A dynamic enterprise credit scoring method and system based on an artificial intelligence model

Abstract

This invention relates to the field of credit scoring technology and discloses a dynamic enterprise credit scoring method and system based on an artificial intelligence model. The method includes: collecting event records within the scoring window of a manufacturing production line, containing workpiece identifiers, local event times, and gateway reception times; binding the actual workpiece path based on the directed graph of the production line and synchronous on-site evidence; obtaining the correction edge propagation time; and generating a material flow constraint process chain event time self-consistency coefficient representing the time-estimated coverage ratio, end-to-end closure residual, and edge-by-edge propagation deviation. The method further generates a time-series graph input and an unordered aggregated input, obtains two branch risk probabilities through a solidified dual-branch artificial intelligence model, performs deterministic degradation when there is no valid time-series input, and performs time-reliable fusion and conflict correction according to the coefficient, outputting the dynamic enterprise credit score and enterprise credit rating of the scoring window.

Description

Technical Field

[0001] This invention belongs to the field of credit scoring technology, specifically relating to a dynamic enterprise credit scoring method and system based on an artificial intelligence model. Background Technology

[0002] Manufacturing companies typically need to continuously monitor their current production continuity in credit granting, post-loan monitoring, supply chain finance, and performance evaluation. Traditional credit scoring relies heavily on financial statements, transaction records, invoices, tax records, and manually set operating indicators. This data has a long update cycle and a time lag compared to the real-time operational status of the production site. With the application of Industrial Internet of Things (IIoT), RFID, PLC, edge gateways, encoders, photoelectric triggers, vibration and power acquisition in production lines, the production site can now generate event records related to workpiece flow, equipment start-up and shutdown, buffer backlog, and production line release. These records provide the basic conditions for assessing enterprise operational risks.

[0003] In actual deployments, data acquisition nodes on the manufacturing floor often do not share the same local clock. RFID readers, PLCs, sensors, conveyor controllers, and edge gateways may come from different manufacturers, and their sampling periods, caching strategies, communication links, and reporting paths may also differ. If the system directly splices multi-device events based on server reception time, device reporting time, or a fixed time grid, it is easy to introduce time deviations caused by gateway caching, network queuing, wireless retransmission, and local clock drift into the model input. For the same workpiece flowing along the production line, there should be a stable sequence between upstream departure, conveying, buffering, and downstream arrival; when the data acquisition link causes time misalignment, the model may identify this misalignment as a change in production line cycle time or a change in the risk of interruption.

[0004] Even when existing enterprise credit scoring systems incorporate Industrial Internet of Things (IIoT) data, they often input statistical indicators such as equipment runtime, output, energy consumption, or alarm frequency into the model, rarely verifying the reliability of cross-device event timings themselves. Some time-series models can handle irregular sampling or missing data, but their inputs typically still rely on an established timeline; if the timeline is affected by local clock differences and transmission delays during the acquisition phase, the risk probability output by the model may still be skewed. Using communication synchronization messages, data integrity rates, or model confidence levels alone is insufficient to determine whether the records of the same workpiece across different nodes on the production line conform to the actual flow patterns. Summary of the Invention

[0005] The present invention aims to provide a dynamic enterprise credit scoring method and system based on an artificial intelligence model, which solves the technical problems mentioned in the background art.

[0006] To achieve the above-mentioned objectives, the present invention provides the following technical solution.

[0007] A dynamic enterprise credit scoring method based on an artificial intelligence model is applied to a manufacturing production line where the entry and exit reference nodes are in the same hardware clock domain and intermediate acquisition nodes retain their local clocks. The method includes: acquiring event records containing workpiece identifiers, local event times, and gateway reception times within the scoring window; using the gateway reception time to form the event candidate order and relative clock correction reference; binding the actual workpiece path based on the workpiece identifier, the directed graph of the production line, and synchronous physical evidence; obtaining the correction edge propagation time based on physical propagation time, edge-level relative clock residuals, relative clock drift rate, and window start-point relative clock offset; and generating a representation time that can estimate the coverage ratio based on the entry reference event time, exit reference event time, physical propagation time, and correction edge propagation time. The event time self-consistency coefficient of the material flow constraint process chain is calculated based on the end-to-end closure residual and the edge-by-edge physical propagation deviation. A time series graph input is generated based on the correction edge propagation time, and an unordered aggregated input is generated from the event records. A fixed dual-branch artificial intelligence model is used to obtain the time series branch risk probability and the unordered aggregated branch risk probability based on the two inputs. When the time series graph input is empty, the branch risk probabilities are made consistent. Time-reliable fusion and conflict correction are performed based on this coefficient. When the coefficient increases, the contribution of the time series branch risk probability is improved; when the coefficient decreases, the contribution of the unordered aggregated branch risk probability is improved. When the coefficient decreases and the divergence of the branch risk probabilities increases, conservative risk correction is performed to obtain the robust risk probability. The robust risk probability is mapped to the dynamic enterprise credit score and enterprise credit rating of the scoring window.

[0008] Compared with the prior art, the present invention has the following substantial features and significant progress:

[0009] This invention incorporates field events generated by RFID readers, PLCs, encoders, edge gateways, photoelectric triggers, vibration acquisition, and power acquisition into a single scoring process. It verifies cross-device event times using the actual workpiece path, adjacent node propagation time, entry reference event time, and exit reference event time. Furthermore, it controls the contribution relationship between the probability of sequential branch risks and the probability of disordered aggregation branch risks using a self-consistent coefficient for process chain events constrained by material flow. This ensures that when the time is reliable, the impact of continuous process changes on the scoring is preserved; when the time is unreliable or the timing diagram input is empty, it switches to current running statistics independent of cross-device order. This reduces abnormal fluctuations in dynamic enterprise credit scores caused by local clock drift, buffer retransmission, network queuing, event mismatch, and missing records, and maintains a verifiable processing relationship between the scoring output, the production site acquisition link, and the enterprise credit rating. Attached Figure Description

[0010] Figure 1 This invention relates to a dynamic enterprise credit scoring system architecture and data flow diagram based on an artificial intelligence model. Detailed Implementation

[0011] The following description is provided in conjunction with the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the invention; those skilled in the art can make equivalent substitutions or combinations for the specific implementations.

[0012] This implementation is applicable to manufacturing production lines where the inlet and outlet reference nodes are in the same hardware clock domain and intermediate acquisition nodes retain their local clocks. This manufacturing production line can be a discrete manufacturing production line, assembly line, processing line, sorting line, inspection line, or a hybrid production line with a buffer zone. Workpieces, carriers, pallets, or batch carriers within the production line are uniquely bound through workpiece identifiers.

[0013] like Figure 1 As shown, the manufacturing production line 100 includes an inlet reference node 110, an outlet reference node 120, and an intermediate acquisition node 130. The inlet reference node 110 and the outlet reference node 120 are connected to the same hardware timing gateway 160 to form an end-to-end reference in the same hardware clock domain. The intermediate acquisition node 130 may include an RFID reader / writer 131, a PLC 132, a conveyor encoder 133, a boundary photoelectric trigger 134, a vibration sensor 135, an active power acquisition module 136, and an edge gateway 137. The intermediate acquisition node 130 retains a local clock to reflect the distributed acquisition conditions in the field.

[0014] As workpiece 140 flows along manufacturing production line 100, an event log 150 is generated. Event log 150 includes at least the workpiece identifier, local event time, and gateway reception time, and may further include the device identifier, event type, boundary photoelectric status, vibration root mean square, active power, encoder cumulative pulses, buffer count, and data source status code. Event log 150 is received by physical operation data acquisition module 200 and sequentially enters workpiece path binding module 210, time self-consistent calculation module 220, and anchored time axis and dual-input generation module 230 in scoring processing unit 400.

[0015] In the credit scoring calculation unit 500, the fixed dual-branch AI reasoning module 240 includes a temporal branch 241 and a disordered aggregation branch 242. The temporal branch 241 receives temporal graph input, and the disordered aggregation branch 242 receives disordered aggregation input. The time-reliable fusion and conflict correction module 250 generates robust risk probabilities based on the material flow constraint process chain event time self-consistency coefficient, the temporal branch risk probability, and the disordered aggregation branch risk probability. The order-preserving calibration and credit scoring output module 260 outputs a credit output 600 based on the robust risk probability. The credit output 600 includes a dynamic enterprise credit score 610 and an enterprise credit rating 620.

