior descriptions. For example, "the average temperature adjustment rate is 2.1 degrees Celsius per minute with a standard deviation of 0.3" is transformed into "when adjusting the heating knob, keep it rotating slowly and uniformly, with each rotation corresponding to a temperature change of about 2 degrees Celsius, and avoid rapid and large adjustments," or "the formaldehyde release interval follows a normal distribution with a mean of 180 seconds" is transformed into "after each formaldehyde release, wait about 3 minutes to observe the diffusion of white mist inside the cabinet, and then carry out the next release after the white mist is basically evenly distributed."

[0064] A complete knowledge transfer program is formed by structurally combining concise causal constraints with implicit operational guidelines. This program can be organized as a sequence of procedural steps, with each step associated with corresponding causal constraints and implicit operational guidelines. For example, a step described as "adjusting the heating plate temperature to the range of 65 to 70 degrees Celsius" has the associated causal constraint "must be performed after the relative humidity inside the cabinet has stabilized above 70%," and the implicit operational guideline is "when adjusting the knob, maintain a speed of one-quarter rotation every 10 seconds, and judge whether a stable upward trend has been reached by observing the rhythm of the temperature display's numbers." This knowledge transfer program not only clarifies the causal logic of the operation but also transmits the operational skills and experience accumulated by senior operators through long-term practice. This enables the learner to accurately reproduce the operation based on an understanding of the causal mechanism, improving the success rate and consistency of sterilization tasks.

[0065] In one optional implementation, reverse tracing is performed from the target sterilization effect node along the causal dependency of a stable causal path. During the reverse tracing process, causal necessity verification is performed for each intermediate node encountered, and the intermediate nodes that pass the verification constitute a set of causal backbone nodes, including: The process involves tracing back from the target sterilization effect node along a stable causal path towards the operation initiation direction, identifying the preceding intermediate nodes that have a direct causal dependency with the target sterilization effect node during the tracing back process. For each identified precursor intermediate node, a causal verification scenario is constructed. The causal verification scenario obtains the predicted sterilization effect value after shielding the causal transmission of the precursor intermediate node in the stable causal path and extrapolating to the target sterilization effect node. Calculate the effect deviation between the predicted sterilization effect value after shielding and the predicted sterilization effect value without shielding. When the effect deviation exceeds a preset deviation threshold, it is determined that the causal contribution of the precursor intermediate node is indispensable and passes the causal necessity verification. The predecessor intermediate node that passes the causal necessity verification is added to the causal backbone node set. The predecessor intermediate node in the causal backbone node set is used as the new tracing starting point to continue tracing back to the operation starting direction and repeating the causal necessity verification process until the tracing back to the operation starting node forms a complete causal backbone node set.

[0066] like Figure 2 As shown, the method includes: In the actual operation of the intelligent management system for disinfection and sterilization of biosafety cabinets, for specific target sterilization effect nodes, it is necessary to trace the key causal path leading to the effect from the trusted knowledge subgraph. This tracing process adopts a reverse tracing mechanism from effect to cause, and performs rigorous causal necessity verification on each intermediate node during the tracing process to ensure that each node in the final extracted set of causal backbone nodes has an irreplaceable causal contribution to the target sterilization effect.

[0067] Assume the target sterilization effect node is defined as the number of residual colonies after sterilization reaching a certain threshold. The sterility assurance level is determined. Starting from the target sterilization effect node, the process traces back to the operation initiation direction along a verified stable causal path in the trusted knowledge subgraph. In the stable causal path, each node has a causal dependency with its predecessor node, which is quantified by a causal strength coefficient. During reverse tracing, all predecessor intermediate nodes with a direct causal dependency on the target sterilization effect node are first identified. These predecessor nodes include nodes where the hydrogen peroxide spray concentration reaches the set value, nodes where the UV lamp irradiation intensity is stable within the specified range, and nodes where the air circulation flow rate reaches the standard value.

