Multi-party accident analysis distributed storage and attribution reasoning method and computer device

Abstract

The application discloses a multi-party accident analysis distributed storage and attribution reasoning method and computer equipment. A chain-on-chain collaborative architecture is adopted, evidence fingerprint construction and minimum exposure evidence are completed on the chain, and through encryption identification, abstract fingerprint and global index, evidence credible evidence storage and audit trace are realized; evidence subgraph reconstruction and time sequence fusion are carried out on the chain, cross-domain entity alignment is executed, and fact grouping and reputation weighting mechanism are adopted to carry out quality screening and conflict arbitration; based on the fused subgraph, attribution reasoning is carried out, conflict evidence is corrected by defining edge weight and calculating causal transmission strength, and attribution conclusion with confidence is generated; finally, the conclusion abstract is written back to the chain to form an audit closed loop. The application realizes deep coupling of blockchain private sharing and knowledge graph reasoning, effectively suppresses low-quality evidence interference, improves the security and credibility of cross-subject evidence management, and makes the attribution result have the characteristics of explainability, verifiability and auditability.

Description

Technical Field

[0001] This invention relates to the field of industrial internet data processing and information security technology, specifically to a distributed storage and attribution reasoning method and computer equipment for multi-party incident analysis. Background Technology

[0002] Safety incidents in complex industrial scenarios are often triggered by the non-linear coupling of multiple factors, including equipment status, environmental parameters, and operational behaviors. A single threshold alarm is insufficient to explain the underlying evolutionary logic of the anomaly. In actual operations involving multi-party collaboration, data flow frequently suffers from delayed reporting, inconsistent reporting methods, and version conflicts, leading to broken chains of evidence. Subsequent manual comparison of fragmented data is insufficient for objectively verifying the boundaries of responsibility. Traditional centralized data governance models suffer from a lack of trust in cross-institutional scenarios, making data susceptible to unilateral tampering and lacking traceable underlying audit support.

[0003] While blockchain technology provides a trusted foundation due to its immutability, its throughput and privacy compliance requirements limit its ability to handle massive amounts of raw industrial data, typically necessitating a collaborative approach of on-chain indexing and off-chain storage. Existing technologies, such as using blockchain for controlled distribution of rule models or combining knowledge graphs with large models for fault attribution, exist, but these solutions are relatively fragmented: the former focuses on asset transfer and ownership confirmation, failing to address dynamic privacy boundary control during the reasoning process; the latter is largely based on centralized data assumptions, lacking a systematic design for minimizing data exposure, adjudicating conflicting evidence, and providing non-repudiation auditing in multi-party collaborations. Currently, the industry lacks a solution that can deeply integrate on-chain permission governance with off-chain semantic reasoning, effectively achieving cross-entity incident causal chain reconstruction and full-process verifiability without disclosing raw private data. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a distributed storage and attribution reasoning method and computer equipment for multi-party accident analysis, resolving the contradiction between data privacy protection and reliable traceability of evidence chains in industrial accident analysis under multi-party collaborative environments.

[0005] The first aspect of this invention provides a distributed storage and attribution reasoning method for multi-party incident analysis based on on-chain and off-chain collaboration, comprising the following steps:

[0006] Data structuring and preprocessing stage: Acquire multi-source raw data from the industrial site and establish multi-protocol heterogeneous data acquisition channels. Calculate state events of sensor analog data using discretization rules based on threshold and rate of change judgments, and map unstructured text into standardized event descriptions through regular expression matching. Construct structured vectors containing head entities, relationships, and tail entities; extract the system time at the time the raw data was generated as a precise timestamp; combine this with unique identifiers from the data source to construct contextual metadata; and finally, clean and map the multi-source raw data into a standard triplet evidence set.

[0007] Topological fingerprint generation and encryption stage: Asymmetric encryption is performed on the entities and relations in the triplet evidence set to generate encrypted identifiers. A digest fingerprint for each piece of evidence is calculated using a digest algorithm that concatenates the encrypted identifier with context metadata on the fly. The predecessor and successor sets of the current evidence node in the logical sequence graph are identified, and the aggregate value of the digest fingerprints of all nodes in the predecessor and successor sets is calculated using a bitwise XOR operation algorithm. A topological dependency hash for locking the causal structure is calculated using a secondary hash algorithm concatenating the current node's digest fingerprint with the aggregate values ​​of predecessors and successors, thus constructing the evidence fingerprint. A corresponding privacy level identifier is generated based on data sensitivity characteristics.

[0008] Blockchain Consensus Notification and Mapping Phase: Construct a notification payload including encrypted identifiers, topology-dependent hashes, privacy levels, and digital signatures. Calculate the signature validity using an elliptic curve digital signature verification algorithm and broadcast it to the blockchain network for network-wide consensus and distributed ledger recording, achieving minimal exposure and trusted traceability of evidence. The system monitors transaction receipts returned by the network, extracts the transaction hash and block height as on-chain anchors, and binds these on-chain anchors to the original evidence records stored in the local database, establishing a mapping relationship between the on-chain global index and distributed off-chain data.

[0009] Uncertainty-Driven Reasoning and Dynamic Update Phase: In response to incident attribution requests, an initial low-permission index and key are applied for from the blockchain to reconstruct the initial evidence subgraph. Multi-source cross-domain entity alignment is performed on the evidence subgraph. A weighted summation algorithm is used to calculate the comprehensive similarity score of the node pairs to be aligned across three dimensions: character morphology, attribute features, and topological structure. After comparison with the alignment judgment threshold, nodes are merged and adjacent edges are updated, generating a disambiguated fused evidence subgraph. Temporal fusion, quality screening, and conflict resolution are performed, employing a fact-based grouping and reputation-weighted quality assessment mechanism. The basic path association strength is calculated by multiplying the edge confidence weights and node prior importance scores. A time decay coefficient is calculated using an exponential decay function based on time difference, and a conflict penalty factor is calculated using a logical mutual exclusion check algorithm based on fact-based grouping combined with a reputation-weighted mechanism. The basic association strength, decay coefficient, and conflict penalty factor are multiplied to obtain the path causal transmission strength. The path strength with the same causal factor is accumulated to obtain the attribution score.

[0010] Subsequently, the attribution score probability distribution is calculated using a normalization algorithm, and the current system inference entropy is calculated based on this distribution using the Shannon entropy formula. If the system inference entropy value is higher than the stable convergence threshold, a state proof package containing the system inference entropy, iteration count, and signature is constructed. After verification through blockchain smart contract inequality comparison, the decryption key for the next privacy level is released, incremental evidence is obtained to update the subgraph, and the inference process is iterated. If the system inference entropy value is lower than or equal to the stable convergence threshold, inference is determined to be converged, conflicting evidence is corrected, and a confidence-based attribution conclusion is generated.

[0011] Closed-loop audit and verification phase: An attribution report is generated. A digest value containing the binary stream of the attribution report, the final system inference entropy value at the inference convergence moment, and the concatenated data from the full-process evidence index list is calculated using a cryptographic hash function to ensure the uniqueness of the binding of attribution content, inference state, and access path. The digest hash and digital signature are used to construct a write-back payload and submitted to the blockchain for evidence storage, forming an audit closed loop. The closed-loop audit verification module pulls the original on-chain payload from the isolated sandbox based on the on-chain index list and reconstructs the logical state of the inference process. Historical system inference entropy values ​​are extracted and compared with the stable convergence threshold to verify whether the current privacy disclosure operation is algorithmically necessary, thereby confirming the legality of the operation and the consistency of the attribution conclusion.

[0012] A second aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the distributed storage and attribution reasoning method for multi-party incident analysis based on on-chain and off-chain collaboration described in the first aspect.