[0016] According to Figure 1 Final drawing or reference Figure 1At that time, the final grade output within the credit output 600 is uniquely labeled as enterprise credit grade 620, and the sequence calibration and credit scoring output module 260 is uniquely labeled as a functional module. 260 may not be used simultaneously for enterprise credit grade. The correspondence between credit output 600, dynamic enterprise credit score 610, enterprise credit grade 620, and the reference numerals in the accompanying drawings may not be replaced, compressed, or reused interchangeably. This constraint is used to maintain consistency of reference numerals between the patent drawings, the patent drawing MD file, the reference numerals description, and this specific embodiment.

[0017] The basic idea of ​​this implementation is not to simply splice together the multi-device events collected from the manufacturing production line 100 into a unified time series, but rather to verify the cross-device event times using the actual flow patterns of the workpiece 140, the physical propagation time of adjacent node connections, the entry reference event time, and the exit reference event time before generating the time series diagram input. The self-consistency coefficient of the material flow constraint process chain event times obtained from this verification is then used in the inference stage of the artificial intelligence model. Thus, when the event times are reliable, the contribution of the time series branch risk probability is fully preserved; when the event times are unreliable or the time series diagram input is empty, the scoring process continuously shifts to unordered aggregation of branch risk probabilities.

[0018] The physical operation data acquisition module 200 is used to collect event records 150 within the scoring window and establish a directed graph of the production line. The directed graph of the production line represents the topology of the manufacturing production line 100 that allows the actual flow of workpieces 140, and includes a set of nodes and a set of adjacent node connections. The set of nodes includes the equipment boundary nodes and workstation boundary nodes corresponding to the entry reference node 110, the exit reference node 120, the intermediate acquisition node 130, and the adjacent node connection set includes the directed edges from upstream nodes to downstream nodes for workpieces 140.

[0019] During implementation, the physical operation data acquisition module 200 receives data from the RFID reader 131, PLC 132, conveyor encoder 133, boundary photoelectric trigger 134, vibration sensor 135, active power acquisition module 136, and edge gateway 137. The RFID reader 131 is used to identify workpiece markings; the PLC 132 provides information on workpiece position changes, equipment start / stop status, and release status; the conveyor encoder 133 generates cumulative encoder pulses; the boundary photoelectric trigger 134 generates continuous blocking intervals; the vibration sensor 135 generates the root mean square of vibration; the active power acquisition module 136 generates active power; and the edge gateway 137 receives data reported by each node and writes it into the gateway reception time.

[0020] The local event time in event log 150 is recorded by the source node's local clock, and the gateway reception time is written by the edge gateway 137 when it receives the event. The local event time is used for relative time calculations on subsequent connections between adjacent nodes, while the gateway reception time is used to form the event candidate order and as a reference for relative clock correction. The edge gateway 137 must not overwrite the local event time with the gateway reception time because the gateway reception time is affected by network queuing, buffer retransmission, and operating system scheduling delays, and cannot be directly used as the physical time of the event occurring at the intermediate acquisition node 130 of the artifact 140.

[0021] Event log 150 includes workpiece identifier, device identifier, event type, local event time, gateway reception time, boundary photoelectric status, vibration root mean square, active power, encoder cumulative pulses, buffer count, and data source status code. The workpiece identifier is used to associate events of the same workpiece 140 at different nodes; the device identifier is used to determine the event source; the event type is used to distinguish between entry, exit, arrival, departure, processing start, processing end, release, or abnormality; the data source status code is used to indicate normal sampling, no workpiece on the device side, sensor self-test failure, gateway buffer retransmission, or message verification failure.

[0022] Event record 150, indicating a message verification failure, is deleted and does not proceed to the workpiece path binding module 210, nor to the timing diagram input or unordered aggregation input. Event record 150, indicating a gateway cache retransmission, is not deleted, but the original local event time is retained and marked as a gateway cache retransmission using the data source status code. This allows subsequent models to know the source status of the event and avoids substituting the actual occurrence time of the event with the retransmission arrival time.

[0023] In one implementation, event log 150 reports to edge gateway 137 using one of the industrial communication methods chosen from OPCUA, ModbusTCP, or EtherNet / IP. Directed graph versions of the same production line within the same scoring window should use the same field definitions. The refresh cycle of RFID reader 131 and boundary photoelectric trigger 134 is no greater than 20 milliseconds; the refresh cycle of PLC 132, conveyor encoder 133, vibration sensor 135, and active power acquisition module 136 is no greater than 100 milliseconds; and the resolution of the gateway reception time written by edge gateway 137 is no less than 1 millisecond.

[0024] Each event log 150 carries the message length, topology version identifier, device identifier, event type, local event time, gateway reception time, payload field, and cyclic redundancy check (CRC) field. The CRC field is calculated using CRC32, and the calculation covers the entire message except for the CRC field itself. If the calculated CRC32 value does not match the CRC field value carried in event log 150, the message verification is considered to have failed.

[0025] The event record 150 for gateway cache retransmission is not deleted, but the original local event time is retained and marked as gateway cache retransmission through the data source status code. When retransmitting, edge gateway 137 only updates the gateway reception time and must not rewrite the local event time, device identifier, event type, workpiece identifier, encoder cumulative pulses, and buffer count.

[0026] The scoring window is a data grouping window for a dynamic enterprise credit score. Scoring windows can be divided by fixed time length, fixed number of workpieces, or production batches, but the scoring window boundaries are only used for data grouping and not as a basis for inter-device synchronization. This limitation avoids misusing grouping boundaries as a time alignment reference for cross-device events, thus preserving the real engineering problems caused by the local clock of the intermediate acquisition node 130.

[0027] The physical operation data acquisition module 200 outputs verified event records. Figure 1 In the process, event log 150 flows from manufacturing production line 100 to scoring processing unit 400, and physical operation data acquisition module 200 outputs verified event log to workpiece path binding module 210. The arrow indicates data input and processing result transmission.

[0028] The workpiece path binding module 210 is used to generate the actual workpiece path based on the workpiece identifier, topological predecessor relationships, and synchronous physical verification. The actual workpiece path represents the confirmed actual node sequence and actual edge sequence of workpiece 140. Figure 1 In the process, the workpiece path binding module 210 receives the verified event records output by the physical operation data acquisition module 200 and outputs the actual workpiece path to the time self-consistent calculation module 220.

[0029] The workpiece path binding module 210 first defines the effective reading range of the RFID using a continuous obstruction zone formed by the boundary photoelectric trigger 134. When the same RFID reader 131 reads the same workpiece identifier multiple times within the same continuous obstruction zone, only the first verified read record is retained. This process addresses the possibility of repeated readings by the RFID reader 131 when the workpiece 140 passes through the reading / writing area, preventing the same physical passage event from being mistakenly identified as multiple node events.

[0030] The workpiece path binding module 210 then executes topological constraints based on the directed graph of the production line. For workpiece events of downstream nodes, association with unclosed events of the same identifier in the predecessor node set is only allowed; matching across non-adjacent nodes is prohibited. An unclosed event refers to an event in which the same workpiece identifier appears in an upstream node but has not yet formed a legal adjacent node connection with that downstream node. This rule ensures that event matching conforms to the directed structure of the manufacturing production line 100, preventing non-adjacent nodes from being incorrectly spliced ​​due to fluctuations in gateway reception time.

[0031] Equipment boundary RFID events also require synchronous physical corroboration. Synchronous physical corroboration includes at least one of the following: conveyor encoder displacement, PLC workpiece position change, vibration moving from the idle zone to the operating zone, or active power moving from the standby zone to the processing zone. If only the workpiece identifier is the same but there is no synchronous physical corroboration, the workpiece path binding module 210 will not directly recognize this event as a node event in the actual workpiece path. This constraint can reduce erroneous paths caused by tag misreading, bypass reading, workpiece dwell, and equipment idling.

[0032] The quantitative judgment rules for synchronous physical evidence are solidified before deployment. The judgment condition for the encoder displacement to be valid is that the cumulative pulse change of the encoder within the continuous blocking interval is not less than the minimum effective pulse change; the judgment condition for the PLC workpiece position change to be valid is that the workpiece position of PLC132 completes one effective flip within the continuous blocking interval; the judgment condition for the vibration to enter the operating area from the idle area is that the root mean square of the vibration exceeds the vibration operating threshold for no less than three consecutive sampling points; the judgment condition for the active power to enter the processing area from the standby area is that the active power exceeds the active power operating threshold for no less than three consecutive sampling points.

[0033] The vibration operation threshold and active power operation threshold are fixed during the installation and commissioning phase. The vibration operation threshold is determined by the midpoint between the 95th percentile of the vibration root mean square of the unloaded sample and the 5th percentile of the vibration root mean square of the normal processing sample; the active power operation threshold is determined by the midpoint between the 95th percentile of the active power of the standby sample and the 5th percentile of the active power of the normal processing sample. If the two intervals overlap, the sensor self-test is considered to have failed, and re-collection of installation and commissioning samples is required.