[0068] For each identified precursor intermediate node, a dedicated causal verification scenario is constructed to evaluate the causal necessity of that node. Taking the hydrogen peroxide spray concentration node as an example, a shielding mechanism is employed when constructing the causal verification scenario. This involves artificially blocking the causal influence transmission of the hydrogen peroxide spray concentration node during the transmission of the stable causal path. Specifically, the shielding operation is achieved by setting the output causal intensity coefficient of this node to zero, preventing it from transmitting any causal contribution to downstream nodes. After shielding this node, the deduction continues along the stable causal path towards the target sterilization effect node. The predicted sterilization effect value after shielding is calculated using the causal transmission model established in the causal knowledge graph. This deduction process comprehensively considers the causal contributions of other unshielded precursor nodes, and the final predicted sterilization effect value is calculated through the cumulative transmission of causal intensity coefficients.

[0069] Simultaneously, under normal conditions without shielding, the predicted sterilization effect value is obtained by extrapolating from all precursor intermediate nodes to the target sterilization effect node along the same stable causal path. This unshielded predicted value reflects the sterilization effect level achievable when all precursor nodes contribute normally. The effect deviation between the shielded predicted sterilization effect value and the unshielded predicted sterilization effect value is calculated; this deviation quantitatively reflects the actual causal contribution of the shielded node to the final sterilization effect. If the hydrogen peroxide spray concentration node is shielded, the predicted sterilization effect value changes from... The level has been reduced to At the level of [level], the deviation in effectiveness is reflected in a decrease of three orders of magnitude in the sterilization assurance level.

[0070] The causal necessity verification of a precursor intermediate node is determined based on a preset deviation threshold. The deviation threshold is set by comprehensively considering the safety level requirements of the sterilization task and the tolerance range of actual operation. For sterilization tasks in high-level biosafety cabinets, the deviation threshold is typically set to a decrease in sterilization assurance level of no more than one order of magnitude; while for sterilization tasks at the routine cleaning level, the deviation threshold can be appropriately relaxed to two orders of magnitude. When the effect deviation caused by shielding a precursor intermediate node exceeds the preset deviation threshold, it is determined that the causal contribution of that precursor intermediate node is indispensable, and the node passes the causal necessity verification and is added to the causal backbone node set. Taking the aforementioned hydrogen peroxide spray concentration node as an example, the three-order-of-magnitude decrease caused by its shielding significantly exceeds the deviation threshold, confirming that the node as a causal backbone node.

[0071] For precursor intermediate nodes that fail the causal necessity verification (i.e., nodes whose effect deviation after shielding does not exceed a preset deviation threshold), their causal contribution to the target sterilization effect is determined to be compensated or replaced by the causal effects of other nodes. These nodes are not included in the causal backbone node set. For example, a certain auxiliary fan speed node, after shielding, only causes the sterilization effect to decrease from... Descending to If the deviation does not exceed a threshold of one order of magnitude, the node is determined to be an unnecessary node.

[0072] After adding all the precursor intermediate nodes that pass the causal necessity verification to the causal backbone node set, these backbone nodes are used as new tracing starting points to continue reverse tracing back to the operation initiation direction. For each precursor intermediate node in the causal backbone node set, upstream precursor nodes with direct causal dependencies are identified again, and the causal necessity verification process is repeated for these upstream nodes. Taking the verified hydrogen peroxide spray concentration node as an example, its upstream precursor nodes include the storage tank pressure setting node, the nozzle diameter selection node, and the atomization time control node. A causal verification scenario is constructed for each precursor node. After shielding the storage tank pressure setting node, it is found that the hydrogen peroxide spray concentration cannot reach the set value, which leads to a significant decrease in the target sterilization effect exceeding the deviation threshold, confirming the storage tank pressure setting node as a new causal backbone node.

[0073] The process involves reverse tracing back to the starting point of the operation, with each layer rigorously verifying causal necessity and retaining only verified nodes to continue tracing upwards. In cases of branch convergence where multiple precursor nodes jointly influence the same downstream node, each precursor node undergoes independent causal necessity verification to ensure each verified node has an independent and indispensable causal contribution. For example, if both the UV lamp irradiation intensity node and the hydrogen peroxide spray concentration node are precursor nodes to the target sterilization effect node, and their respective shielding leads to a significant decrease in effect, both are verified and added to the causal backbone node set.

[0074] The process continues to trace back to the starting point of the operation until it reaches the initial node. This initial node typically corresponds to the initial operations of the sterilization process, such as the cabinet door closing confirmation node or the sterilization program initiation node. These initial nodes, as the source of the causal chain, are incorporated into the initial part of the causal backbone node set after passing causal necessity verification. This forms a complete causal backbone node set from the initial node to the target sterilization effect node. Each node in this set undergoes rigorous causal necessity verification to ensure its indispensable causal contribution to the target sterilization effect.