[0013] The technical solution provided by this invention achieves deep coupling between private, controllable sharing of blockchain and semantic reasoning of knowledge graphs, constructing a transferable attribution analysis framework decoupled from specific industries. Adopting an on-chain and off-chain collaborative architecture, it achieves evidence fingerprint construction and minimal exposure storage through encrypted identifiers, digest fingerprints, and global indexes, ensuring the credibility and security of cross-entity data management. An uncertainty-driven system inference entropy control mechanism is introduced to achieve dynamic permission disclosure of data at multiple privacy levels. In the underlying attribution reasoning, a quality constraint mechanism integrating cross-domain entity similarity assessment, exponential decay over time, and fact grouping and reputation weighting is integrated to effectively suppress the interference of low-quality and conflicting evidence on the reasoning direction. Finally, through attribution summary write-back and audit verification based on an isolated sandbox, a complete audit closed loop is formed, enabling the complex large-scale cross-domain incident attribution results to possess strict interpretability, verifiability, and auditability characteristics.

[0014] This invention provides a distributed storage and attribution reasoning method and computer equipment for multi-party incident analysis. It has the following beneficial effects: 1. This invention achieves causal structure locking of evidence data by introducing a topological dependency hashing mechanism based on the XOR aggregation of predecessor and successor sets. This mechanism mathematically binds the content hash of a single piece of evidence to the hash values ​​of its logically adjacent nodes, ensuring that any tampering with a single evidence node will cause the topological verification of its associated nodes to fail. This cryptographically solidifies the integrity of the industrial accident evidence chain and the event sequence logic, effectively preventing data forgery and shirking of responsibility.

[0015] 2. This invention establishes a dynamic privacy disclosure mechanism based on Shannon entropy, enabling on-demand allocation of data access permissions. The system calculates the uncertainty entropy value during the inference process in real time, and only requests the next level of sensitive data keys from the blockchain smart contract when the entropy value exceeds a set threshold. This method uses the algorithmic quantification results as the basis for permission escalation, strictly adhering to the principle of minimizing data access while ensuring the accuracy of incident attribution, thus resolving the conflict between data value mining and core privacy protection in multi-party collaborative scenarios.

[0016] 3. This invention constructs a closed-loop audit system based on on-chain index backtracking and entropy value verification, ensuring the compliance of the reasoning process. The audit module utilizes the evidence index list stored on the chain and historical reasoning entropy values ​​to reconstruct the reasoning sequence without accessing the original plaintext data. It also verifies the legality of the permission disclosure operation by verifying the inequality relationship between historical entropy values ​​and thresholds, thereby achieving credible verification and responsibility determination throughout the entire process of incident attribution. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2This is an interaction diagram of the system logic architecture of the present invention. Detailed Implementation

[0018] The technical solutions in 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.

[0019] To better understand the present invention, the above content will be described in detail below with reference to specific embodiments.

[0020] Please see the appendix Figure 2 The present invention provides a distributed storage and attribution reasoning method and computer equipment for multi-party incident analysis, including: a data structuring and preprocessing module, a topological fingerprint generation and encryption module, a distributed off-chain storage module, a transaction construction unit of a blockchain consensus notarization module, an uncertainty-driven reasoning module, and a closed-loop audit verification module.

[0021] The data structuring and preprocessing module is configured to access multi-source heterogeneous industrial field data, perform cleaning operations, and convert the raw data into a standard triplet structure containing head entities, relations, tail entities, and metadata.

[0022] The topology fingerprint generation and encryption module is connected to the data structuring and preprocessing module. It is configured to perform asymmetric encryption on entities and relations in triples to generate encrypted identifiers, calculate the content hash of a single piece of evidence, and calculate the topology dependency hash based on the predecessor and successor sets of the time-series causal graph. At the same time, it sets the privacy level according to the data sensitivity.

[0023] The distributed off-chain storage module is configured to receive the encrypted raw data payload, distribute it across off-chain storage nodes, and return the corresponding data index address.

[0024] The transaction construction unit of the blockchain consensus notarization module is connected to the topology fingerprint generation and encryption module and the distributed off-chain storage module. It is configured to record notarization payloads including encrypted identifiers, topology dependency hashes, privacy levels and digital signatures. It maintains a global index and access control list through smart contracts and releases the decryption key when the verification conditions are met.

[0025] The uncertainty-driven inference module is configured to respond to attribution analysis requests, obtain the index and key corresponding to the authorization level through the trusted execution environment, reconstruct the local evidence subgraph, calculate the path causal strength, and dynamically determine whether to trigger the decryption request of the next level of data based on the system entropy value during the inference process.

[0026] The closed-loop audit verification module is configured to verify the consistency between the attribution conclusion summary and the on-chain record, and to review the legality of the reasoning process based on the record's citation list.

[0027] Please see the appendix Figure 1 The workflow of the distributed storage and attribution reasoning method for multi-party accident analysis provided by this invention includes: S1, the data structuring and preprocessing module, acquires multi-source raw data from the industrial site, extracts key elements, and converts them into a set of triplet evidence containing source identifiers and timestamps.

[0028] S2, the topology fingerprint generation and encryption module independently encrypts and generates object identifiers for each entity and relationship in the triple evidence set, calculates the content hash for each piece of evidence, and aggregates the fingerprints of the predecessor and successor nodes of the evidence through XOR operation to generate topology dependency hash, and constructs on-chain evidence storage payload containing privacy level markers.

[0029] S3. The system sends the evidence storage payload containing only the encrypted index and hash fingerprint to the transaction construction unit of the blockchain consensus evidence storage module for network-wide consensus and distributed ledger recording. The encrypted original data entity is sent to the distributed off-chain storage module for storage, establishing a mapping relationship between on-chain index and off-chain data.

[0030] S4. After receiving the accident attribution request, the uncertainty-driven reasoning module applies for initial permissions from the transaction construction unit of the blockchain consensus and evidence storage module, obtains the corresponding on-chain index set and the first-level decryption key, pulls the encrypted data from the distributed off-chain storage module, decrypts and reconstructs the initial evidence subgraph in the local trusted environment.

[0031] S5, the uncertainty-driven reasoning module performs entity alignment and temporal fusion on the initial evidence subgraph, searches for causal paths from potential causal nodes to accident target nodes, calculates the causal transmission strength based on the path's topology, time decay characteristics, and evidence conflict, and aggregates the attribution scores of each causal factor.

[0032] S6. The uncertainty-driven reasoning module calculates the current system reasoning entropy based on the distribution of attribution scores and compares the entropy value with a preset stability threshold. If the entropy value is higher than the stability threshold, a state proof is generated and submitted to the transaction construction unit of the blockchain consensus notarization module to apply for data access permissions for the next privacy level, obtain incremental data to update the evidence subgraph, and return to the execution step S5. If the entropy value is lower than or equal to the stability threshold, the reasoning is determined to be converged, and the final attribution conclusion is generated.

[0033] S7. The uncertainty-driven reasoning module generates an attribution report containing a list of causes and a full-process reference list, calculates the attribution conclusion summary hash and digital signature, constructs a write-back payload, and submits it to the transaction construction unit of the blockchain consensus notarization module for notarization.

[0034] S8. The closed-loop audit verification module reads the attribution conclusion summary hash and reference list on the chain during the review phase, reconstructs the reasoning path based on the reference list, and verifies the consistency between the calculation results and the evidence stored on the chain.

[0035] To address the difficulties in causal inference caused by inconsistent data source formats in industrial scenarios, the following steps are included: S101. Establish multi-protocol heterogeneous data acquisition channels and standardized mapping. The data structuring and preprocessing module connects to the underlying data sources in the industrial field through a pre-built adapter interface. These data sources include, but are not limited to, sensor time-series data transmitted via the OPCUA protocol, SCADA system alarm logs transmitted via the Syslog protocol, and equipment maintenance ledgers stored in a relational database. For data from these different sources, the module performs format cleaning operations, removing invalid characters and redundant whitespace, and maps unstructured text description fields to standardized event descriptions using regular expression matching and keyword extraction algorithms. For sensor analog data, the module converts it into discrete state events according to preset discretization rules, ensuring that all input data has a unified semantic granularity. Specific industrial communication protocol adaptation and analog discretization processing methods are well-known technologies in this field and will not be elaborated upon here.