[0034] When the entry or exit reference event time is missing, the workpiece path binding module 210 marks the corresponding workpiece 140 as an incomplete path. If an intermediate node event is missing and cannot be determined by a unique predecessor, a unique successor, and the same photoelectric blocking interval, the corresponding workpiece 140 will also be marked as an incomplete path. Only when each adjacent node on a path has at least two complete workpiece event pairs connected within the current scoring window is the path marked as a time-estimateable path.

[0035] In terms of implementation sequence, the server can first group event records 150 by workpiece identifier, then form a candidate order by gateway reception time, and then use the production line topology hierarchy to eliminate reverse-order events and cross-edge events, and confirm physical passage by boundary photoelectric status, encoder cumulative pulses, workpiece position changes in PLC132, vibration root mean square or active power. Gateway reception time is only used for candidate matching and is not used as the event timeline of the artificial intelligence model.

[0036] The output of the workpiece path binding module 210 includes the actual workpiece path, complete path marker, and time-estimated path marker for each workpiece 140. The complete path marker is used to determine whether the entry reference event time and exit reference event time are complete; the time-estimated path marker is used to determine whether the actual path of the workpiece can enter the subsequent calculations of physical propagation time, edge-level relative clock residual, relative clock drift rate, and window start-point relative clock offset.

[0037] The time self-consistency calculation module 220 is used to obtain the correction edge propagation time and generate the time self-consistency coefficient of the material flow constraint process chain event. This module is the key module in this embodiment that connects the physical flow law of the manufacturing production line 100 and the solidified dual-branch artificial intelligence reasoning module 240. Figure 1 In the process, the time self-consistency calculation module 220 receives the actual path of the workpiece and outputs the correction edge propagation time to the anchored time axis and dual-input generation module 230, while outputting the material flow constraint process chain event time self-consistency coefficient to the time reliable fusion and conflict correction module 250.

[0038] For workpiece 140 marked as having a time-estimateable path within the scoring window, the time-consistent calculation module 220 calculates the physical propagation time for each adjacent node connection on the actual path of the workpiece. The physical propagation time is the time from node 140 of workpiece 140. To the node The sum of kinematic delivery time and queuing time.

[0039]

[0040] In the above formula, Indicates physical propagation time; Represents the effective distance, taken from the node. To node The calibrated distance along the conveying path; The average linear velocity is taken from the cumulative encoder pulse conversion result of the conveyor encoder 133 during the period when the workpiece 140 passes through the adjacent node connection; This indicates the number of preceding workpieces, taken from the workpiece 140 that is located in front of the adjacent node when it enters the connection and has not yet been connected by the node. The count of allowed buffers; Represents the actual release rate, taken from the node. The actual release rate is formed by continuous release events before and after workpiece 140.

[0041] The average linear velocity must be greater than zero, and the actual release rate must be greater than zero. If a valid value cannot be obtained for either the average linear velocity or the actual release rate, the workpiece event pair connected to that adjacent node will not participate in the subsequent estimation of the relative clock drift rate and the relative clock offset of the window start point. This constraint prevents distortion of physical propagation time due to encoder stop reading, abnormal buffer counts, or missing release events.

[0042] The effective distance is calibrated during the deployment phase of the manufacturing production line 100. During calibration, the distance is measured in nodes. Trigger centerline and nodes Using the trigger center line as the measurement reference, the length of the conveying trajectory between the two is measured along the actual conveying path of the workpiece 140. For straight conveying sections, the travel is calibrated using a steel ruler, laser rangefinder, or encoder. For curved conveying sections, segments are accumulated along the trajectory center line, with each segment no longer than 0.5 meters. Each adjacent node connection is measured three times. If the maximum difference between the three measurements does not exceed 0.5% of the average of the three measurements, the average of the three measurements is fixed as the effective distance; if it exceeds 0.5%, the measurement is repeated.

[0043] The average linear velocity is calculated from the cumulative encoder pulses of the conveyor encoder 133. The encoder pulse equivalent is fixed during the deployment phase for each adjacent node connection. The encoder pulse equivalent represents the conveyor path length corresponding to one encoder pulse, in meters per pulse. For the process of workpiece 140 passing through adjacent node connections, the starting encoder cumulative pulses upon entering the adjacent node connection, the ending encoder cumulative pulses upon leaving the adjacent node connection, and the corresponding encoder timing difference are taken to calculate the average linear velocity.

[0044]

[0045] In the above formula, Indicates average linear velocity; Indicates the connection between adjacent nodes encoder pulse equivalent; This indicates the accumulated encoder pulses at the starting point when workpiece 140 enters the adjacent node connection; This indicates the cumulative encoder pulses at the endpoint when workpiece 140 leaves the adjacent node connection; This indicates the encoder timing difference, in seconds. The average linear speed is considered invalid if the encoder timing difference is less than 1 millisecond, the cumulative pulse change of the encoder is zero, or the average linear speed exceeds the rated linear speed range of the equipment.

[0046] The number of preceding workpieces is determined at the moment workpiece 140 enters the connection of the adjacent node. If the buffer photoelectric counter directly outputs the buffer count, then the number of preceding workpieces is the number of workpieces that entered before workpiece 140 and have not yet been connected to the node. The buffer count for release; if the buffer photoelectric counter outputs entry and exit events, then the number of preceding workpieces is equal to the number of adjacent nodes connected before the entry time that have not yet generated a node. The number of workpieces for which a release event occurred. If two count sources exist simultaneously and the difference is greater than 1, then mark the buffer count as abnormal.

[0047] The actual release rate is determined by the node. The continuous release events before and after workpiece 140 are formed. Specifically, workpiece 140 at node... Centered on the release event, select no more than 5 release events forward and no more than 5 release events backward to form a release event set; when the number of valid release events in the release event set is not less than 3, calculate the actual release rate according to the following formula.

[0048]

[0049] In the above formula, Indicates the actual release rate; This indicates the number of valid release events in the release event set; Indicates the local event time of the earliest released event in the released event set; This indicates the local event time of the latest released event in the set of released events; This indicates protection against zero positive numbers. If the number of valid release events is less than 3, the latest release event is not later than the earliest release event, or the actual release rate exceeds the node's limit... If the rated release rate of the equipment is below the upper limit, the actual release rate is deemed invalid.

[0050] The significance of physical propagation time lies in incorporating the actual variable-speed transport, congestion buffering, and downstream release rhythm into the temporal interpretation of adjacent node connections. If physical propagation time is not calculated, and nodes are directly compared... and nodes If the local event time difference is too large, the actual buffer wait may be mistaken for a node clock error, thus contaminating the timing graph input.

[0051] After obtaining the physical propagation time, the time self-consistency calculation module 220 subtracts the physical propagation time from the local event time difference to obtain the edge-level relative clock residual. The edge-level relative clock residual mainly reflects the clock offset between nodes, the relative clock drift rate, and the impact of abnormal transmission.

[0052]

[0053] In the above formula, This represents the relative clock residual at the edge level; Represents a node Local event time; Represents a node Local event time; This represents the physical propagation time. Since the physical propagation time already includes transport motion and queuing, the edge-level relative clock residual is more suitable for relative clock correction.

[0054] Relative clock drift rate represents the node Relative to node The clock drift rate. The time self-consistent calculation module 220 uses the relative clock residuals of the edge level that change with the timing of the gateway reception of multiple complete workpiece event pairs connected to the same adjacent node, and uses an anti-outlier statistical method to determine the relative clock drift rate.

[0055]

[0056] In the above formula, Indicates the relative clock drift rate; and This represents two different complete workpieces connected by the same adjacent node; and This indicates the midpoint of the gateway reception time for the corresponding workpiece event pair; and This represents the relative clock residual at the corresponding edge level. The median slope, as an anti-outlier statistical method, can suppress outlier residuals caused by buffer retransmissions, wireless retransmissions, network queuing, and operating system scheduling delays.

[0057]

[0058] In the above formula, Indicates the midpoint of the gateway's reception time; Represents a node The gateway reception time for the corresponding event; Represents a node The gateway reception time for the corresponding event. If either of the two events lacks a gateway reception time, then the event pair is not included in the relative clock drift rate estimation.

[0059] The number of complete artifact event pairs used for relative clock drift rate estimation must be no less than 5. If fewer than 5 complete artifact event pairs meet the condition, the adjacent node connection is marked as time-unpredictable within the current scoring window. If the difference between the gateway reception times of two artifact event pairs is less than 1 second, these two artifact event pairs are not used for median slope calculation. Before median slope calculation, artifact event pairs with invalid physical propagation times, unresolved gateway buffer retransmission status, abnormal buffer counts, and failed message verification are removed.