[0075] A complete set of causal backbone nodes constitutes the core supporting structure of the necessary causal path to achieve the target sterilization effect. This set eliminates redundant and substitutable nodes from the trusted knowledge subgraph, retaining each backbone node as a key control point in the sterilization operation. Through a combination of reverse tracing and causal necessity verification, the extracted causal backbone node set ensures that it fully covers all necessary conditions for achieving the target sterilization effect while avoiding increased operational complexity caused by including non-critical nodes. This causal backbone node set provides a concise yet sufficient causal basis for subsequently generating a succession scheme containing causal guarantee constraints, enabling the succession scheme to focus on the key operational steps that truly affect the sterilization effect, thus improving the executability and effectiveness guarantee of the succession scheme.

[0076] In one optional implementation, a deviation analysis is performed between the actual sterilization effect feedback data and the expected effect of the inheritance scheme. When the deviation exceeds the tolerance, the causal dependency relationship of the corresponding causal path in the causal knowledge graph is corrected, and an adversarial scenario perturbation is triggered to re-verify the corrected causal path, including: The actual sterilization effect feedback data is compared with the expected effect of the inheritance plan to obtain the effect deviation amount. When the effect deviation amount exceeds the preset tolerance, the deviation causal path corresponding to the effect deviation amount is located in the causal knowledge graph, and the deviation causal dependency relationship that leads to the effect deviation amount in the deviation causal path is identified. The biased causal dependencies are corrected to reduce the effect bias. The corrected causal dependencies replace the biased causal dependencies in the biased causal path to form a corrected causal path. The causal dependencies of the corresponding causal path in the causal knowledge graph are updated based on the corrected causal path. Adversarial scenario perturbations are triggered for the modified causal path in the causal knowledge graph. The adversarial scenario perturbations verify the causal stability of the modified causal path under perturbation conditions by introducing abnormal operating conditions into the modified causal path and extrapolating along the modified causal path to the target sterilization effect.

[0077] During the sterilization process in biosafety cabinets, to ensure the accuracy and adaptability of the causal knowledge graph, the effectiveness of the transfer protocol needs continuous monitoring and dynamic correction. During the implementation phase of the transfer protocol, multiple temperature sensors, pressure sensors, gas concentration detectors, and ultraviolet intensity detectors deployed inside the biosafety cabinet collect real-time data on environmental parameter changes during sterilization. These sensors continuously record the temperature distribution, hydrogen peroxide vapor concentration gradient, negative pressure fluctuations, and ultraviolet irradiation intensity decay curves at various locations within the sterilization chamber at a sampling frequency of 100 milliseconds, forming a high-density data stream of actual sterilization effect feedback.

[0078] To quantify the sterilization effect, biological indicators carrying known concentrations of tolerant spores were placed at six standard test sites within the sterilization chamber. After the sterilization cycle, the microbial survival rate at each site was measured using a culture method. The measured logarithmic microbial kill values, time-series curves of key parameters, sterilization time, and other data were compared item by item with the expected sterilization effect derived from causal path deduction in the inheritance protocol. The expected effect of the inheritance protocol is a theoretical value calculated based on the causal dependencies in a trusted knowledge subgraph; for example, when the hydrogen peroxide concentration is maintained at 450 ppm and the contact time reaches 25 minutes, the expected logarithmic microbial kill value should reach 6 log.

[0079] Deviation analysis employs a multi-dimensional evaluation strategy, establishing a method for calculating the effect deviation of microbial eradication efficacy. The difference between the measured logarithmic kill values ​​and the expected values ​​at six test sites is calculated, and the largest deviation is taken as the primary evaluation indicator. Simultaneously, the mean deviation and standard deviation of all test sites are calculated. When the absolute value of the primary evaluation indicator exceeds 0.5 log, or the mean deviation exceeds 0.3 log, the effect deviation is considered to have exceeded the preset tolerance range, triggering a causal path correction mechanism. This tolerance threshold is set based on the aseptic assurance level requirements in the biosafety cabinet sterilization validation standard, ensuring the timeliness and necessity of correction triggering.