[0036] S102. Constructing a mathematical model for the triple evidence set. To support subsequent topological dependency calculations and graph reasoning, the system abstracts the cleaned discrete events into a triple evidence set. For any piece of evidence extracted from the original data The module strictly defines it as a structured vector containing head entities, relations, and tail entities: ; in, The head entity refers to the active trigger or preceding state in a causal relationship, such as main ventilator No. 1 or abnormal voltage fluctuations. Indicates a relationship, referring to the logical connection or physical interaction between entities, such as causing, due to, belonging to, or located in; The term "tail entity" refers to the passive recipient or subsequent result in a causal relationship, such as shutdown protection or excessive gas concentration.

[0037] S103. Associating contextual metadata to construct a complete evidence object. A simple triple structure lacks spatiotemporal constraints and is insufficient to support dynamic reasoning. Therefore, the data structuring and preprocessing module generates structured vectors. Simultaneously, environmental information is extracted from the original logs to construct contextual metadata uniquely bound to this evidence. : ; in, It serves as a unique identifier for the data source, used to record the physical source channel of evidence and support source tracing in subsequent audits; This is the precise timestamp of the event, taken from the system time when the original data was generated, and used for subsequent time alignment and causal sequence verification.

[0038] Ultimately, each complete piece of data to be stored is encapsulated into a composite object containing structured semantics and environmental context. The data is then output to the next-level encryption module. Through this step, unstructured, multi-source heterogeneous data is transformed into graph nodes and edges that can be directly parsed and computed by a computer, realizing the mapping of physical world facts to digital space logic.

[0039] To ensure that specific industrial parameters cannot be deduced when the evidence is publicly stored on the blockchain, the following steps are taken: S201. Configure the encryption environment and key parameters. The topology fingerprint generation and encryption module loads the preset encryption algorithm configuration and obtains the encryption key of the current data owner. This key is used to define the mapping rules from plaintext space to ciphertext space, ensuring that only authorized parties holding the corresponding decryption private key can restore the original semantics in subsequent processes. The specific asymmetric encryption algorithm library calls and key pair generation and management are well-known technologies in this field and will not be elaborated upon here.

[0040] S202. Perform independent discrete encryption of entities and relations. For each independent semantic unit in the triplet evidence, the module does not use a whole-unit encryption approach, but instead calls encryption functions independently for the head entity, relation predicate, and tail entity. Through this discretization process, the system can generate the corresponding encrypted identifier. The specific calculation formula is as follows: ; ; ; in, Defined as the cryptographic identifier of the header entity, it serves as a unique index ID representing the source node in the on-chain storage structure; Defined as a relation encryption identifier, used to mark the type attribute of the connecting edges between nodes; Defined as the tail entity cryptographic identifier, it serves as a unique index ID representing the target node in the on-chain storage structure; Represents key-based The encryption mapping function is configured to have deterministic output properties, meaning that the ciphertext identifiers generated by the same plaintext entity under the same key remain consistent.

[0041] Through the processing in step S202 above, the system transforms the original semantic network into an abstract network composed of ciphertext identifiers. While this object-level encryption hides the specific physical meaning, it fully preserves the connection topology between entities. For example, if two different evidence events both involve the same device, then their corresponding... This ensures that all generated cryptographic identifiers will be identical, enabling the system to identify common nodes and connection structures in the graph with zero knowledge, providing the necessary structured foundation for subsequent construction of topology dependency hashes. , , It is stored in a temporary build buffer, awaiting hash aggregation processing.

[0042] To address the risk of data tampering during batch data reveal in a distributed environment, the specific execution steps are as follows: S203. Calculate the integrity fingerprint of the evidence content. The module reads the generated encrypted identifier combination and corresponding context metadata from the buffer, and calls the preset anti-collision hash function. Generate the content hash of a single piece of evidence This step aims to solidify the current node's own attribute information, ensuring that any tampering with the entity ID, relationship type, or timestamp will cause fingerprint verification to fail. The specific calculation logic is shown in the following formula: ; in, As evidence Content hash value; Collision-resistant hash functions for cryptographic digest algorithms, such as SHA-256 or SM3; This represents the concatenation operation of binary data, that is... , , and metadata The bytes are concatenated end-to-end in a predetermined order to form a complete byte stream input.

[0043] S204. Aggregate the fingerprint features of sequential neighbor nodes. To construct a topology locking mechanism, the system needs to identify the predecessor set of the current evidence node in the logical sequence graph. With successor set Precursor Collection The term refers to the set of historical evidence that precedes the current node in time and shares entity connections with the current node; the successor set. This refers to the set of future evidence that is temporally later than the current node and is related to it. The module traverses these two sets, extracts the content hash values ​​of all neighboring nodes, and aggregates them using an XOR operation. The reason for using the XOR operation is that it satisfies the commutative and associative laws, making the aggregation result independent of the order of the neighboring nodes and only dependent on the composition of the neighboring set, thus ensuring the verification stability of the graph structure.

[0044] S205. Generate a topological dependency hash to lock the causal structure. Based on the above calculation results, the module concatenates and hashes the current node's own content hash with the aggregate fingerprints of its predecessor and successor nodes to generate the final topological hash. This hash value not only represents the data integrity of the node itself, but also implicitly contains information about the node's positional constraints within the entire evidence chain. The specific formula for calculating the topological dependency hash is as follows: ; in, As evidence Topology dependency hash value; This indicates the XOR aggregation operation, which is used to perform a bitwise XOR operation on the hash values ​​of all elements in the set; Representatives include the precursor set Any preceding evidence node in the process; Representatives are included in the successor set. Any subsequent evidence node in the process.

[0045] Through the processing of steps S203 to S205 above, the system achieves topological locking of the data. When any piece of evidence... If the content of something changes, or its adjacency relationship in the causal graph is modified, not only its own... Dramatic changes will occur, and will also be through Operational cascading affects the hash calculation results of all associated nodes that consider it a neighbor. This mechanism constructs an indivisible chain of evidence at the cryptographic level, enabling auditors to verify the overall structural integrity of the data solely based on hash values.

[0046] After generating a hash fingerprint with topological locking properties, the transaction construction unit of the blockchain consensus notarization module performs on-chain notarization operations, encapsulating the encrypted identifier and topological hash generated in the aforementioned steps into a standardized blockchain transaction payload to ensure the immutability and permanent traceability of the evidence data. The specific execution steps are as follows: S301. Constructing a standardized notarization transaction payload. The transaction construction unit of the blockchain consensus notarization module extracts the calculated encrypted fingerprint data from the upstream buffer and assembles them into a transaction payload object to be uploaded to the chain according to a predefined serialization protocol. This payload object does not directly contain the original plaintext data, but only contains a ciphertext digest used for verification and indexing. Its structure is defined as follows: ; in, This represents the transaction payload object to be broadcast; The topological dependency hash value serves as the unique integrity verification root of this evidence in the current spatiotemporal state. The timestamp signature generated for the payload by the trusted timestamp service center is used to provide authoritative third-party time verification for the data and prevent the data owner from forging the generation time.

[0047] S302. Perform transaction digital signature and tamper-proof encapsulation. To verify the legitimacy of the transaction initiator and prevent the payload from being intercepted and tampered with during network transmission, the module calls the data owner's private key. For transaction payload objects The digest is digitally signed. This process follows the Elliptic Curve Digital Signature Algorithm or the Chinese national cryptographic standard SM2 signature standard, generating a digital signature. The calculation is as follows: ; in, This indicates a signature operation performed using a private key. For hash digest functions. Digital signatures. With transaction payload objects Together they form the final on-chain transaction structure. This ensures that only authorized devices holding the corresponding private key can initiate valid evidence storage requests.

[0048] S303. Execute distributed storage by calling the smart contract interface. The module connects to the blockchain network node via an RPC interface and calls the method of the pre-deployed on-chain evidence storage smart contract. This method receives... As input parameters, the distributed ledger's state update mechanism is triggered. At the contract execution layer, the transaction is packaged into the latest block through a consensus mechanism and broadcast to the entire network. The specific coding implementation of the smart contract and the P2P broadcasting mechanism of the blockchain nodes are well-known technologies and will not be elaborated upon here.