[0060] After determining the relative clock drift rate, the time self-consistency calculation module 220 estimates the relative clock offset of the window start point. The relative clock offset of the window start point represents the node at the start of the scoring window. Relative to node Clock offset.

[0061]

[0062] In the above formula, This indicates the offset of the window's starting point relative to the clock. Indicates the midpoint of the gateway's reception time; This indicates the gateway time at the start of the scoring window. By first estimating the relative clock drift rate and then estimating the relative clock offset at the start of the window, the time-varying frequency error can be separated from the fixed phase error.

[0063] If the relative clock drift rate cannot be estimated, the relative clock offset at the window start point does not use the historical value of the previous scoring window, and the connection of the current adjacent nodes is directly marked as time unpredictable. This process avoids the error of historical clock state being carried over to the current scoring window.

[0064] The time self-consistency calculation module 220 calculates the corrected edge propagation time after eliminating the effects of relative clock offset and relative clock drift rate at the window start point. The corrected edge propagation time is neither the gateway reception time difference nor the interpolation time on the fixed sampling grid, but rather the edge propagation time after clock error correction of adjacent node connections while retaining the original local event time.

[0065]

[0066] In the above formula, Indicates the corrected side propagation time; and This indicates the local event time recorded by two adjacent nodes for the same workpiece 140; This indicates the offset of the window's starting point relative to the clock. Indicates the relative clock drift rate; Indicates the midpoint of the gateway's reception time; This indicates the gateway time at the start of the scoring window.

[0067] Subsequently, the time self-consistent calculation module 220 calculates the end-to-end closure residual using the entry reference event time and the exit reference event time. The entry reference event time is recorded by the entry reference node 110, and the exit reference event time is recorded by the exit reference node 120. Since both are in the same hardware clock domain, they can serve as an end-to-end time reference that is not affected by the local clock of the intermediate acquisition node 130.

[0068]

[0069] In the above formula, This represents the end-to-end closure residual; Indicates the entry reference event time; Indicates the export baseline event time; Indicates the actual path of the workpiece; Indicates the corrected side propagation time; This indicates a positive number that is protected against zero. The larger the end-to-end closure residual, the more likely that the end-to-end additive closure relationship cannot be satisfied even after edge-by-edge time correction, which may indicate missing events, mismatches, clock anomalies, encoder anomalies, buffer count anomalies, or transmission link anomalies.

[0070] Before calculating the end-to-end closure residual, the time-consistent calculation module 220 first determines whether the exit reference event time is later than the inlet reference event time. When the condition is met... When the inlet and outlet events of workpiece 140 are marked as end-to-end baseline anomalies, they are not included in the calculation of the end-to-end closure residual and material flow constraint process chain event time self-consistency coefficient, and are included in the unpredictable samples of the time-predictable coverage ratio. When the conditions are met... Only then should the end-to-end closure residual be calculated according to the above formula.

[0071] Zero-prevention positive numbers are positive numbers with a time dimension, measured in seconds, and are fixed before system deployment. Zero-prevention positive numbers are used to prevent the time denominator from being zero, and their values ​​must not be interpreted as dimensionless positive numbers. In one implementation, the zero-prevention positive number is fixed according to the following formula.

[0072]

[0073] In the above formula, Indicates a positive number that is protected against zero, in seconds; This indicates the number of seconds that corresponds to 1 millisecond; This represents the median physical propagation time in the deployment and debugging samples. If the deployment and debugging samples cannot yet generate physical propagation time, a positive value of 1 millisecond is initially used to prevent zero propagation, and the value is re-fixed after at least 100 complete artifact event pairs are generated.

[0074] The time self-consistency coefficient of the material flow constrained process chain event is jointly determined by the time-predictable coverage ratio, end-to-end closure residual, and edge-by-edge physical propagation deviation. The time-predictable coverage ratio represents the proportion of complete and time-predictable workpieces at the inlet and outlet. The end-to-end closure residual represents the closure difference between the end-to-end elapsed time in the same hardware clock domain and the accumulated correction edge propagation time along the actual path of the workpiece. The edge-by-edge physical propagation deviation represents the deviation between the correction edge propagation time and the physical propagation time.

[0075]

[0076] In the above formula, This represents the time consistency coefficient of the material flow constraint process chain events; The coverage ratio can be estimated by indicating the time frame; This represents the end-to-end closure residual; Indicates the actual path of the workpiece; Indicates the corrected side propagation time; Indicates physical propagation time; This indicates a zero positive number. If the number of complete workpieces at the entrance and exit of the current scoring window is zero, then the self-consistency coefficient of the material flow constraint process chain event time is directly set to zero.

[0077]

[0078] In the above formula, The coverage ratio can be estimated by indicating the time frame; This indicates the number of workpieces whose time is estimable and whose inlet and outlet are complete within the scoring window; This indicates the number of complete workpieces at the entrance and exit within the scoring window. If If it is zero, then the self-consistency coefficient of the material flow constraint process chain event time is directly set to zero.

[0079] because , and All are time quantities. The edge-to-edge physical propagation deviation is a dimensionless quantity, the exponential term is a dimensionless quantity, and the material flow constraint process chain event time self-consistency coefficient is a dimensionless quantity.

[0080] In one implementation, a first self-consistency coefficient threshold and a second self-consistency coefficient threshold are fixed before deployment. The first self-consistency coefficient threshold is set to 0.8, and the second self-consistency coefficient threshold is set to 0.5. When the self-consistency coefficient of the material flow constrained process chain event time is not less than the first self-consistency coefficient threshold, the event time is determined to be reliable; when the self-consistency coefficient of the material flow constrained process chain event time is less than the second self-consistency coefficient threshold, the event time is determined to be unreliable; when the self-consistency coefficient of the material flow constrained process chain event time is between the two, continuous fusion is performed according to the formula without hard switching.

[0081] The self-consistency coefficient of the material flow constraint process chain event time is not the same as the data integrity rate, the synchronous message round-trip time, or the confidence level of the artificial intelligence model. The data integrity rate only indicates whether the data exists, the synchronous message round-trip time only indicates the round-trip delay at the communication layer, and the confidence level of the artificial intelligence model reflects the distribution of the model output; the self-consistency coefficient of the material flow constraint process chain event time comes from the material flow structure, physical propagation time, and end-to-end closure relationship of the real workpiece 140 within the current scoring window.

[0082] The anchored time axis and dual-input generation module 230 are used to generate a timing diagram input and an unordered aggregate input based on the correction side propagation time. For example... Figure 1 As shown, the anchored time axis and dual-input generation module 230 receive the corrected side propagation time output by the time self-consistent calculation module 220, and output the timing graph input and the unordered aggregation input to the solidified dual-branch artificial intelligence inference module 240, respectively.

[0083] Before generating the timing diagram input, the anchoring time axis and the dual-input generation module 230 use the entry reference event time as the time starting point of the workpiece 140, and accumulate the correction edge propagation time along the actual path of the workpiece to obtain the physical anchoring event time.

[0084]

[0085]

[0086] In the above formula, and Indicates the physical anchoring event time; Indicates the correction edge propagation time; when node When the ingress reference node is 110, Equal to the entry baseline event time Physically anchoring event times means that the events at intermediate acquisition node 130 no longer directly depend on the original order of local event times, but rather on the order after material flow constraint verification and clock correction.

[0087] The timing diagram input uses only workpiece 140 with a time-estimated path. The timing event vector in the timing diagram input is fixed and includes equipment identification code, event type code, adjacent anchor event interval, vibration root mean square, active power, buffer count, average linear velocity, data source status code, and time validity mask. Both the equipment identification code and event type code use a pre-deployed one-bit valid code; the time validity mask is 0 or 1, where 1 indicates that the event time can be used for timing sequencing, and 0 indicates that the event time is only used as a missing or invalid status input. This setting allows timing branch 241 to simultaneously read the process sequence, equipment operating status, and data validity status.

[0088] The sequence diagram input sorts events by their physical anchor time and connects adjacent nodes according to the directed graph of the production line. When multiple events occur at the same physical anchor time, they are sorted sequentially according to the deterministic lexicographical order of the production line topology hierarchy and equipment identifier; random sorting is prohibited. Missing events are not filled with mean, zero, or linear interpolation; instead, the missing status is represented by the data source status code and a time validity mask.