[0080] After detecting excessive deviations, it is necessary to locate the specific causal path leading to the deviation in the causal knowledge graph. The causal knowledge graph is stored in a triplet structure of node-edge-attribute. Nodes represent sterilization parameters or effectiveness indicators, edges represent causal dependencies, and attributes record the dependency strength and confidence level. Through backtracking analysis, starting from the sterilization effectiveness node corresponding to the effectiveness deviation, all upstream parameter nodes affecting that effectiveness are traced backward along the causal path. For example, when the microbial killing effect at a certain test point is insufficient, tracing back reveals that the causal path corresponding to that point involves three key parameters: hydrogen peroxide injection rate, air circulation velocity, and sterilization chamber sealing.

[0081] A deviation contribution quantification analysis was performed on each traced causal path to calculate the degree of deviation between the actual value of each parameter node and the expected value of the causal dependency. Sensitivity analysis was used to assess the contribution weight of each parameter deviation to the final effect deviation, identifying the deviation causal path with the highest contribution. In the above case, if the measured value of the air circulation wind speed was found to be 0.35 m / s, while the effective wind speed range recorded in the causal dependency was 0.40 to 0.45 m / s, and the sensitivity analysis showed that the wind speed deviation had a weight of 62% on the kill effect, then the causal dependency corresponding to that parameter was marked as a deviation causal dependency.

[0082] To address the identified causal dependencies, an adaptive parameter correction strategy was employed. First, actual data collected during this sterilization cycle was used to establish an empirical mapping between parameters and effects. Virtual experiments were conducted with air circulation speeds ranging from 0.30 to 0.50 m / s in 0.02 m / s increments. Combined with other measured parameter values, the expected sterilization effect under different wind speeds was simulated and calculated. Fitting analysis determined that, under the current equipment conditions and environmental conditions, the critical wind speed for achieving a 6 log sterilization effect was 0.42 m / s. This value showed a significant difference from the records in the causal dependency analysis.

[0083] The expression of causal dependencies is revised based on empirical data, and the original fixed interval constraints are adjusted to dynamic adaptive functions. The revised causal dependencies incorporate equipment aging coefficients and environmental temperature and humidity correction factors, forming a causal relationship model that better reflects actual working conditions. For example, the revised relationship states that when the ambient temperature is below 18 degrees Celsius, the air circulation speed needs to be increased to 0.44 meters per second to compensate for the decrease in heat transfer efficiency. This revised causal dependency replaces the corresponding edge attributes in the original biased causal path, and the confidence score and applicability label of that edge are updated to form the revised causal path.

[0084] The revised causal path needs to be validated under adversarial perturbation scenarios to ensure its robustness under non-ideal conditions. The adversarial perturbation design simulates extreme situations encountered during sterilization, setting multiple sets of abnormal operating conditions for the corrected air circulation velocity parameter. The first perturbation scenario simulates a sudden fan failure causing the air velocity to drop instantaneously to 0.25 m / s during the middle of sterilization and remain there for 3 minutes. The second scenario simulates operator error causing periodic fluctuations in air velocity within the range of 0.38 to 0.48 m / s. The third scenario simulates airflow turbulence caused by a leak in the negative pressure system.

[0085] After introducing adversarial perturbations, forward extrapolation was performed along the modified causal path to calculate the expected sterilization effect under each perturbation scenario. The extrapolation process comprehensively considered the cascade response of all nodes in the causal path, simulating how parameter anomalies propagate to the final effect index through causal dependencies. For the first set of perturbation scenarios, the extrapolation results showed that a sudden drop in wind speed led to uneven distribution of hydrogen peroxide concentration in local areas, and the expected kill log value decreased to 4.2 log, which is lower than the safety threshold of 6 log, indicating that the modified causal path has a risk of failure under this perturbation condition.

[0086] The causal stability of the modified causal path is evaluated based on the simulation results. The stability evaluation index is defined as the probability that the sterilization effect remains above a safe threshold within a set perturbation intensity range. When the stability score is below 85%, the causal path is considered to require further optimization. To address the vulnerability exposed by sudden wind speed drops, an emergency compensation mechanism is added to the causal path, such as automatically extending the sterilization time or increasing the hydrogen peroxide injection concentration when the detected wind speed falls below a critical value. This compensation mechanism is embedded into the modified causal dependency, forming an enhanced causal path with adaptive fault tolerance.