[0049] S304. Parse the receipt and establish a local index mapping. After the transaction is confirmed and packaged by the blockchain network, the transaction construction unit of the blockchain consensus and notarization module listens for and captures the transaction receipt information returned on-chain. This receipt contains two key parameters: transaction hash. With block height The system uses these two parameters as on-chain anchors, binding them to the original evidence records in the local database and updating the local index records. : ; By establishing this mapping, the system implements a hybrid storage architecture that uses off-chain storage of plaintext, on-chain storage of fingerprints, and bidirectional index anchoring. When auditing is required later, the system can rely on... Quickly locate on-chain evidence data and use the immutability of the data on the chain to reverse verify the authenticity of off-chain data.

[0050] To balance data privacy and security with the feasibility of inference computation, the specific implementation steps are as follows: S401. Initialize Inference Request and Permission Verification. The Uncertainty-Driven Inference module receives an attribution request containing a description of the incident target. In response to this request, the module initiates an initial access request to the transaction construction unit of the blockchain consensus and evidence storage module. This request, by default, only requests the minimum privacy level. Data access permissions. After verifying the legitimacy of the requester's identity signature, the smart contract returns the data marked as such in the current ledger. A list of evidence storage indexes at each level and their corresponding first-level decryption keys. The first-level decryption key here. It can only be used to decrypt non-sensitive basic data that can be obtained through public channels, ensuring that sensitive production data is not exposed in the initial stage.

[0051] S402. Retrieve on-chain indexes and off-chain ciphertext. The module parses the list of indexes returned from the blockchain and filters out those that meet the conditions. Global index of all evidence Based on these indexes, the module sends batch data read commands to the distributed off-chain storage module. The distributed off-chain storage module, as a content-addressable storage layer, returns the corresponding encrypted triplet data packets based on the indexes. Each returned data packet contains a header, entity, and ciphertext. Confidential Relationships and tail entity ciphertext .

[0052] S403. Perform secure decryption and entity reconstruction. In a locally isolated trusted execution environment, the uncertainty-driven inference module loads the first-level decryption key obtained in step S401. Perform symmetric or asymmetric decryption on the retrieved ciphertext data. This allows us to recover the original plaintext meaning. The specific decryption and recovery calculations are as follows: ; ; ; in, The first-level decryption key has been configured. The decryption function must have an algorithm logic that strictly corresponds to the algorithm used in the encryption stage.

[0053] S404. Construct the initial evidence subgraph. Using the set of decrypted plaintext triples, the module instantiates a graph data structure in memory to construct the initial evidence subgraph. .in, Representative of plaintext entities The set of nodes that constitutes Representative by relationship The set of edges formed. The subgraph constructed at this point. It only includes publicly available facts with low sensitivity, serving as the basic framework for subsequent iterative reasoning and multi-source fusion. For those that fail to utilize... For decrypted highly sensitive data, the system maintains its encrypted state or does not load it at the current stage until the subsequent inference entropy triggers the dimensionality-upgrading mechanism. Through this process, the system achieves an initial mapping between physical data and logical graph structure, laying a structured foundation for subsequent heterogeneous data fusion.

[0054] To eliminate the graph structure break caused by this semantic heterogeneity, the specific execution steps are as follows: S501, Calculate multidimensional entity similarity. Module traverses subgraph. Node pairs in The comprehensive similarity score between the two is calculated in three dimensions: character morphology, attribute features, and topological structure. This step aims to quantify the probability that two independent nodes point to the same physical entity. The specific formula for calculating the overall similarity is as follows: ; in, For nodes With nodes The overall similarity score between them, with a value range of [0,1]; Name morphological similarity calculated based on string edit distance is used to identify spelling variations or abbreviation differences; This is the attribute similarity calculated based on the node's associated metadata. This value is only high when both nodes have a compatible set of attributes. Topological similarity based on graph structure is obtained by calculating the Jaccard coefficient of the neighbor sets of two nodes, which is used to take advantage of the association feature that nodes with similar neighbors are often the same node. The preset weighting coefficients are used, and they satisfy the following conditions: This is used to adjust the influence of different similarity indicators according to specific application scenarios.

[0055] S502, Perform node fusion and subgraph update. The module will calculate the comprehensive similarity score. Alignment threshold with preset Comparison. When At that time, the system determines the node. and This is a duplicate mapping for the same entity. The module performs a merging operation, combining two nodes into a single unique identifier node and performing a union operation on their adjacent edge sets to eliminate redundant nodes. The merged node inherits all attribute information and historical associations from both original nodes, thus generating a disambiguated fused evidence subgraph. The specific algorithm implementation for character distance calculation and the set operation logic for Jaccard coefficients are well-known technologies in this field and will not be elaborated upon here. Through this step, the system ensures that subsequent causal inference is based on physically consistent entity connections, rather than discrete symbolic records.

[0056] Given the strong dependence of industrial accident analysis on causal time series and the asynchronous nature of multi-source data on time bases, the specific execution steps are as follows: S503, Time Base Unification and Time Series Index Construction. The module reads the context metadata associated with each node in the fusion subgraph. Extract the timestamp information. To address the heterogeneous time formats used by different data sources, the module calls a pre-built time parser to uniformly convert them into high-precision standard absolute time. Based on this, the module establishes an inverted time-series index for all event nodes in the subgraph, mapping discrete graph nodes to a globally unified time axis. Through this step, the system can quickly locate the set of events before and after any given time, providing a basis for subsequent calculations of the time decay function for the path. Provide precise time difference Input the basics to ensure the validity of causal inferences over time.

[0057] S504. Implementing a sliding window mechanism and causal time series constraints. To filter out noisy data that loses its correlation due to excessively large time spans and to ensure the correctness of the causal direction, the module deploys a sliding window mechanism on top of the time series index. The system sets an effective time window length. Define the time of the current incident node as The module only retains data within the specified time interval. The predecessor nodes within the range are used as the valid candidate set, and outdated historical data outside the range and data with timestamps later than [a certain value] are removed. The system then examines the future data. For nodes falling within the window, the system further checks whether their connectivity satisfies the causal time constraints, i.e., for any directed edge in the subgraph... Mandatory timestamp Less than or equal to timestamp If connections that violate temporal logic are found, the module will mark or prune such abnormal edges, thereby generating a temporally aligned, normalized evidence network, providing a logically rigorous graph data foundation for the next stage of path search and strength calculation.

[0058] S505, Candidate Causal Path Search. The uncertainty-driven reasoning module uses nodes marked as the starting point of an incident. As the root node, a reverse graph traversal algorithm is executed while satisfying topological temporal constraints. To address the high density of node connections in industrial IoT data, the system employs a constrained depth-first search or bundle search strategy, setting a maximum search depth. traverse all pointers The incoming edges. During the search process, the module, based on the time window determined in step S504, retains only paths that satisfy the time reversal constraint, and generates paths from potential causal nodes. Pointing to the accident node Set of feasible causal paths .

[0059] S506, Calculation of path causality transit strength. For sets... Each path in The module calculates its causal transmission strength. This indicator measures the likelihood of an accident propagating along a specific path. The calculation process comprehensively considers the confidence level of each piece of evidence along the path, the prior importance of each node, the decay characteristics over time, and logical conflicts. The specific calculation formula is as follows: ; in, Representing a path The strength of causal transmission; Represents a single edge in the path The confidence weight, which is determined by the data source reliability level in the metadata, reflects the credibility of evidence in single-step reasoning; Represents nodes on the path The prior importance score, which can be derived based on historical statistics from the PageRank algorithm or the asset criticality set by experts; The adjustment index is used to non-linearly adjust the weight of the impact of node importance on the overall score. Here, is the time decay function, used to simulate the physical law that causal effects weaken with increasing time intervals; its calculation formula is: ,in The time difference between the start and end points of the path. The preset time decay constant; This is a conflict penalty factor used to handle logical mutual exclusion in a chain of evidence. It applies when the path contains nodes that contradict known facts. The value should be close to 0 to block the score for that path; otherwise, the default value is 1.