[0089] The unordered aggregation input is generated from event log 150, but the local event time, gateway reception time, event sequence number, and physically anchored event time are removed. The unordered aggregation input includes equipment statistics, including effective running duration, workpiece throughput count, vibration energy quantile, active power quantile, buffer peak value, and data validity percentage. This design ensures that unordered aggregation branch 242 is independent of cross-equipment time sequence and is suitable for scoring windows with low self-consistency coefficients in material flow-constrained process chain events or when the timing diagram input is empty.

[0090] When no time-estimated path exists for the current scoring window, the anchored timeline and dual-input generation module 230 sets the timing diagram input to empty and sets a window-level time invalidation flag. This process does not delete the scoring window, but rather provides deterministic degradation conditions for the solidified dual-branch AI inference module 240.

[0091] The window-level time invalidation flag is a boolean field. A value of 1 indicates that the timing diagram input is empty and triggers deterministic degradation, while a value of 0 indicates that the timing diagram input is available.

[0092] The fixed dual-branch AI inference module 240 includes a temporal branch 241 and an unordered aggregation branch 242. This module is used to output the temporal branch risk probability and the unordered aggregation branch risk probability based on the temporal graph input and the unordered aggregation input, and to keep the two branch risk probabilities consistent when the temporal graph input is empty.

[0093] In this embodiment, the solidified dual-branch AI model consists of a temporal branch 241 and an unordered aggregation branch 242, and is executed by the solidified dual-branch AI inference module 240. The parameters of the solidified dual-branch AI model are solidified after training and validation, and are not updated online during the scoring process based on the dynamic enterprise credit score, enterprise credit rating, or robust risk probability of the current scoring window.

[0094] Temporal branch 241 outputs the risk probability of the temporal branch according to the input of the temporal graph. Temporal branch 241 includes, in sequence, an event linear encoding layer, a one-way gated recurrent unit, a device topology graph attention layer, a global average pooling layer, and a sigmoid output layer. The event linear encoding layer maps the temporal event vector to a 64-dimensional latent vector; the one-way gated recurrent unit has a hidden dimension of 64 and processes events strictly according to the historical direction of physically anchored event time; the device topology graph attention layer propagates features only on the actual adjacent node connections in the directed graph of the production line.

[0095] Before entering the device topology attention layer, the adjacent anchor event interval, vibration root mean square, active power, buffer count, and average linear velocity in the time-series event vector are first standardized according to their corresponding feature dimensions, and then mapped by the event linear encoding layer into dimensionless node latent vectors. and node hidden vectors All are located in the dimensionless hidden space. , and They are all located in the dimensionless hidden space, therefore the attention score It is a dimensionless quantity, which satisfies the requirements for subsequent exponential function calculations.

[0096]

[0097]

[0098]

[0099] In the above formula, Represents a node With nodes In time location Attention score; , and This represents the matrix parameters that are fixed after training; and Represents the hidden vector of a node; Indicates hidden dimensions; Represents nodes in a directed graph of production lines. The set of downstream adjacent nodes; Indicates attention weight; This represents the attention layer output of the device topology graph; This represents the ReLU activation function.

[0100] The unordered aggregation branch 242 outputs the risk probability of the unordered aggregation branch based on the device statistics. The unordered aggregation branch 242 sequentially includes a device statistics standardization layer, a first fully connected layer, a second fully connected layer, a device-dimensional average pooling layer, and a Sigmoid output layer. The hidden dimension of both the first and second fully connected layers is 64, and the activation function is ReLU. The device statistics standardization layer uses the training set mean and standard deviation; after training, the training set mean and standard deviation are fixed and cannot be updated during scoring.

[0101]

[0102] In the above formula, This represents standardized equipment statistics. Indicates the first The first sample Statistical data for each device; Indicates the training set number The mean of the statistics for each device; Indicates the training set number Standard deviation of individual device statistics; Indicates the first Standardized zero-prevention quantity of individual equipment statistics. The standardized zero-prevention quantity of the statistical quantity of each device has the same characteristics as the first The statistics of each device have the same dimensions, therefore they can be compared with the training set. The standard deviations of the statistics for each device are summed.

[0103]

[0104] In the above formula, Indicates the first Standardized zero-prevention quantity of individual equipment statistics; Indicates the first Reference standard deviation of individual device statistics; Indicates the first The minimum zero-prevention quantity for each device's statistics. Reference standard deviation of the statistics for each device and the first The minimum zero-prevention value of each device statistic is fixed according to the unit of the corresponding device statistic before training, and written into memory 310 together with the standardized parameters of the training set. If the first If the statistical quantity of the device is the average linear velocity, then the... The unit of the standardized zero-prevention quantity of the equipment statistics is meters per second; if the first... If the statistical quantity of the device is active power, then the first device... The unit for the standardized zero-prevention quantity of each device's statistical quantity is watts; if the first... If the device statistic is the buffer count, then the th device statistic is... The standardized unit for zero-prevention statistical quantity of each device is the piece.

[0105] Both branches predict the same target event: whether, within the observation period following the scoring window, manufacturing line 100 experiences a continuous material flow interruption lasting longer than the minimum recoverable period specified in the equipment maintenance manual. This tag is determined by RFID end-to-end workpiece flow, PLC operating status, and equipment physical operation records, without using financial default amounts, personal credit data, or biometric data.

[0106] The observation period is one scoring window length; the minimum recoverable period is the minimum recovery time recorded in the equipment maintenance manual for 100 bottleneck devices of the manufacturing production line. If the equipment maintenance manual does not record it, the 5th percentile of the recovery record is solidified before deployment through no less than 20 downtimes.

[0107] The temporal branch 241 and the unordered aggregation branch 242 are trained using the same historical window and the same physical interruption label. The training set, validation set, and test set are split chronologically, and windows associated with the same observation period must not cross sets. After training and validation are completed, the parameters of the fixed dual-branch AI inference module 240 must not be updated online during the scoring run.

[0108] The solidified dual-branch artificial intelligence model is trained using a binary cross-entropy loss function.

[0109]

[0110] In the above formula, Indicates training loss; Indicates the number of samples in the training batch; Indicates the first Physical interruption label for each sample; This represents the risk probability output by the time-series branch 241 or the unordered aggregation branch 242. The training set, validation set, and test set are split sequentially in a ratio of 7:2:1, and scoring windows associated with the same observation period must not cross sets.

[0111] In one implementation, the training epochs are capped at 100, the batch size is 128, the optimizer is Adam, the initial learning rate is 0.001, and early stopping occurs when the validation set loss does not decrease for 10 consecutive epochs. After training, the model parameters with the minimum validation set loss are selected as the fixed bi-branch AI model parameters, and these model parameters, along with the training set normalized parameters, the event type encoding table, the device identifier encoding table, and the directed graph version of the production line, are written into memory 310.

[0112] When the time series graph input is not empty, time series branch 241 outputs the time series branch risk probability, and unordered aggregation branch 242 outputs the unordered aggregation branch risk probability. When the time series graph input is empty, the unordered aggregation branch risk probability is first obtained by unordered aggregation branch 242 based on the unordered aggregation input, and then the time series branch risk probability is deterministically set to equal the unordered aggregation branch risk probability, without using random values, the risk value of the previous scoring window, or manually defaulted credit values.

[0113]

[0114] In the above formula, This represents the probability of time-series branch risk. This represents the probability of unordered aggregation branch risk. This formula is only executed when the time series graph input is empty, to ensure that the subsequent reliable time fusion and conflict correction module 250 obtains a unique and reproducible robust risk probability.

[0115] Both the temporal branch risk probability and the unordered aggregation branch risk probability are limited to open intervals by numerical truncation boundaries fixed during training. Numerical truncation is only used for numerical stability in subsequent logarithmic calculations and does not change the fact that both branches predict the same physical interruption label.

[0116] In one implementation, positive numbers are probabilistically truncated. The probability of risk of time-series branching and the probability of risk of unordered aggregation branching are both truncated according to the following formula.

[0117]

[0118]

[0119] The time-based reliable fusion and conflict correction module 250 is used to generate robust risk probabilities based on the time self-consistency coefficient of material flow-constrained process chain events, the probability of time-series branch risks, and the probability of disordered aggregation branch risks. For example... Figure 1 As shown, the module simultaneously receives the timing branch risk probability output by timing branch 241, the disordered aggregation branch risk probability output by disordered aggregation branch 242, and the material flow constraint process chain event time self-consistency coefficient output by time self-consistency calculation module 220.

[0120] The temporal trusted fusion and conflict correction module 250 first performs temporal trusted fusion in the logarithmic odds domain to obtain the trusted fusion risk probability.