[0087] The enhanced causal path was again subjected to adversarial scenario perturbation verification. The verification results showed that after introducing emergency compensation, even in the event of a sudden drop in wind speed, the final sterilization effect still reached 6.1 log by automatically extending the action time by 5 minutes, meeting the safety requirements. The verified corrected causal path, its applicable conditions, and compensation strategies were stored in the causal knowledge graph, and the attribute labels of the corresponding nodes and edges were updated. The correction history and verification record of the path were marked in the trusted knowledge subgraph. This closed-loop mechanism of correction and verification ensures that the causal knowledge graph can be continuously optimized as the performance of sterilization equipment evolves and the operating environment changes, providing a more accurate and reliable causal reasoning basis for the generation of subsequent inheritance schemes, and realizing the iterative inheritance of sterilization knowledge and the continuous improvement of quality assurance capabilities.

[0088] A second aspect of this invention provides an intelligent management and traceability system for the disinfection and sterilization of biosafety cabinets, comprising: The causal modeling unit is used to collect multimodal sterilization data streams, perform causal decoupling modeling on the multimodal sterilization data streams, construct a causal knowledge graph by identifying the causal dependency between sterilization parameter changes and sterilization effect response, and verify the effectiveness of causal paths in the causal knowledge graph. By introducing adversarial scenario perturbations, the failure boundaries of causal paths under parameter offsets or environmental changes are tested, and the verified stable causal paths are retained to form a reliable knowledge subgraph. The feature recognition unit is used to extract implicit operation features corresponding to operator interaction behavior data from the trusted knowledge subgraph, and to identify distinguishing features of different operation modes by comparing the differences in implicit operation features of different operators under the same sterilization task. The scheme generation unit is used to trace the necessary causal path set that satisfies the target sterilization effect from the causal path in reverse from the trusted knowledge subgraph for the sterilization task to be inherited, and to perform operation feasibility screening on the necessary causal path set in combination with the distinguishing features to generate an inheritance scheme that includes causal guarantee constraints and implicit operation guidance. The feedback correction unit is used to collect actual sterilization effect feedback data during the execution of the inheritance scheme, analyze the deviation between the actual sterilization effect feedback data and the expected effect of the inheritance scheme, and correct the causal dependency relationship of the corresponding causal path in the causal knowledge graph when the deviation exceeds the tolerance, and trigger adversarial scenario perturbation to re-verify the corrected causal path.

[0089] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0090] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0091] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent management and traceability of disinfection and sterilization of biosafety cabinets, characterized in that, include: Multimodal sterilization data streams are collected, and causal decoupling modeling is performed on the multimodal sterilization data streams. A causal knowledge graph is constructed by identifying the causal dependency between changes in sterilization parameters and sterilization effect response. The causal paths in the causal knowledge graph are validated. The failure boundaries of the causal paths under parameter offset or environmental change are tested by introducing adversarial scenario perturbation. Stable causal paths that pass the validation are retained to form a reliable knowledge subgraph. The implicit operation features corresponding to the operator interaction behavior data are extracted from the trusted knowledge subgraph. By comparing the differences in implicit operation features of different operators under the same sterilization task, the distinguishing features of different operation modes are identified. For the sterilization task to be inherited, the necessary causal path set that satisfies the target sterilization effect is traced backward from the causal path in the trusted knowledge subgraph, and the feasibility of the necessary causal path set is screened by combining the distinguishing features to generate an inheritance scheme that includes causal guarantee constraints and implicit operation guidance. During the execution of the inheritance plan, actual sterilization effect feedback data is collected. Deviation analysis is performed between the actual sterilization effect feedback data and the expected effect of the inheritance plan. When the deviation exceeds the tolerance, the causal dependency relationship of the corresponding causal path in the causal knowledge graph is corrected, and adversarial scenario perturbation is triggered to re-verify the corrected causal path.