[0060] S507, Causative Factor Attribution Score Aggregation. Due to a single potential causative node. The accident can occur through multiple intermediate paths. The module needs to aggregate the scores of all valid paths belonging to the same causative factor to calculate the comprehensive attribution score of that factor as the root cause of the accident. The aggregation calculation formula is as follows: ; in, Representing a path The starting node. By accumulating all nodes... Causal transit strength of a path originating from The system eliminates noise caused by multipath redundancy and obtains a quantitative ranking of candidate causes under the current evidence. The multiplication and summation operations in the above formulas are well-known techniques in this field and will not be elaborated upon here.

[0061] The uncertainty-driven reasoning module needs to assess the confidence level of the current reasoning result and decide whether to request further data disclosure from the blockchain evidence storage layer. The specific execution steps are as follows: S601. Normalization of the causative probability distribution. To quantify the overall uncertainty of the system, the module first normalizes the probability distribution of each potential causative node calculated in step S507. Comprehensive attribution score Transform it into a standardized probability distribution. Let... For any factor, the set of all currently identified potential causes is given. Its attribution probability The calculation formula is as follows: ; in, Characterization in existing evidence subgraphs Under the constraints, the causal node This is the posterior probability estimate of the root cause of the accident. The denominator is the sum of the scores of all candidate factors, ensuring that the sum of all probability values ​​is 1, thus constructing a valid probability space.

[0062] S602, System Uncertainty Entropy Calculation. This module uses Shannon entropy as a metric to calculate the inference entropy of the current inference state. This indicator directly reflects the degree of ambiguity of current evidence in identifying a single root cause of the accident. The calculation formula is as follows: ; in, A higher value indicates a more uniform probability distribution, meaning the system struggles to choose between multiple potential causes, leading to greater uncertainty in the inference results; conversely, a lower value indicates a more uniform probability distribution. The lower the value, the more focused the system is on a specific cause, and the higher the confidence level of the inference result.

[0063] S603, Convergence Judgment and Decision Branch. The system will calculate the entropy value. With the preset convergence threshold Perform a comparison. If... If the current evidence is sufficient to support the established reasoning conclusion, the system terminates the iteration and directly enters the audit report generation process; if The system determines that the current chain of evidence is insufficient to eliminate interference, and needs to further expand its data scope to proceed to the next round of data request process.

[0064] S604, High-Value Boundary Node Identification and Request Generation. When the system determines that further iteration is needed, the module needs to consider the boundary node set of the current subgraph. In the process, the optimal boundary node with the greatest information gain value is identified. This selection strategy aims to minimize system entropy by querying the node's adjacency relationships or attribute details. Optimal boundary node. The selection is based on the following expected information gain formula: ; in, This is the set of boundary nodes of the current evidence subgraph, i.e., those nodes that exist in the subgraph but whose complete adjacency relationships have not yet been fully decrypted or loaded; For pointing to the cause The set of known paths; This represents the current probability of causing the problem. For nodes The information content estimate is usually determined by the degree centrality of the node or the amount of data in the evidence index.

[0065] Selecting the optimal boundary node Then, the module generates a new data access request for that node ID and its associated time window. The request includes the hash identifier of the selected node, the required time range, and the verification signature for this iteration. It is then sent to the blockchain oracle node to trigger the next round of decryption and data retrieval operations. The logarithmic operations and optimization search in the above formula are well-known techniques and will not be elaborated upon here.

[0066] S605. Constructing a State Proof and Requesting Authorization. When the system determines that the current evidence is insufficient to support a clear attribution conclusion, the uncertainty-driven reasoning module automatically constructs an encrypted proof package containing the current reasoning state. This proof package serves as credentials for requesting higher-level data permissions from on-chain smart contracts. It proves the necessity of the request at the current moment, preventing malicious unauthorized access by the reasoning party under conditions of uncertainty. (State Proof) The data structure is defined as follows: ; in, The globally unique identifier hash of the original incident attribution request; The system inference entropy value calculated at the moment quantifies the uncertainty of the current inference state and is the core parameter for determining whether the conditions for on-demand disclosure are met. This counts the current inference iteration rounds to prevent replay attacks. The uncertainty-driven reasoning module uses its private key to generate a digital signature for the above content, which is used to verify the legitimacy of the applicant's identity.

[0067] Once constructed, the uncertainty-driven reasoning module uses the proof package as input parameters to call the permission control smart contract pre-deployed in the transaction construction unit of the blockchain consensus notarization module.

[0068] S606, On-chain Verification and Key Primitive Release. The access control smart contract receives the state proof. Subsequently, the authenticity and logical compliance of the proof content are atomically verified within the execution environment of the blockchain node. The verification logic comprises two parts: first, confirming that the request originates from a legitimate instance of the uncertainty-driven reasoning module through a signature verification algorithm; second, comparing the input entropy value. With the convergence threshold pre-set in the contract ,confirm The smart contract triggers a state transition and retrieves the next sensitivity level from the key-value store only when all of the above conditions are met. Corresponding encryption key The data is then sent to the secure trusted area of ​​the uncertainty-driven inference module. The deployment and execution of the aforementioned smart contracts are well-known technologies and will not be elaborated upon here.

[0069] S607, Incremental Evidence Encryption. The uncertainty-driven reasoning module obtains the decryption key. Then, based on the global index FID, an inquiry is sent to the distributed off-chain storage module targeting the sensitivity level. The distributed off-chain storage module responds to the data request, returning the corresponding level of encrypted evidence data blocks. The uncertainty-driven inference module utilizes the decryption key within a secure sandbox. The data block is decrypted to extract the newly added set of evidence triples, denoted as the incremental evidence subgraph. At this point, the system has completed the process of acquiring plaintext incremental data from state proof and is ready to enter the subsequent subgraph fusion calculation stage.

[0070] S608, Incremental Subgraph Fusion and State Update. The uncertainty-driven reasoning module decrypts the obtained incremental evidence data. Evidence subgraph in current memory Perform topology fusion. This fusion process is not a simple addition of nodes, but rather an execution of a union update operation on the graph data, mathematically expressed as: ; in, This represents the evidence subgraph before the update; This indicates that the newly decrypted data is at the next level of sensitivity. Incremental evidence subgraph; The union operator represents graph data. This operation merges newly added entity nodes, relation edges, and attribute information into the original graph and automatically establishes potential connections between the newly added nodes and the original nodes, thereby expanding the topological space of causal search.

[0071] After the fusion is complete, the uncertainty-driven inference module will be updated. As a new baseline map, the completeness of evidence in the system state is improved, providing new structural information to support the elimination of attribution uncertainty.

[0072] S609, Recursive Reasoning Iteration and Convergence Determination. Based on the Updated Evidence Subgraph. The system triggers recursive computation logic, returning to execute step S502 and its subsequent steps. The uncertainty-driven inference module then re-searches for a set of potential causal paths in the expanded graph space. The causal transmission strength of each path is recalculated based on the path scoring formula defined above, thereby updating the attribution scores of each potential causative factor. .

[0073] Subsequently, the system refreshes the attribution probability again using the Softmax function. Calculate the entropy of the new round of system inference. At this point, the system executes the following convergence determination logic: If the updated Still greater than the stable convergence threshold And the current permission level has not reached the maximum permission level set. The system then remains in a dynamic disclosure cycle, continuing to generate new state proofs to request the next level of data permissions; If the updated Less than or equal to the stable convergence threshold If no higher-level data is available for request, the reasoning process is considered converged. At this point, the uncertainty-driven reasoning module locks the current state of the evidence subgraph and the causal distribution results, terminates the recursive loop, and marks the current calculation results in memory as the final conclusion to be solidified, preparing for the subsequent evidence preservation stage. The graph traversal algorithm and recursive control logic described above are well-known technologies in this field and will not be elaborated upon here.