[0121]

[0122]

[0123]

[0124] In the above formula, Indicates the probability of trustworthy fusion risk; This represents the time consistency coefficient of the material flow constraint process chain events; Indicates the probability of time-series branch risk; This represents the probability of risk in disordered aggregation branches. Log-odds domain fusion combines the strengths of evidence from two branches on the same additive scale, rather than a simple probability average.

[0125] When the self-consistency coefficient of the material flow-constrained process chain event time increases, the risk probability contribution of the timing branch increases, indicating that the multi-device event time in the current scoring window is more in line with the material flow propagation and end-to-end closure rules. The system can make greater use of the ability of timing branch 241 to identify continuous process changes, buffer accumulation, and equipment linkage. When the self-consistency coefficient of the material flow-constrained process chain event time decreases, the risk probability contribution of the disordered aggregation branch increases, indicating that the cross-device event time may be affected by clock drift, buffer retransmission, transmission delay, event mismatch, or event missing. The system should rely more on device statistics that do not include cross-device time sequence.

[0126] After obtaining the reliable fusion risk probability, the time reliable fusion and conflict correction module 250 performs branch conflict correction, and performs conservative risk correction to obtain robust risk probability when the self-consistency coefficient of the material flow constraint process chain event time decreases and the branch risk probability divergence increases.

[0127]

[0128] In the above formula, Indicates the probability of robust risk; Indicates the probability of trustworthy fusion risk; This represents the time consistency coefficient of the material flow constraint process chain events; Indicates the probability of time-series branch risk; This represents the probability of risk from unordered aggregation branches. This formula must be executed after the probability of trusted fusion risk has been calculated; the order cannot be reversed.

[0129] When the self-consistency coefficient of the material flow-constrained process chain event time is not less than the first self-consistency coefficient threshold or the difference in risk probabilities between the two branches is not greater than 0.05, the branch conflict correction has a small additional impact on the probability of reliable fusion risk. When the self-consistency coefficient of the material flow-constrained process chain event time is less than the second self-consistency coefficient threshold and the difference in risk probabilities between the two branches is greater than 0.2, the system considers this difference as a technical risk that the cross-device time structure may contaminate the model judgment, and improves the robust risk probability through conservative risk correction.

[0130] When there is no valid timing input, the timing branch risk probability equals the unordered aggregated branch risk probability, the branch risk probability divergence is zero, the branch conflict correction term is automatically zero, and the robust risk probability uniquely degenerates to the result corresponding to the unordered aggregated branch risk probability. This processing maintains the determinism of the scoring results during equipment maintenance, partial sensor failures, or missing paths with unpredictable timing.

[0131] The order calibration and credit scoring output module 260 is used to output dynamic corporate credit scores and corporate credit ratings based on robust risk probabilities. For example... Figure 1As shown, the time-reliable fusion and conflict correction module 250 outputs robust risk probability to the order-preserving calibration and credit scoring output module 260, and the order-preserving calibration and credit scoring output module 260 outputs dynamic enterprise credit score 610 and enterprise credit rating 620 to the credit output 600.

[0132] The system consists of a sequence calibration and credit scoring output module 260, a credit output module 600, a dynamic enterprise credit score module 610, and an enterprise credit rating module 620. Figure 1 The fixed sequence of labels in the output link on the right. Among them, 260 corresponds only to the sequence calibration and credit scoring output module 260, 600 corresponds only to the credit output 600, 610 corresponds only to the dynamic enterprise credit score 610, and 620 corresponds only to the enterprise credit rating 620; no final submitted figure may use 260 for enterprise credit rating, nor may it use 620 for sequence calibration and credit scoring output module 260.

[0133] The order-preserving calibration and credit scoring output module 260 first inputs the robust risk probability into a monotonic non-decreasing order-preserving regression calibrator trained on an independent validation set, and then obtains the calibrated enterprise physical operation continuity risk probability.

[0134]

[0135] In the above formula, This indicates the probability of physical operational continuity risk for the enterprise after calibration. Indicates the probability of robust risk; This indicates a solidified order-preserving regression calibrator. The solidified order-preserving regression calibrator maintains the model's risk ranking and converts the output of the solidified dual-branch AI inference module 240 into a probabilistic scale consistent with the historical physical interruption frequency.

[0136] The solidified order-preserving regression calibrator can only be built using physical interruption tags defined with the same type as the solidified dual-branch AI inference module 240, and must not include financial default amounts, personal credit data, or human data. The solidified order-preserving regression calibrator must not be updated online during the scoring process.

[0137] The solidified order-preserving regression calibrator is trained using the pooled adjacent violation algorithm. During training, the independent validation set samples are first arranged in ascending order of robust risk probability. If the mean of physical interruption labels of adjacent groups does not satisfy a monotonically non-decreasing relationship, the adjacent groups are merged, and the mean of physical interruption labels within the merged group is used as the calibrated enterprise physical operational continuity risk probability for that group. This merging is repeated until all groups satisfy a monotonically non-decreasing relationship. The robust risk probability interval, sample size, and calibrated enterprise physical operational continuity risk probability of each final group are written into memory 310 and solidified.

[0138] During the scoring run, if the robust risk probability falls within the robust risk probability range of a certain fixed group, the calibrated enterprise physical operational continuity risk probability corresponding to that group is output; if the robust risk probability is less than the lower bound of the minimum fixed group, the calibrated enterprise physical operational continuity risk probability of the minimum fixed group is output; if the robust risk probability is greater than the upper bound of the maximum fixed group, the calibrated enterprise physical operational continuity risk probability of the maximum fixed group is output. The fixed order-preserving regression calibrator must not be updated based on the current scoring window results during the scoring run.

[0139] The validation criteria for the solidified order-preserving regression calibrator are that the Brier score on the independent test set is not higher than the Brier score corresponding to the uncalibrated robust risk probability, and the maximum absolute difference between the predicted probability after binning and the actual physical interruption frequency is not greater than 0.15.

[0140] Subsequently, the Sequence Calibration and Credit Score Output Module 260 maps the calibrated probability of enterprise physical operational continuity risk into a dynamic enterprise credit score.

[0141]

[0142] In the above formula, This indicates the dynamic corporate credit score; Indicates the lower bound of the credit score; Indicates the upper limit of the credit score; This represents the probability of risk to the physical continuity of enterprise operations after calibration. The lower and upper bounds of the credit score are fixed before system deployment, and the upper bound of the credit score is greater than the lower bound.

[0143] Enterprise credit rating ranges are fixed before system deployment. These ranges are non-overlapping and provide complete coverage. A single dynamic corporate credit score can only fall into one corporate credit rating. If a dynamic corporate credit score hits a credit rating boundary, the corporate credit rating is uniquely determined according to the pre-defined closed-open interval rule or the final closed interval rule.

[0144] In one implementation, the lower bound of the credit score is 0, and the upper bound is 100. The enterprise credit rating range is divided into four levels—D, C, B, A, and AA—based on the dynamic enterprise credit score from low to high, with D being the lowest. Grade C is Grade B is Grade A is Grade AA is The aforementioned ranges are fixed before system deployment, and the same dynamic enterprise credit score can only fall into one enterprise credit rating.

[0145] This step ultimately outputs only the dynamic enterprise credit score and enterprise credit rating for the current scoring window. Model latent vectors, window start-point relative clock offset, relative clock drift rate, device parameters, intermediate probabilities, and data quality status are not included in the final business output. This limitation protects enterprise device privacy and the status of the data acquisition link while ensuring the clarity of the final business output.

[0146] At the start of a scoring window, the entry reference node 110, intermediate acquisition node 130, and exit reference node 120 on the manufacturing production line 100 continuously generate event logs 150. The entry reference node 110 and the exit reference node 120 are provided with a time reference by the same hardware timing gateway 160, and the RFID reader 131, PLC 132, conveyor encoder 133, boundary photoelectric trigger 134, vibration sensor 135, active power acquisition module 136, and edge gateway 137 in the intermediate acquisition node 130 generate data according to their respective sampling methods.

[0147] The physical operations data acquisition module 200 receives event records 150 within the scoring window, deletes event records 150 that failed message verification, retains the original local event times of event records 150 that were retransmitted from the gateway cache, and marks them with data source status codes. Subsequently, the physical operations data acquisition module 200 establishes a directed graph of the production line based on the current production line topology version and outputs the verified event records to the workpiece path binding module 210.

[0148] The workpiece path binding module 210 defines the effective reading range of RFID using a continuous obstruction interval formed by the boundary photoelectric trigger 134. Within the same continuous obstruction interval, only the first verified reading record for the same workpiece identifier is retained. Subsequently, the workpiece path binding module 210 associates the workpiece events of downstream nodes only with unclosed events of the same identifier in the predecessor node set, and requires that the equipment boundary RFID events have at least one synchronous physical corroboration.