2. The method according to claim 1, characterized in that, The validity of causal paths in the causal knowledge graph is verified by introducing adversarial scenario perturbations to test the failure boundaries of causal paths under parameter offsets or environmental changes. Stable causal paths that pass verification are retained to form a reliable knowledge subgraph, including: An adversarial detection mechanism is constructed for each causal path in the causal knowledge graph. The adversarial detection mechanism backpropagates the deviation gradient from the sterilization effect response node along the causal dependency to the parameter node and the environmental state node, calculates the gradient sensitivity of the parameter node and the environmental state node to the sterilization effect response node, and generates the perturbation direction and perturbation intensity that maximizes the prediction deviation based on the gradient sensitivity. The worst-case perturbation scenario is constructed by injecting directional perturbations into the parameter node and the environmental state node using the perturbation direction and the perturbation intensity. Under the worst-case perturbation scenario, the predicted sterilization effect value is obtained by forward deduction along the causal path and the corresponding actual sterilization effect measurement value is collected. The failure deviation between the two is calculated. When the failure deviation exceeds the preset failure threshold, the gradient sensitivity of the adversarial detection mechanism is used to trace and locate the weak node that caused the failure, record the failure mode formed by the weak node and the perturbation features of the worst perturbation scenario, and mark the causal path as a vulnerable path and remove it. When the failure deviation is lower than the preset failure threshold, the causal path is marked as a stable causal path, and all stable causal paths and their associated nodes are combined to form a trusted knowledge subgraph.

3. The method according to claim 2, characterized in that, The gradient sensitivity of the adversarial detection mechanism is used to locate the weak nodes that cause the failure. The weak nodes and the perturbation features of the worst-case perturbation scenario constitute the failure mode. The causal path is marked as a vulnerable path and removed, including: The contribution weights of each node in the causal path to the failure deviation are extracted from the gradient sensitivity calculated by the adversarial detection mechanism. The contribution weights are sorted in descending order and the set of sensitive nodes with the highest contribution weights is identified. For each sensitive node in the set of sensitive nodes, a reverse compensation perturbation is applied under the worst perturbation scenario and the causal path is re-deduced to obtain the predicted sterilization effect value after compensation. Calculate the failure mitigation degree between the compensated predicted sterilization effect value and the original predicted sterilization effect value. When the failure mitigation degree exceeds a preset mitigation threshold, the sensitive node is determined to be a weak node that causes failure. Extract the topological location and node type of the weak node in the causal path, and extract the perturbation direction and perturbation amplitude acting on the weak node from the worst perturbation scenario as perturbation features. Combine the topological location, the node type and the perturbation features to form a failure mode. The failure mode is bound and labeled with the causal path, and the causal path is marked as a vulnerable path and removed from the causal knowledge graph.

4. The method according to claim 1, characterized in that, Implicit operational features corresponding to operator interaction behavior data are extracted from the trusted knowledge subgraph. By comparing the differences in implicit operational features among different operators under the same sterilization task, distinguishing features for different operation modes are identified, including: Operator interaction behavior data is extracted from the stable causal path of the trusted knowledge subgraph. Behavioral pattern mining is performed on the operator interaction behavior data to identify the operation timing selection pattern and parameter adjustment response pattern. The operation timing selection pattern and the parameter adjustment response pattern are encoded into implicit operation feature vectors. For the same sterilization task, implicit operation feature vectors of multiple operators are collected to construct a feature comparison space, and feature difference measure between implicit operation feature vectors of different operators is calculated in the feature comparison space. Based on the feature difference measure, the implicit operation feature vectors are clustered to form operation pattern clusters. Inter-cluster separation analysis is performed on the operation pattern clusters to identify the feature dimensions that contribute the most to the clustering results. The implicit operation features corresponding to the feature dimensions are used as distinguishing features to differentiate different operation patterns.

5. The method according to claim 1, characterized in that, The set of necessary causal paths that satisfy the target sterilization effect are traced backward along the causal path from the trusted knowledge subgraph, and the feasibility of the necessary causal path set is screened by combining the distinguishing features to generate a succession scheme that includes causal guarantee constraints and implicit operational guidance, including: The target sterilization effect of the sterilization task to be inherited is located as the target sterilization effect node in the trusted knowledge subgraph. The causal dependency relationship along the stable causal path is traced backward from the target sterilization effect node. During the reverse tracing process, the causal necessity is verified for each intermediate node. The intermediate nodes that pass the verification are retained to form a set of causal backbone nodes. The causal dependency relationship connecting each node in the set of causal backbone nodes is connected to form a set of necessary causal paths. For each necessary causal path in the set of necessary causal paths, the operation node is matched with the operation mode recognition boundary corresponding to the operation node and the implicit operation feature carried by the operation node. When the matching degree reaches the preset matching threshold, the necessary causal path is determined to be executable at the operation level. Necessary causal paths with operational feasibility are selected to form an executable causal path set. The causal dependencies between the causal backbone nodes are parsed from the executable causal path set, and the causal dependencies are condensed into causal guarantee constraints. The implicit operation characteristics of the operation nodes in the executable causal path set are parsed and transformed into implicit operation guidelines. The causal guarantee constraints and the implicit operation guidelines are combined to generate a succession scheme.

