A predictive maintenance method for intelligent logistics systems

Abstract

This invention discloses a predictive maintenance method for intelligent logistics systems, belonging to the field of intelligent logistics equipment operation and maintenance technology. Under the premise of ensuring that the original component-level data does not leave the local area, the method of this invention efficiently aggregates the scattered, heterogeneous and real-time changing equipment health information into degradation cognition that can be directly executed at the system level, and drives causal fault chain reasoning covering components, subsystems and business consequences based on this cognition, thereby realizing the synchronous closed loop of automated operation and maintenance and production line scheduling.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics equipment operation and maintenance technology, and specifically to a predictive maintenance method for intelligent logistics systems. Background Technology

[0002] In publicly available information and industry implementation cases, predictive maintenance in automated warehousing typically employs a centralized monitoring + cloud analytics approach. Various sensors on-site continuously collect high-frequency signals such as vibration, temperature rise, current, and position via PLCs or gateways, which are then aggregated to a data platform using protocols such as OPC-UA and MQTT. The maintenance platform builds device-level digital twins in the cloud, performs feature engineering and fault label cleaning based on a big data lake, and then uses models such as random forests, LSTM, or Gradient Boosting to predict remaining lifespan and output health scores. The prediction results are returned to the WMS or MES via API to trigger manual maintenance plans and spare parts requests; the decision-making chain still relies primarily on manual review.

[0003] To alleviate bandwidth and latency pressures, some solutions offload lightweight inference models to edge servers, retaining only model training and version management on the cloud side. The process involves uploading historical curves in batches during device maintenance, iterating the model in the cloud, and then redistributing the weights as a whole. Edge nodes perform inference on real-time data based on the latest weights and periodically report health status via REST or gRPC. This solution reduces the frequency of raw data transmission to some extent, but still requires centralized collection of all samples during the training phase, making it difficult to meet the requirements of scenarios with strict confidentiality of local data.

[0004] There are also practices in the industry utilizing federated learning frameworks, the idea being to share gradients across multiple sites rather than raw data. A typical implementation relies on the open-source FedAvg algorithm: each gateway updates its model locally, then sends the gradients to the cloud for aggregation, and finally generates a global model that is broadcast back to the edge. To adapt to the diverse equipment in logistics scenarios, a common practice is to maintain separate sub-models for different machine types, or to introduce equipment type weights during aggregation to reduce interference from heterogeneous samples. However, existing literature mainly focuses on single-layer equipment granularity, lacking a hierarchical relationship between conveyor sections, sorting machines, and the overall warehouse operational consequences. Federated outputs typically remain at the level of single-machine health scores and have not yet formed a closed-loop linkage with production scheduling.

[0005] In addition, some digital twin platforms focus on 3D visualization and real-time monitoring. After acquiring PLC data through the OPC-UA channel, they map component status to virtual model nodes and trigger alarms through a rule engine. While this type of solution achieves the integration of data and model in system-level display and alarm logic, it relies more on static thresholds or defect libraries. It does not incorporate dynamically learned degradation trends into causal chain reasoning, nor does it perform probabilistic analysis and ranking of multi-level impacts from equipment and subsystems to the business layer.

[0006] While centralized monitoring combined with cloud analytics can enable predictive maintenance for individual devices, the complete sensor curves still need to be uploaded to a remote data lake during peak data transmission periods, consuming bandwidth and increasing the risk of information leakage. When there are tens of thousands of devices in a warehouse, this centralized aggregation method cannot meet the real-time requirements of second-level timing. Improved solutions combining edge inference with cloud training reduce communication pressure to some extent, but still rely on centralized collection of historical samples for model iteration, failing to completely prevent the cross-domain flow of raw data and making them unsuitable for scenarios that strictly adhere to OT isolation procedures.

[0007] While existing federated learning frameworks reduce the outflow of raw data through gradient aggregation, they mostly remain at a single-layer granularity of aggregation by device type, lacking contextual modeling of subsystem topology and business consequences. Therefore, they cannot directly provide actionable health decisions for production line-level scheduling. Digital twin platforms rely heavily on static thresholds for 3D visualization and rule-based alerts, failing to combine multi-source health assessments with causal chain reasoning. This makes it difficult to provide clearly defined, prioritized maintenance recommendations before faults propagate to the business layer. Furthermore, the various technical elements are independent and have not yet formed a real-time closed loop from edge awareness and health fusion to business linkage. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a predictive maintenance method for intelligent logistics systems to solve the technical problems mentioned in the background art.