[0074] When uncertainty drives the inference module to determine the system inference entropy Meeting convergence criteria or reaching the maximum privilege level At that time, perform the following operations: S701. Attribution Report Generation and Full-Process Data Traceability. The uncertainty-driven inference module structures and encapsulates the final calculated causative factor distribution results in memory to generate the final attribution report. The report It does not contain only a single conclusion, but rather includes attribution scores. The system provides a list of causes, sorted from highest to lowest, a set of key evidence chains corresponding to each cause, and parameter snapshots during the reasoning process. Additionally, to support subsequent auditing and compliance verification, the system extracts a list of evidence indices used throughout the entire process from the runtime logs, denoted as... This complete evidence index list It records a global index, in chronological order, of all evidence fragments accessed by the uncertainty-driven reasoning module in each iteration. This is used to demonstrate that the system follows the principle of minimizing data access during the reasoning process, that is, it does not request redundant privacy data that is irrelevant to the attribution of the incident.

[0075] S702. Conclusion Digest Hash Calculation. To ensure the immutability of the attribution results and their consistency with the original on-chain data, the system performs cryptographic digest calculation on the attribution report and source data to generate a conclusion digest hash. This calculation process uses a collision-resistant hash function. The implementation, mathematically expressed as: ; in, Indicates attribution report The data stream after binary serialization; The system inference entropy value at the time of inference convergence is incorporated into the hash calculation to solidify the uncertain state of the system when making decisions and prevent the fabrication of high-confidence vacation images afterward. This is the complete evidence index list for the entire process mentioned above.

[0076] Using the above formula, any minor alteration to the content of the attribution report, falsification of the inference confidence level, or modification of data access records will result in changes to the final generated attribution report. Dramatic changes occur, thus ensuring the uniqueness and completeness of the conclusions.

[0077] S703, Write-back Payload Construction and Digital Signature. After completing the digest computation, the uncertainty-driven inference module constructs the write-back payload for on-chain persistent storage. This payload not only contains the fingerprint of the computation results, but also the identity credentials of the uncertainty-driven inference module, and its structure is defined as follows: ; in, Use its private key pair for uncertainty-driven inference modules Digital signatures generated by combining data.

[0078] This signature mechanism authenticates the identity of computing nodes, ensuring that only registered and authorized uncertainty-driven inference modules can write valid attribution conclusions to the blockchain, preventing malicious nodes from injecting false reports. The serialization algorithm and digital signature generation process described above are well-known technologies in this field and will not be elaborated upon here. After completing the payload construction, the system prepares to... Transaction construction unit submitted to the blockchain consensus and evidence storage module.

[0079] S704, On-chain notarized transaction consensus and state solidification. The uncertainty-driven inference module will construct the write-back payload. The transaction is encapsulated as a blockchain transaction and broadcast to the transaction construction unit of the blockchain consensus and notarization module. The consensus node first verifies the digital signature. The validity of the transaction is verified to confirm that the payload was indeed issued by an authorized uncertainty-driven inference module and has not been modified during transmission. Once verified, the transaction is packaged into a block, the smart contract updates the on-chain state, and... The data is persistently recorded in the distributed ledger as key-value pairs. At this point, the hash of the incident's cause conclusion and the evidence call records of the reasoning process have publicly available timestamps and proof of existence. Any deletion or modification to the original report will cause a mismatch between the off-chain calculated hash and the on-chain evidence storage hash, thus being identified by the system as an illegal operation.

[0080] S801, Audit Data Extraction and Environment Reset. When regulatory agencies or third-party auditors need to review incident attribution conclusions, the closed-loop audit verification module is activated. This module first identifies the incident request identifier... Query the associated final evidence-preserving transactions in the blockchain ledger and parse to obtain the conclusion digest hash. and a complete list of evidence throughout the process. Subsequently, the closed-loop audit verification module is based on The global index FID of the evidence recorded in the blockchain is used to batch retrieve the corresponding original on-chain payload packets from the transaction construction unit of the blockchain consensus and evidence storage module. Within the locally built isolated sandbox, the closed-loop audit verification module does not directly access the original plaintext data. Instead, it uses the aforementioned indexes and hash information to initialize the audit environment, preparing to reenact the reasoning and decision-making logic.

[0081] S802. Dynamic Disclosure of Legality Verification and Result Review. To prevent the uncertainty-driven inference module from maliciously requesting high-level privacy data without meeting necessary conditions, the closed-loop audit verification module must strictly verify the compliance of each privilege escalation operation during the inference process. The closed-loop audit verification module is based on... The recorded time series is used to reconstruct the logical state of the reasoning process step by step, and to address the dynamic reveal operation at the first step. In each iteration, the system inference entropy value recorded at that time is extracted. With a preset stable convergence threshold Perform a comparison and verification. This verification logic must satisfy the following inequality condition: ; If the above conditions are met, it indicates that in the first... The time-sharing system is indeed in a state of high uncertainty; request the next level of data access. If the algorithm is deemed necessary, the privacy disclosure operation is deemed legal; otherwise, the uncertainty-driven reasoning module is deemed to have engaged in unauthorized or illegal behavior, and the audit fails.

[0082] After confirming that all permission calls are legitimate, the closed-loop audit verification module further generates a local digest hash based on the simulation calculation results in the sandbox. and read it from the chain. Perform an equivalent comparison. If If the original attribution report is found to be complete, the calculation process compliant, and untampered, an audit pass report is output; otherwise, the system marks the attribution conclusion as unreliable and triggers an anomaly alarm. Through this dual verification based on cryptographic hashing and entropy logic, auditability of the calculation process is achieved while protecting the privacy of the original data.

[0083] S803, Audit Conclusion Feedback and Anomaly Location. The system generates a standardized audit report based on the comparison results in S802. If the local digest hash... With on-chain evidence hash If all results are completely consistent and all dynamic disclosure operations pass the entropy threshold verification, the system determines that the attribution and tracing is valid, issues a trusted attribution certificate, and notifies the relevant business parties. This certificate serves as the final basis for incident handling and liability determination.

[0084] If a hash mismatch or privacy overreach is detected, the closed-loop audit verification module initiates anomaly location procedures, utilizing a full-process evidence index list. The system records the timeline path, replays the reasoning process frame by frame, and locates the first iteration where the computational state deviated. The system marks the various context parameters of this abnormal iteration as disputed data points and automatically triggers an alarm transaction to the transaction construction unit of the blockchain consensus and evidence storage module. This alarm transaction locks the smart contract account involved in the reasoning and temporarily freezes the reputation scores of relevant nodes until manual intervention for further judicial evidence collection. This precise location mechanism based on a cryptographic evidence chain enables the system not only to detect tampering with the results but also to specifically identify when, where, and who committed the violation.

[0085] Through the attribution conclusion write-back and auditing method in Example 5, this invention realizes a novel data collaboration model where the computation process is invisible but verifiable. This model utilizes the avalanche effect of one-way hash functions and the non-repudiation of digital signatures to project complex off-chain reasoning logic into a fixed-state fingerprint on the chain. Even without exposing the original sensitive data plaintext, regulatory nodes can still rigorously prove, through mathematical methods, whether the derivation process from input data to output conclusion follows preset algorithm rules and privacy protocols, thereby solving the technical problems of opaque cross-entity data sharing and difficulty in determining responsibility in industrial accident analysis.

[0086] To verify the effectiveness of the method of the present invention in complex industrial scenarios, an illegal shutdown accident of the main ventilation fan in a large coal mine was selected as the specific implementation object. This scenario involves three heterogeneous business domains: environmental monitoring, electromechanical control, and personnel scheduling, and has typical characteristics of multi-source heterogeneous data and cross-domain privacy sensitivity.

[0087] Heterogeneous data acquisition and triple mapping. Within the time window preceding the accident, the system first acquires timing readings from the methane sensor in the gas monitoring subsystem, operating status commands from the main fan motor in the PLC control subsystem, and operation logs from the maintenance crew in the personnel positioning system. To adapt to the aforementioned general reasoning logic, the above unstructured industrial protocol data needs to be mapped into standardized evidence triples.

[0088] Specifically, for The system uses real-time sampling of methane concentration exceeding the limit as evidence. Head entity Mapped to a unique identifier for sensor devices ,relation Mapped to the operation predicate Report_Value, tail entity Mapped to a specific value of 2.5%; associated metadata. Includes the data source IP address and nanosecond-level timestamp.