[0149] After path binding is completed, the time self-consistency calculation module 220 calculates the physical propagation time, edge-level relative clock residual, relative clock drift rate, window start-point relative clock offset, correction edge propagation time, and end-to-end closure residual for the time-estimated path, ultimately obtaining the time self-consistency coefficient of the material flow constraint process chain event. This coefficient, along with the correction edge propagation time, is passed to subsequent modules. The correction edge propagation time enters the anchor time axis and dual-input generation module 230, while the time self-consistency coefficient of the material flow constraint process chain event is connected to the time reliable fusion and conflict correction module 250.

[0150] The anchoring time axis and dual-input generation module 230 uses the entry reference event time as the starting point and accumulates the correction edge propagation time along the actual workpiece path to generate the physical anchoring event time. Based on the physical anchoring event time, the anchoring time axis and dual-input generation module 230 generates the timing diagram input; based on the equipment statistics in the event record 150, the anchoring time axis and dual-input generation module 230 generates the unordered aggregated input.

[0151] The fixed dual-branch AI inference module 240 receives a timing graph input and an unordered aggregation input. When the timing graph input is not empty, timing branch 241 outputs the timing branch risk probability, and unordered aggregation branch 242 outputs the unordered aggregation branch risk probability. When the timing graph input is empty, unordered aggregation branch 242 first outputs the unordered aggregation branch risk probability, and then deterministically makes the timing branch risk probability equal to the unordered aggregation branch risk probability.

[0152] The Time-Based Reliable Fusion and Conflict Correction Module 250 first generates a reliable fusion risk probability based on the material flow constraint process chain event time self-consistency coefficient, the time-series branch risk probability, and the disordered aggregation branch risk probability, and then performs branch conflict correction to obtain a robust risk probability. The Order Preservation Calibration and Credit Score Output Module 260 converts the robust risk probability into a calibrated enterprise physical operation continuity risk probability, then maps it to a dynamic enterprise credit score, and outputs the enterprise credit rating.

[0153] When multiple abnormal states occur simultaneously within the same scoring window, the system processes them in the following order: message verification failure, sensor self-test failure, incomplete path, unpredictable time, empty timing diagram input, and model numerical boundary hit. Event record 150 for message verification failure is directly deleted; the acquisition entity for sensor self-test failure does not provide synchronous physical evidence within the current scoring window; workpiece 140 with an incomplete path does not participate in end-to-end closure residual calculation; paths with unpredictable time do not participate in the numerator of the time-predictable coverage ratio in the time self-consistency coefficient of the material flow constraint process chain event; deterministic degradation is performed when the timing diagram input is empty; and probabilistic truncation is performed when the model numerical boundary hits.

[0154] Invalid average linear velocity, invalid actual release rate, abnormal buffer count, gateway cache retransmission, and sensor self-test failure all enter the timing diagram input or unordered aggregated input through the data source status code or time valid mask. If all inlet and outlet complete workpieces in a certain scoring window cannot form a time-estimated path, the time self-consistency coefficient of the material flow constraint process chain event is zero, the timing diagram input is empty, and the solidified dual-branch artificial intelligence reasoning module 240 deterministically makes the timing branch risk probability equal to the unordered aggregated branch risk probability.

[0155] Under normal steady-state conditions, workpiece 140 flows sequentially along the inlet reference node 110, intermediate acquisition node 130, and outlet reference node 120. Boundary photoelectric triggers 134 form clear and continuous blocking intervals, the conveyor encoder 133 records stable encoder cumulative pulses, the PLC 132 records the workpiece's position changes, and the vibration sensor 135 and the active power acquisition module 136 display device move from the standby area to the operating area or processing area.

[0156] Under this condition, the workpiece path binding module 210 can form a complete actual workpiece path, and the time self-consistency calculation module 220 can obtain the effective physical propagation time, relative clock drift rate, window start-point relative clock offset, and correction edge propagation time. The end-to-end elapsed time between the entry reference event time and the exit reference event time is basically closed with the edge-by-edge accumulated correction edge propagation time. The end-to-end closure residual is small, and the edge-by-edge physical propagation deviation is also small. Therefore, the material flow constraint process chain event time self-consistency coefficient is not less than the first self-consistency coefficient threshold.

[0157] When the time self-consistency coefficient of the material flow constraint process chain event is not less than the first self-consistency coefficient threshold, the time reliable fusion and conflict correction module 250 improves the risk probability contribution of the time branch. At this time, the time branch 241 can use the time sequence diagram input to identify the impact of continuous process changes, buffer accumulation and equipment linkage on the risk of physical operation continuity, and the dynamic enterprise credit score can more fully reflect the real operating status of the manufacturing production line 100.

[0158] In the case where the intermediate acquisition node 130 generates a relative clock drift rate, the local event time of an RFID reader 131 or edge gateway 137 may gradually shift relative to the upstream node due to temperature changes, crystal oscillator errors, or equipment aging. If multiple device events are directly spliced ​​according to the local event time, digital time distortion may occur where the downstream node event precedes the upstream node event.

[0159] In this embodiment, the time self-consistent calculation module 220 first subtracts the actual transport motion and queuing waiting time from the physical propagation time, then estimates the relative clock drift rate based on the edge-level relative clock residuals of multiple complete workpiece event pairs, and estimates the relative clock offset of the window start point. By correcting the edge propagation time, the system eliminates the influence of the relative clock drift rate and the relative clock offset of the window start point on the edge propagation time.

[0160] If the corrected inlet and outlet reference event times still satisfy the end-to-end closure relationship with the sum of the edge-by-edge correction propagation time, the self-consistency coefficient of the material flow constrained process chain event time remains high, and the system will not misjudge the numerical misalignment of local event times as a sudden change in production rhythm. If the corrected end-to-end closure residual makes the self-consistency coefficient of the material flow constrained process chain event time less than the second self-consistency coefficient threshold, the system reduces the self-consistency coefficient of the material flow constrained process chain event time and increases the contribution of disordered aggregation branch risk probability in time-reliable fusion.

[0161] In the context of gateway buffer retransmission and event loss, edge gateway 137 may experience delays in the arrival of some event records 150 due to communication link congestion or wireless retransmission; some sensors may experience localized failures, resulting in the loss of ingress reference event time, egress reference event time, or intermediate node events.

[0162] For event record 150 that is cached and retransmitted by the gateway, the physical operation data acquisition module 200 retains the original local event time and marks it with the data source status code. For event record 150 that fails message verification, the physical operation data acquisition module 200 deletes the event record 150. For workpiece 140 with missing ingress or egress reference event time, the workpiece path binding module 210 marks it as an incomplete path; for scoring windows that cannot form a time-estimated path, the anchor time axis and dual-input generation module 230 sets the timing diagram input to empty.

[0163] When the time sequence graph input is empty, the solidified dual-branch AI inference module 240 deterministically sets the time sequence branch risk probability equal to the disordered aggregate branch risk probability. Thus, the branch conflict term in the time-credible fusion and conflict correction module 250 is zero, and the robust risk probability uniquely degenerates to the result corresponding to the disordered aggregate branch risk probability. In this condition, the system does not use random risk values, the risk value from the previous scoring window, or manually defaulted credit values.

[0164] In buffer congestion and real variable speed conveying conditions, the actual elapsed time at adjacent node connections in manufacturing production line 100 may be prolonged due to slower release of downstream equipment, backlog of workpieces 140 in the buffer, or changes in conveying speed. Without incorporating effective distance, average linear velocity, number of preceding workpieces, and actual release rate, the system may misinterpret this delay as node clock error.

[0165] This implementation incorporates both kinematic transport time and queuing time into the physical propagation time. The average linear velocity is derived from the transport encoder 133, the number of preceding workpieces is derived from the buffer count, and the actual release rate is derived from the node release event. Therefore, actual variable-speed transport and buffer congestion are reflected in the physical propagation time, without directly causing abnormalities in the relative clock residual at the edge level.

[0166] If the end-to-end closure residual and the edge-by-edge physical propagation deviation still ensure that the time self-consistency coefficient of the material flow constrained process chain event is not less than the first self-consistency coefficient threshold, and the time self-consistency coefficient of the material flow constrained process chain event remains high, then timing branch 241 can incorporate buffer congestion as real operational evidence into the timing diagram input. If an abnormal encoder cumulative pulse or buffer count causes the physical propagation time to be unavailable, then the workpiece event pairs connected to the corresponding adjacent nodes will not participate in the estimation of relative clock drift rate and window start point relative clock offset.