[0009] To achieve the above objectives, the method of this invention, while ensuring that the original component-level data does not leave the local machine, efficiently aggregates scattered, heterogeneous, and real-time changing equipment health information into a degradation cognition that can be directly executed at the system level. Based on this cognition, it drives causal fault chain reasoning covering components, subsystems, and business consequences, thereby achieving a synchronous closed loop between automated operation and maintenance and production line scheduling. Specifically, the method includes the following steps:

[0010] S1. Perform coding scans on all electromechanical components in the warehouse. In the digital twin platform, instantiate the coded electromechanical components into component-level images, subsystem-level virtual entities, and system-level parent instances in sequence. For each component, open a one-way write-only data channel on the corresponding PLC or edge gateway to write the original sampling information including vibration, temperature rise, and current. At the same time, record the entity serial number, installation coordinates, and unique binding identifier of component-level image-data channel in the index table established on the platform side.

[0011] S2. Generate a degradation reference set by aggregating the historical operation and maintenance records of the warehousing system. Divide all components in the index table into multiple categories according to the dominant sensing mode. Construct a lightweight model blueprint with parameter constraints for each category. Encapsulate the blueprint into a container image and distribute it to the twins of each component. Use the degradation reference set to train the image locally and output the health scale. Then synchronize it to the subsystem scheduling layer through the read-only indicator channel.

[0012] S3. Deploy federated learning scheduling services on the execution system or cloud side. Distribute participation identifiers and gradient upload quotas to all registered component twins in the index table according to the warehouse shift rhythm. After each component twin completes its local weight update, each component twin uploads its local gradient. The scheduling service calls the built-in hierarchical scenario weighted aggregation operator to obtain the system-level degradation parameter set based on the local gradients uploaded by each component twin. The system-level degradation parameter set with version number is broadcast for each component to replace at once.

[0013] S4. Instantiate the component nodes, subsystem nodes, and business consequence nodes into a three-layer causal Bayes graph. Update the underlying prior using the global health vector obtained from the federated learning scheduling service in step S3. Infer the synchronization probability surface from the bottom up. Perform depth-first path scoring on the causal graph after pruning with a sliding window and business occupancy threshold. Obtain high-risk fault chains and encapsulate them into event streams to synchronize with the work order system and scheduling system to automatically generate maintenance work orders and adjust picking or stacking schedules.

[0014] Its beneficial effects are as follows: This invention establishes a unidirectional write-only channel on each electromechanical component side and generates an independent image for it on a digital twin platform, achieving complete local encapsulation of the original vibration, temperature rise, and current waveforms. This structure avoids the bandwidth occupation and leakage risks caused by large-scale cross-network aggregation of raw data, but can still complete data traceability in milliseconds through unified identification and chain indexing. From a mechanism perspective, it simultaneously meets the dual requirements of information isolation and real-time perception, making up for the gap in existing technologies where both are difficult to achieve simultaneously. Facing the scenario of large-scale deployment of heterogeneous devices, this invention introduces a hierarchical scenario weighted aggregation operator in the federated learning scheduling stage. Based on the local gradient uploaded by each component twin, a system-level degradation parameter set is obtained, thereby obtaining a system-level degradation vector covering the entire warehouse topology and business load. For fault chain identification, the component health, subsystem status, and business consequences are associated into a directed graph, and depth-first scoring is performed after real-time edge pruning in a sliding window, so that potential downtime links are sorted and output in the probability domain.

[0015] Further, in step S1, the process of sequentially instantiating component-level images, subsystem-level virtual entities, and system-level parent instances of the electromechanical components undergoing encoding scanning on the digital twin platform includes: first, instantiating a component-level image for each electromechanical component on the digital twin platform, simultaneously writing the entity serial number and installation coordinates; after completing the instantiation of component-level images for all electromechanical components, calling the platform aggregation interface to sequentially merge the instantiation of component-level images for all electromechanical components into multiple subsystem-level virtual entities; then, mounting all subsystem-level virtual entities to the system-level parent instance to form the system-level parent instance.