6. The method according to claim 5, characterized in that, The causal dependencies of the target sterilization effect node are traced backward along a stable causal path. During the backward tracing process, the causal necessity of each intermediate node is verified. The intermediate nodes that pass the verification constitute the causal backbone node set, including: The process involves tracing back from the target sterilization effect node along a stable causal path towards the operation initiation direction, identifying the preceding intermediate nodes that have a direct causal dependency with the target sterilization effect node during the tracing back process. For each identified precursor intermediate node, a causal verification scenario is constructed. The causal verification scenario obtains the predicted sterilization effect value after shielding the causal transmission of the precursor intermediate node in the stable causal path and extrapolating to the target sterilization effect node. Calculate the effect deviation between the predicted sterilization effect value after shielding and the predicted sterilization effect value without shielding. When the effect deviation exceeds a preset deviation threshold, it is determined that the causal contribution of the precursor intermediate node is indispensable and passes the causal necessity verification. The predecessor intermediate node that passes the causal necessity verification is added to the causal backbone node set. The predecessor intermediate node in the causal backbone node set is used as the new tracing starting point to continue tracing back to the operation starting direction and repeating the causal necessity verification process until the tracing back to the operation starting node forms a complete causal backbone node set.

7. The method according to claim 1, characterized in that, The actual sterilization effect feedback data is compared with the expected effect of the inheritance plan. When the deviation exceeds the tolerance, the causal dependency relationship of the corresponding causal path in the causal knowledge graph is corrected, and an adversarial scenario perturbation is triggered to re-verify the corrected causal path, including: The actual sterilization effect feedback data is compared with the expected effect of the inheritance plan to obtain the effect deviation amount. When the effect deviation amount exceeds the preset tolerance, the deviation causal path corresponding to the effect deviation amount is located in the causal knowledge graph, and the deviation causal dependency relationship that leads to the effect deviation amount in the deviation causal path is identified. The biased causal dependencies are corrected to reduce the effect bias. The corrected causal dependencies replace the biased causal dependencies in the biased causal path to form a corrected causal path. The causal dependencies of the corresponding causal path in the causal knowledge graph are updated based on the corrected causal path. Adversarial scenario perturbations are triggered for the modified causal path in the causal knowledge graph. The adversarial scenario perturbations verify the causal stability of the modified causal path under perturbation conditions by introducing abnormal operating conditions into the modified causal path and extrapolating along the modified causal path to the target sterilization effect.

8. A biosafety cabinet disinfection and sterilization intelligent management and traceability system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The causal modeling unit is used to collect multimodal sterilization data streams, perform causal decoupling modeling on the multimodal sterilization data streams, construct a causal knowledge graph by identifying the causal dependency between sterilization parameter changes and sterilization effect response, and verify the effectiveness of causal paths in the causal knowledge graph. By introducing adversarial scenario perturbations, the failure boundaries of causal paths under parameter offsets or environmental changes are tested, and the verified stable causal paths are retained to form a reliable knowledge subgraph. The feature recognition unit is used to extract implicit operation features corresponding to operator interaction behavior data from the trusted knowledge subgraph, and to identify distinguishing features of different operation modes by comparing the differences in implicit operation features of different operators under the same sterilization task. The scheme generation unit is used to trace the necessary causal path set that satisfies the target sterilization effect from the causal path in reverse from the trusted knowledge subgraph for the sterilization task to be inherited, and to perform operation feasibility screening on the necessary causal path set in combination with the distinguishing features to generate an inheritance scheme that includes causal guarantee constraints and implicit operation guidance. The feedback correction unit is used to collect actual sterilization effect feedback data during the execution of the inheritance scheme, analyze the deviation between the actual sterilization effect feedback data and the expected effect of the inheritance scheme, and correct the causal dependency relationship of the corresponding causal path in the causal knowledge graph when the deviation exceeds the tolerance, and trigger adversarial scenario perturbation to re-verify the corrected causal path.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images