[0089] for The system constructs the stop command triggered by the PLC as evidence. Head entity Mapped to logic controller identifier ,relation Mapped to the control predicate Execute_Interlock, tail entity Mapped to the action object Cut_Off_Power.

[0090] This mapping mechanism transforms physically independent sensor values ​​and electrical switch actions into semantically related graph node data at the logical level.

[0091] Topology-dependent encryption and fingerprint authentication. Based on data sensitivity standards, the system categorizes gas concentration data into low-privacy levels. The PLC's internal control logic code blocks and the maintenance team's detailed operation records are classified into high-privacy categories. and .

[0092] Regarding the generated evidence and The system performs topology-dependent hash calculations. For object-level encryption, the national standard SM2 elliptic curve algorithm is used as the encryption method. The specific implementation, generation and .

[0093] In calculating evidence When performing topological hashing, since the PLC's actions are triggered by gas exceeding limits in the temporal causal logic, the evidence... Identified as evidence predecessor node set The elements in the hash table are calculated according to the aforementioned topological dependency hash calculation formula.

[0094] This is the evidence. The result of XORing the content hash value with zero.

[0095] The SHA-256 algorithm was selected.

[0096] This calculation transforms the causal sequence of gas exceedance causing wind turbine shutdown in the physical world into a cryptographic hash constraint. If a malicious attacker subsequently attempts to delete the gas exceedance record to conceal the true cause of the shutdown, the evidence... Topological hash verification will be hampered by the lack of a predecessor. The failure resulted in cross-system data dependency locking.

[0097] Distributed storage and index publishing. After hash calculation, the system constructs a structure containing privacy tags. Public index information and containing The system provides hidden index information. For the original encrypted data, the system processes it differently based on its size: structured sensor triplet encrypted data is written to a time-series database, while unstructured maintenance work order scanned encrypted data is stored in a distributed file system. The specific read / write interface calls and sharding storage strategies for the database and file system are well-known technologies and will not be elaborated upon here.

[0098] Ultimately, the system will generate a global index. Topological hash The corresponding encrypted storage address pointer is packaged into the on-chain payload. The data is then broadcast to the mining consortium blockchain network. At this point, the blockchain ledger records the existence proof and structural fingerprint of all relevant data related to the incident, but does not disclose specific numerical values ​​or operational details, thus completing the environmental preparation before attribution inference.

[0099] After completing on-chain notarization and off-chain storage of the data, the system begins to process specific main fan abnormal shutdown requests. Perform causal inference.

[0100] Initial subgraph reconstruction and multi-source entity alignment. When the uncertainty-driven inference module receives an incident investigation command, it obtains only the minimum privacy level according to the preset initial security policy. Data access permissions. The engine utilizes on-chain indexes. Retrieve methane concentration records from the environmental monitoring subsystem and switch status from the power monitoring subsystem from the time-series database.

[0101] Due to differences in device naming conventions among different suppliers—for example, a gas sensor might be identified as CH4_Sens_001 in the monitoring system but recorded as Sen_Gas_N_Face in the asset ledger—the system needs to perform entity alignment. For the two extracted nodes to be aligned... and The system calls the entity similarity calculation formula to calculate it.

[0102] In this application scenario, weighting coefficients are set. . Calculate the edit distance of the device label string. Compare the attribute matching degree of the equipment installation locations. Analyze the structural similarities between the two in cable connection topologies. When the calculation yields... At that point, the system determined that both referred to the same physical entity and merged them in the local memory graph database, completing the initial fusion of cross-system data and constructing the basic evidence subgraph. .

[0103] Temporal causal path search and intensity quantification. Based on a fused full-process evidence index list subgraph. The uncertainty-driven inference module searches for all potential logical paths that could cause the main fan to shut down. The system identifies two main candidate paths: path The gas concentration rose sharply, triggering the power-off interlocking logic and causing the main fan to stop. path The voltage fluctuation in the power supply network caused the inverter protection to activate, resulting in the main fan stopping.

[0104] To determine which path is the true cause, the system calculates the causal transmission strength for each path. .

[0105] If the gas exceedance occurs 30 seconds before shutdown, and the system is set with a decay constant... If the time frame is close to 1, the function value is close to 1, indicating a close temporal correlation; if the time frame exceeds 1 hour ago, the function value approaches 0, essentially ruling out a causal relationship.

[0106] Entropy calculation and dynamic permission disclosure. Based on the path scoring results mentioned above, the system aggregates the comprehensive scores of each causative candidate. And it is transformed into a probability distribution using the Softmax function. .

[0107] Assuming in Probability of gas exceeding limits in the data view The probability of power grid fluctuations Other reasons probability At this point, the system calculates the current inference uncertainty based on the information entropy formula.

[0108] The calculation results show bit. The system compares this value with a preset stability threshold. Compare the bits to determine. This indicates that environmental monitoring data alone cannot clearly distinguish whether the problem is a gas issue or a power grid issue, and the attribution conclusion is highly ambiguous.

[0109] Based on this determination, the system generates a state proof containing the current entropy value of 1.45. Uploading to the blockchain automatically triggers a move to a higher privacy level. Access request. After the smart contract verifies that the entropy condition is met, it releases... The corresponding decryption key. After acquiring incremental data, the uncertainty-driven inference module discovered a clear record in the PLC log of a manual emergency stop button being pressed. The introduction of this evidence led to the path... The rating instantly rose, and and The probability of this significantly decreased. After another iteration of calculation, the new entropy value decreased to 0.2 bits, satisfying the convergence condition, thus accurately pinpointing the cause of the accident as human error, rather than equipment failure or environmental anomaly.

[0110] After the uncertainty-driven reasoning module arrives at the clear conclusion that human error caused the main fan to stop, the system enters the conclusion solidification and credibility verification stage to ensure the non-repudiation of the attribution result in a multi-party collaborative scenario.

[0111] Attribution conclusion encapsulation and summary evidence storage. The system first encapsulates the final results output during the inference phase into an attribution report. The report includes a sorted list of causes, key chains of evidence, and confidence levels of inference.

[0112] To support auditability of subsequent calculation processes, the system generates a full-process evidence index list. This list details the on-chain indices of all data objects invoked by the uncertainty-driven inference module as described above. Including a basic methane sensor index And the PLC log index for dynamic request decryption .

[0113] Next, the system uses a secure hash algorithm to calculate the digest hash of the attribution conclusion. . Subsequently, the system constructs a write-back payload. This payload is submitted to the blockchain network as a final proof of the incident analysis.

[0114] Entropy-based process legality verification. When regulatory agencies need to verify incident reports, they do not directly view the original sensitive data in plaintext, but instead initiate a closed-loop audit verification module. The audit module is downloaded from the blockchain. With initial evidence The reasoning logic is reenacted in the local sandbox environment.

[0115] The core of auditing lies in verifying the necessity of data access and preventing privacy breaches. The system is based on... The records in the file were used to locate the uncertainty-driven reasoning module in the first... Requested during the round of iteration Access permissions for level-one data. At this point, the audit module extracts the level-one data. intermediate state entropy value during round iteration And verify whether it meets the dynamic disclosure conditions: The above verifies that access to highly sensitive data is only permitted under conditions of excessive uncertainty. If, during the audit process, a call to high-priority data is found to have occurred... In cases where sensitive data is requested even after the conclusion is clear, it is considered an overreach of privacy rights. This verification mechanism based on mathematical inequalities allows regulators to determine the compliance of data access without needing to know the specific control code of the PLC.

[0116] Trustworthy closed loop and accountability. After completing the hash comparison and entropy verification, if the topological fingerprint of the input data and the digest hash of the output conclusion match correctly, and all data disclosure operations comply with the minimum necessary principle, the audit module automatically issues a trusted attribution certificate. This certificate confirms the legal validity of the conclusion of human error and excludes the possibility of data tampering or malicious manipulation of the algorithm.

[0117] Conversely, if a hash mismatch or entropy value verification failure is found during verification, the system will mark the abnormal step as a disputed state and proceed accordingly. This precisely identifies which sensor data point within which time window caused the verification failure. This provides a precise data target for subsequent manual intervention investigations, thus striking a balance between protecting mining companies' trade secrets and ensuring transparency in safety supervision, achieving closed-loop management of the entire process for attributing complex industrial accidents. The specific presentation format or notification method after certificate issuance is well-known in this field and will not be elaborated upon here.