[0016] This invention first establishes a strongly coupled, integrated mapping between on-site assets and the digital space. After all electromechanical components are entered into the warehouse once via QR code or RFID, component images, subsystem virtual entities, and system parent instances are sequentially generated on the digital twin platform, forming a hierarchical three-level logical tree. Logical tree nodes are bound to physical entities with globally unique identifiers. These identifiers are synchronously written to an index table upon creation. In addition to recording serial numbers and installation coordinates, the index table also stores chained audit entries for image-channel binding, ensuring that any upper-level algorithm can reverse-locate to the data source without exceeding authorized access permissions.

[0017] Further, in step S2, the process of aggregating historical operation and maintenance records of the warehousing system to generate a degradation reference set includes: aggregating existing operation and maintenance records of the warehousing system into a historical data table; retrieving each record in the historical data table from the index table by entity sequence number to trace back and lock the complete life cycle segment of the component; extracting the failure time point by comparing with the operation and maintenance log during the locking period; binding the failure time point with the sensing curve of the same period in the digital twin platform to generate a degradation label; and writing the degradation label, the sensing curve of the same period, the process link and the load into the metadata to accumulate and form a degradation reference set that can be directly referenced by the model.

[0018] The method of this invention generates a degradation reference set, ensuring that the historical state, failure point and field conditions of each component remain consistent within the platform.

[0019] Furthermore, in step S2, lifecycle monitoring is enabled on the read-only indicator channel. When a component enters a maintenance or replacement state and triggers a flag, the channel is immediately frozen.

[0020] This invention prevents erroneous health status from interfering with subsequent system decisions by enabling lifecycle monitoring on the read-only indicator channel.

[0021] Further, in step S3, after each component twin completes its local weight update, the local gradient upload of each component twin includes: encapsulating the gradient vector obtained after each component twin completes its local weight update into a message; sequentially writing the twin identifier, the subsystem number to which it belongs, and the generation timestamp into the message, and pushing it to the federated learning scheduling service through the encrypted channel between the edge gateway and the federated learning scheduling service; the edge gateway synchronously sets an upload rate threshold, and if the message queue length reaches the upload rate threshold, it will automatically be delayed in queuing.

[0022] When uploading local gradients, this invention ensures that only parameter differences are transmitted without leaking the original vibration, temperature rise, and current data. Furthermore, by setting an upload rate threshold, this invention monitors the message queue length during upload and automatically delays queuing when the message queue length reaches the upload rate threshold, thus prioritizing the bandwidth of sorting and stacking control messages.

[0023] Further, in step S3, the scheduling service calls the built-in hierarchical scene weighted aggregation operator to obtain a system-level degradation parameter set based on the local gradients uploaded by each component twin. This includes: the weighted aggregation operator establishing subsystems for gradient vectors from the same conveyor section, sorting machine, or aisle stacking unit; subsequently, the digital twin platform is called in real time to obtain three field indicators for each device: equipment serialization depth, recent shift order throughput, and historical downtime cost; after weighting is completed at the subsystem level, the weighted aggregation operator reads task priority, spare parts availability, and manual entry difficulty for each subsystem draft and calculates global coefficients; finally, the system-level degradation parameter set is output.

[0024] Furthermore, the calculation formula for the system-level degradation parameter set is obtained as follows:

[0025] ;

[0026] In the formula, g i B is the gradient vector of the i-th device; p For the p-th subsystem; This is the device serial connection depth coefficient; the deeper the connection, the greater the weight. This is the order throughput coefficient for the most recent shift; the busier the shift, the greater the weight. This represents the historical downtime cost coefficient; the higher the cost, the greater the weight. p g represents the priority coefficient for subsystem tasks. p The availability coefficient for spare parts in the subsystem is determined by the scarcity of parts, with a higher weight given to more scarce parts. p The difficulty coefficient for manual intervention in the subsystem is denoted by P, with higher weights for more difficult interventions; the total number of subsystems is denoted by P; and the set of system-level degradation parameters is denoted by G.

[0027] This invention introduces a three-dimensional scenario weighting of serial depth, order throughput, and downtime cost during the federated learning scheduling phase, and performs segmented weighting based on the hierarchical relationship between components, subsystems, and systems, thereby obtaining a system-level degradation vector covering the entire warehouse topology and business load. This approach breaks through the single-layer granularity limitation in the algorithm flow, enabling, for the first time, dispersed health information to be mapped into executable parameters for production line scheduling without exposing the original data.