[0167] This implementation can be deployed on edge servers, industrial servers, cloud servers, or hybrid computing systems consisting of edge and cloud terminals. Figure 1 The computing device 700 includes a processor 300 and a memory 310. The memory 310 stores program instructions, production line topology version, data source status code definitions, model parameters of the fixed dual-branch AI inference module 240, a fixed sequence-preserving regression calibrator, a lower bound for credit scores, an upper bound for credit scores, and enterprise credit rating ranges. The processor 300 executes the program instructions in the memory 310 to implement the physical operations data acquisition module 200, the workpiece path binding module 210, the time self-consistent calculation module 220, the anchored time axis and dual-input generation module 230, the fixed dual-branch AI inference module 240, the time reliable fusion and conflict correction module 250, and the sequence-preserving calibration and credit score output module 260.

[0168] In terms of hardware selection, the RFID reader 131 should be able to read workpiece markings and output the local event time of the RFID reading event; the PLC 132 should be able to provide information on workpiece position changes, equipment operating status, and release events; the conveyor encoder 133 should be able to output the encoder's cumulative pulses; the boundary photoelectric trigger 134 should be able to form a continuous blocking interval; the vibration sensor 135 should be able to output the root mean square of vibration; the active power acquisition module 136 should be able to output active power; and the edge gateway 137 should be able to write the gateway reception time and retain the original local event time.

[0169] In one embodiment, an RFID reader 131 is installed on the side above the workpiece 140 passing through the node boundary, with the main lobe of the reading antenna facing the area traversed by the tag on the workpiece 140, and a reading distance of 0.1 meters to 1.5 meters; a boundary photoelectric trigger 134 is installed at the node trigger center line, with the optical axis perpendicular to the conveying direction and the positional deviation not exceeding 1% of the effective distance; a conveying encoder 133 is installed on the conveying spindle or synchronous pulley, and the encoder pulse equivalent is calibrated during the deployment phase; a vibration sensor 135 is installed on the spindle box of the processing equipment or the workstation support structure; and an active power acquisition module 136 is connected to the power supply circuit of the corresponding processing equipment, with a sampling range covering standby power to rated power.

[0170] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A dynamic enterprise credit scoring method based on an artificial intelligence model, applied to a manufacturing production line where the entry and exit reference nodes are in the same hardware clock domain and intermediate acquisition nodes retain their local clocks, characterized in that... include: The collection and scoring window contains event records with workpiece identification, local event time, and gateway reception time. The gateway reception time is used to form the event candidate order and relative clock correction reference. The actual path of the workpiece is bound based on the workpiece identification, the directed graph of the production line, and the synchronous physical evidence; the corrected edge propagation time is obtained based on the physical propagation time, the edge-level relative clock residual, the relative clock drift rate, and the window start-point relative clock offset. Based on the entry baseline event time, exit baseline event time, physical propagation time, and correction edge propagation time, the self-consistency coefficient of the material flow constraint process chain event time is generated to estimate the coverage ratio, end-to-end closure residual, and edge-by-edge physical propagation deviation. The time sequence diagram input is generated based on the correction edge propagation time, and the disordered aggregation input is generated from the event records. The time sequence branch risk probability and the disordered aggregation branch risk probability are obtained based on the two inputs using a solidified dual-branch artificial intelligence model. When the time sequence diagram input is empty, the branch risk probabilities are made consistent. Based on this coefficient, time-reliable fusion and conflict correction are performed. When the coefficient increases, the contribution of time-series branch risk probability is increased. When the coefficient decreases, the contribution of disordered aggregation branch risk probability is increased. When the coefficient decreases and the divergence of branch risk probability increases, conservative risk correction is performed to obtain robust risk probability. The robust risk probability is mapped to the dynamic corporate credit score and corporate credit rating of the scoring window.

2. The dynamic enterprise credit scoring method based on an artificial intelligence model according to claim 1, characterized in that, The event log includes workpiece identifier, equipment identifier, event type, local event time, gateway reception time, boundary photoelectric status, vibration root mean square, active power, encoder cumulative pulses, buffer count, and data source status code. Event records that fail message verification are deleted, and event records that are cached and retransmitted by the gateway retain the original local event time and are marked with the status code of the data source.

3. The dynamic enterprise credit scoring method based on an artificial intelligence model according to claim 1, characterized in that, Binding the actual path of the workpiece includes: defining the effective reading range of RFID with a continuous occlusion interval formed by the boundary photoelectric trigger; retaining only the first verified reading record for the same workpiece identifier within the same continuous occlusion interval; associating the workpiece events of the downstream node only with the unclosed events of the same identifier in the predecessor node set; and requiring that the equipment boundary RFID events have at least one synchronous physical evidence of the following: displacement of the conveyor encoder, change of workpiece position in PLC, vibration from the idle area to the operating area, or active power from the standby area to the processing area.

4. The dynamic enterprise credit scoring method based on an artificial intelligence model according to claim 1, characterized in that, The physical propagation time is determined based on the effective distance, average linear velocity, number of preceding workpieces, and actual release rate; the edge-level relative clock residual is obtained by subtracting the physical propagation time from the local event time difference; The relative clock drift rate is determined based on the relative clock residuals of the edge level, which vary with the timing of the gateway reception of multiple complete workpiece event pairs connected to the same adjacent node; and is determined using an anti-outlier statistical method. The relative clock offset of the window start point is determined after deducting the influence of the relative clock drift rate.

5. The dynamic enterprise credit scoring method based on an artificial intelligence model according to claim 1, characterized in that, The self-consistency coefficient of the material flow constraint process chain event time is jointly determined by the time estimable coverage ratio, the end-to-end closure residual, and the edge-by-edge physical propagation deviation. The time estimable coverage ratio represents the proportion of complete and time-estimated workpieces at the inlet and outlet. The end-to-end closure residual represents the closure difference between the end-to-end elapsed time in the same hardware clock domain and the correction edge propagation time accumulated along the actual path of the workpiece. The edge-by-edge physical propagation deviation represents the deviation between the correction edge propagation time and the physical propagation time.

6. The dynamic enterprise credit scoring method based on an artificial intelligence model according to claim 1, characterized in that, The generation of the timing diagram input includes: taking the entry reference event time as the starting point of the workpiece time, accumulating the propagation time of the correction edge along the actual path of the workpiece to obtain the physical anchoring event time; making the timing event vector in the timing diagram input contain event type encoding; sorting the events according to the physical anchoring event time, and when multiple events occur at the same physical anchoring event time, sorting them sequentially according to the deterministic dictionary order of the production line topology level and equipment identifier; and not performing mean filling, zero filling, or linear interpolation on missing events, but representing the missing state through the data source status code and time validity mask.

7. The dynamic enterprise credit scoring method based on an artificial intelligence model according to claim 1, characterized in that, The solidified dual-branch artificial intelligence model includes a temporal branch and a disordered aggregation branch. The temporal branch inputs and outputs the risk probability of the temporal branch according to the temporal diagram. The disordered aggregation branch outputs the risk probability of the disordered aggregation branch according to the equipment statistics. The equipment statistics include effective operating duration, workpiece passing count, vibration energy quantile, active power quantile, buffer peak value, and effective data ratio. The equipment statistics do not include local event time, gateway reception time, event sequence number, and physical anchoring event time.

8. A dynamic enterprise credit scoring system based on an artificial intelligence model, characterized in that, include: The physical operations data acquisition module is used to collect event records within the scoring window and build a directed graph of the production line; The module includes a workpiece path binding module, used to generate the actual workpiece path based on workpiece identification, topological predecessor relationships, and synchronous physical evidence; a time self-consistency calculation module, used to obtain the correction edge propagation time and generate the time self-consistency coefficient of material flow constrained process chain events; an anchored time axis and dual-input generation module, used to generate a time sequence graph input and an unordered aggregation input based on the correction edge propagation time; a solidified dual-branch AI inference module, used to output the time sequence branch risk probability and the unordered aggregation branch risk probability based on the time sequence graph input and the unordered aggregation input, and to keep the two branch risk probabilities consistent when the time sequence graph input is empty; and a time reliable fusion and conflict correction module, used to generate robust risk probabilities based on the time self-consistency coefficient of material flow constrained process chain events, the time sequence branch risk probability, and the unordered aggregation branch risk probability. The order calibration and credit scoring output module is used to output dynamic corporate credit scores and corporate credit ratings based on the robust risk probability.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the dynamic enterprise credit scoring method based on an artificial intelligence model as described in any one of claims 1 to 7.

Images