Claims

1. A distributed storage and attribution reasoning method for multi-party accident analysis, characterized in that, Includes the following steps: S1. The data structuring and preprocessing module acquires multi-source raw data from the industrial site, cleans and maps the multi-source raw data from the industrial site into a standard triplet evidence set containing head entities, relations, tail entities and metadata. S2. The topological fingerprint generation and encryption module performs asymmetric encryption on the entities and relations in the triple evidence set, generates encryption identifiers, and constructs evidence fingerprints. The evidence fingerprints include a digest fingerprint calculated by using a collision-resistant hash function, a topological dependency hash used to characterize the logical structure between evidences, and a corresponding privacy level identifier generated based on data sensitivity characteristics. S3. Send the evidence payload containing the encrypted identifier and topology dependency hash to the blockchain consensus evidence storage module for network-wide consensus and distributed ledger recording, achieving minimal exposure of evidence storage and trusted traceability; send the encrypted original data to the distributed off-chain storage module for storage, and establish a mapping relationship between the on-chain global index and the off-chain data. S4. The uncertainty-driven reasoning module responds to the accident attribution request, applies to the blockchain consensus and evidence storage module for the initial low-permission index and key, and reconstructs the initial evidence subgraph. S5. The uncertainty-driven reasoning module performs multi-source cross-domain entity alignment, temporal fusion, quality screening and conflict adjudication on the evidence subgraph, searches for causal paths from potential causal nodes to accident target nodes, defines edge weights and calculates the path causal transmission strength of the causal paths. In the quality screening and evaluation process, a quality evaluation mechanism of fact grouping and reputation weighting is adopted. S6. The uncertainty-driven reasoning module calculates the current system reasoning entropy using the Shannon entropy algorithm based on the attribution score distribution. If the system reasoning entropy value is higher than the stable convergence threshold, it applies to the blockchain for the decryption key of the next privacy level, obtains incremental evidence to update the subgraph, and returns to execute step S5. If the system reasoning entropy value is lower than or equal to the stable convergence threshold, it determines that the reasoning has converged, corrects conflicting evidence, and generates a confidence-based attribution conclusion. S7. The uncertainty-driven reasoning module generates an attribution report, calculates the summary hash and digital signature of the attribution conclusion, constructs a write-back payload and submits it to the blockchain consensus and notarization module for notarization, forming an audit closed loop. S8, the closed-loop audit verification module reconstructs the reasoning process based on the evidence index list recorded on the chain, and verifies the legality of the dynamic permission-revealed operations and the consistency of the attribution conclusions.

2. The distributed storage and attribution reasoning method for multi-party accident analysis according to claim 1, characterized in that, In step S1, mapping the multi-source raw data into a standard triplet evidence set includes: Establish a multi-protocol heterogeneous data acquisition channel, calculate the state events of sensor analog data through discretization rules based on threshold judgment and rate of change judgment, and map unstructured text into standardized event descriptions through regular expression matching; Construct a structured vector containing head entities, relationships, and tail entities, and extract the system time when the original data was generated as a precise timestamp. Combine this with the unique identifier of the data source to construct context metadata bound to the structured vector.

3. The distributed storage and attribution reasoning method for multi-party accident analysis according to claim 1, characterized in that, In step S2, the calculation of the topological dependency hash used to characterize the logical structure between evidences, specifically employing an XOR aggregation method based on the predecessor and successor sets, includes: The digest fingerprint of a single piece of evidence is calculated by a digest algorithm that concatenates the encrypted identifier with the context metadata in real time; Identify the predecessor and successor sets of the current evidence node in the logical sequence diagram, and calculate the aggregate value of the summary fingerprint of all nodes in the predecessor set and the aggregate value of the summary fingerprint of all nodes in the successor set through a bitwise XOR operation algorithm. The topological dependency hash of the causal structure is calculated by concatenating the current node digest fingerprint with the aggregate values ​​of its predecessors and successors using a quadratic hash algorithm.

4. The distributed storage and attribution reasoning method for multi-party accident analysis according to claim 1, characterized in that, In step S3, establishing the mapping relationship between on-chain indexes and off-chain data includes: The transaction construction unit of the blockchain consensus notarization module receives notarization payloads containing cryptographic identifiers, topology dependency hashes, privacy levels, and digital signatures. It calculates the validity of the signature using an elliptic curve digital signature verification algorithm and broadcasts it to the blockchain network. The system listens to the transaction receipts returned by the blockchain network, extracts the transaction hash and block height as on-chain anchors, binds the on-chain anchors to the original evidence records stored in the local database, and updates the local index records.

5. The distributed storage and attribution reasoning method for multi-party accident analysis according to claim 1, characterized in that, In step S5, performing multi-source cross-domain entity alignment includes: The weighted summation algorithm is used to calculate the comprehensive similarity score of the node pair to be aligned in three dimensions: character morphological similarity, attribute feature similarity, and topological structure similarity. Whether to perform node fusion operation is determined by numerically comparing the comprehensive similarity score with the alignment determination threshold; When a decision is made to perform fusion, nodes are merged and the set of adjacent edges is updated to generate a fusion evidence subgraph that has undergone disambiguation processing.

6. The distributed storage and attribution reasoning method for multi-party accident analysis according to claim 1, characterized in that, In step S5, the calculation of the path causal transmission strength of the causal path, using a path scoring algorithm that comprehensively considers confidence weights and time decay, specifically includes: The basic association strength of a path is calculated by multiplying the confidence weight of a single edge in the path and the prior importance score of a node. The decay coefficient in the time dimension is calculated using an exponential decay function based on the time difference between the start and end points of the path. The conflict penalty factor is calculated by a logical mutual exclusion check algorithm based on the fact grouping, and a quality assessment is performed in combination with a reputation weighting mechanism; The causal transmission strength of a single path is calculated by multiplying the basic correlation strength, the attenuation coefficient, and the conflict penalty factor, and the attribution score of the causal factor is calculated by summing the strengths of all paths for the same potential causal factor.

7. The distributed storage and attribution reasoning method for multi-party accident analysis according to claim 1, characterized in that, In step S6, requesting the decryption key for the next privacy level from the blockchain specifically includes: The probability distribution of the attribution scores of each potential causative factor is calculated using a normalization algorithm. The inference entropy of the system is calculated using the Shannon entropy formula, which is a logarithmic weighted sum of probability distributions. When the system inference entropy is higher than the stable convergence threshold, construct a state proof package containing the current system inference entropy, the inference iteration round count, and the digital signature. The blockchain smart contract receives the state proof packet and compares it with the logical calculation system to see if the inference entropy is strictly greater than the stable convergence threshold stored in the contract. If the verification is successful, the decryption key corresponding to the next privacy level is released.

8. The distributed storage and attribution reasoning method for multi-party accident analysis according to claim 1, characterized in that, In step S8, verifying the legality of the dynamic permission disclosure operation includes: The closed-loop audit verification module utilizes the on-chain full-process evidence index list and pulls the original on-chain payload in an isolated sandbox. Based on the logical state of the reasoning process reconstructed in time sequence, for each iteration step in which data disclosure operation has occurred, the historical system reasoning entropy value recorded at that time is extracted; A numerical determination method is used to calculate whether the current privacy disclosure operation is algorithmically necessary by comparing the historical system inference entropy value with the stable convergence threshold: if the historical system inference entropy value is greater than the stable convergence threshold, it is determined to be legal; otherwise, it is determined to be illegal and unauthorized.

9. The distributed storage and attribution reasoning method for multi-party accident analysis according to claim 1, characterized in that, In step S7, the digest hash of the attribution conclusion is obtained by calculating the digest value of the concatenated data containing the binary stream of the attribution report, the final system inference entropy value at the inference convergence time, and the full-process evidence index list through a cryptographic hash function, ensuring the uniqueness of the binding of attribution content, inference determinism state, and data access path.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-9.

Images