[0028] Furthermore, in step S4, the depth-first path scoring algorithm is divided into three segments:

[0029] The first segment, path enumeration: For each bottom component node, push the quadruple of current node, cumulative layer number, cumulative probability, and minimum health degree onto the stack; when expanding child nodes, read the cargo flow carrying value and hub level in the digital twin, and if the product of the two is lower than the threshold, backtrack immediately.

[0030] The second part, real-time pruning: when the search depth reaches the top-level business node and the path is still retained, immediately calculate its conditional probability product and the minimum health of the path. If either of them is lower than the scheduling threshold, the path is directly discarded.

[0031] The third stage, risk assessment: For each path that has passed the filtering in the first two stages, the accessibility coefficient of spare parts and the difficulty coefficient of manual entry are read from the digital twin, and then a unified weight constant of the business layer is introduced to finally determine the risk score.

[0032] This invention achieves real-time pruning of low-impact links by immediately backtracking when the product of the two is below a threshold, eliminating non-critical branches during the enumeration phase, and directly discarding the path if either of the two is below the scheduling threshold.

[0033] Furthermore, the risk score is determined using the following formula:

[0034] ;

[0035] In the formula, D m For path The set of component nodes on h; d The health scale of node d is output by federated learning, ranging from [0,1]. For path The set of directed edges on p; e Let u1 be the conditional probability of edge e, given a priori by the causal graph; u1 is the global weight constant for spare parts accessibility, which takes a larger value as the shortage increases; u2 is the global weight constant for the difficulty of manual entry, which takes a larger value as the difficulty increases. For path Spare parts availability coefficient; For path The difficulty level of manual entry; For path Risk score.

[0036] Based on the above, the operator only references three types of information: health status, conditional probability, and business coefficient, and does not read any raw vibration, temperature rise, or current data. All risk scores... After automatic sorting using a min-heap, a high-to-low risk list is formed and returned for inference, enabling interpretable priority output for the entire portfolio's failure chain.

[0037] Furthermore, it also includes:

[0038] S5. Decompose the high-risk fault chain by component dimension to generate a risk write-back package, write it into the feature risk labeling table of the corresponding twin through an encrypted channel, include features with importance higher than the threshold in the key upload list, increase the frequency of the next round of gradient upload, and implement gradient freezing for features not included.

[0039] S6. The Federation Dispatch Service aggregates the characteristic risk labeling tables of each twin to generate an upload profile. For components marked as high-risk, it enables enhanced synchronization strategies, sets up dedicated messages and bandwidth reservations, and enables economical synchronization strategies to extend the upload cycle for low-risk components. The synchronization cycle and bandwidth changes are written to the dispatch log and broadcast to the operation and maintenance dashboard and network management system to complete the closed loop from risk assessment to communication dispatch.

[0040] This invention writes risk scores back to the component twins, dynamically adjusting the gradient upload frequency and bandwidth reservation: high-risk components use enhanced synchronization, while low-risk components enter conserved synchronization. All bandwidth and cycle changes are written to the scheduling log and displayed in real time on the network management and operation and maintenance dashboards. Combined with direct event stream push to the work order system and scheduling system, and the write-back mechanism that automatically increases the gradient synchronization frequency for high-risk components, the solution forms an operation and maintenance closed loop integrating health fusion, causal reasoning, and adaptive communication strategies.

[0041] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall process of the predictive maintenance method for intelligent logistics systems according to the present invention;

[0043] Figure 2 This is a flowchart illustrating step S3 of the method of the present invention;

[0044] Figure 3 This is a schematic diagram of step S4 of the method of the present invention;

[0045] Figure 4 This is a flowchart of step S5 of the method of the present invention;

[0046] Figure 5 This is a flowchart illustrating step S6 of the method of the present invention;

[0047] Figure 6 This is a schematic diagram comparing the prior art with the present application. Detailed Implementation

[0048] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. However, those skilled in the art should understand that the embodiments described below are only for illustrating the present invention and should not be regarded as limiting the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Examples of Predictive Maintenance Methods for Intelligent Logistics Systems

[0050] The innovation of this invention lies in:

[0051] This technical solution first establishes a strongly coupled, integrated mapping between on-site assets and the digital space. After all electromechanical components are entered into the warehouse once via QR code or RFID, component images, subsystem virtual entities, and system parent instances are sequentially generated on the digital twin platform, forming a hierarchical three-level logical tree. Logical tree nodes are bound to physical entities with globally unique identifiers, which are simultaneously written to an index table upon creation. In addition to recording serial numbers and installation coordinates, the index table also stores chained audit entries for image-channel binding, ensuring that any upper-level algorithm can trace back to the data source without unauthorized queries. Unlike existing centralized monitoring, this solution opens a write-only data channel on each component side. The channel uses bidirectional certificate authentication and a whitelist downlink strategy to block uplink access to raw vibration, temperature rise, and current waveforms. All high-frequency samples are cached locally and input to the component-level twin via a one-way push mechanism, achieving a unified data isolation and real-time perception, laying a secure foundation for subsequent distributed learning.

[0052] This solution leverages modality partitioning and parameter constraint strategies to build lightweight modeling capabilities for tens of thousands of components. Supported by historical operational data and failure tags, the platform prunes convolutional stacks, recurrent units, or self-attention blocks for three dominant modes: rotation, aerodynamics, and low power consumption. This ensures that any single-core edge processor instantiated meets preset parameter limits and has a consistent output interface. A containerized pipeline encapsulates the pruned blueprints into images and deploys them with a single click. Component twins are trained locally and output normalized health metrics. The federated learning scheduler uses shift ticks as round indices, first distributing gradient upload quotas to each component, then performing hierarchical scenario-weighted aggregation in the cloud: within subsystem buckets, three on-site indicators—serialization depth, order throughput, and historical downtime cost—are introduced to calculate local coefficients. At the system level, task priority, spare parts availability, and access difficulty are further combined to obtain global coefficients. Finally, a system-level degradation parameter set is output and broadcast with a version number. This aggregation process incorporates topology depth and business load information into the computation graph, enabling collaborative learning of equipment health and production line constraints. It goes beyond simple weighting based on sample size or equipment type, addressing the pain point of data dispersion and system cognitive gaps from a logical structure perspective.

[0053] After updating the global health vector, this solution connects component health, subsystem status, and business consequences into a directed graph using a three-layer causal Bayesian network. The system periodically updates the underlying priors with a health scale, derives the synchronization probability surface from the bottom up, and uses a sliding window and business occupancy threshold to prune the edge set in real time. A depth-first path scoring operator searches for high-risk fault chains on the pruned graph, calculates a risk score by combining conditional probability, minimum health, and spare parts reachability coefficients, and pushes the path results directly to the work order system and scheduling system via an event stream, forming a closed loop linked to scheduling. This solution writes the risk score back to the component twin, dynamically adjusting the gradient upload frequency and bandwidth reservation: high-risk components use enhanced synchronization, while low-risk components enter conserved synchronization. All bandwidth and periodic changes are written to the scheduling log and displayed in real time on the network management and operations dashboard. Through risk-communication coupling, this solution achieves a closed-loop process of health fusion, fault chain reasoning, and resource scheduling without touching the original data, forming a core technology system that is significantly different from existing solutions that only perform single-layer health scoring or static threshold alarms.

[0054] like Figure 1 As shown, the predictive maintenance method for intelligent logistics systems in this embodiment includes the following steps:

[0055] S1. Encode and scan all electromechanical components in the warehouse. In the digital twin platform, instantiate the component-level image, subsystem-level virtual body and system-level parent instance in sequence. For each component, open a one-way write-only data channel in the corresponding PLC or edge gateway to write the original samples such as vibration, temperature rise and current. At the same time, record the entity serial number, installation coordinates and image-channel unique binding identifier in the index table.

[0056] S2. Generate a degradation reference set by aggregating historical operation and maintenance records, construct a lightweight model blueprint with parameter constraints according to rotation mode, aerodynamic mode and low power mode, encapsulate the blueprint as a container image and distribute it to the twins of each component, complete the training locally and output the health scale, and then synchronize it to the subsystem scheduling layer through the read-only indicator channel.

[0057] S3. Deploy federated learning scheduling services on the execution system or cloud side, distribute participation identifiers and gradient upload quotas according to shift rhythm, and upload local gradients to the twins of each component; the scheduling service calls the hierarchical scenario weighted aggregation operator, combines indicators such as device serialization depth, order throughput, downtime cost and spare parts availability to obtain a system-level degradation parameter set, and broadcasts global model weights with version numbers for each component to replace at once;

[0058] S4. Instantiate the component node, subsystem node and business consequence node into a three-layer causal Bayes graph, update the underlying prior using the global health vector obtained in step S3, and obtain the synchronization probability surface by bottom-up reasoning. After pruning with sliding window and business occupancy threshold, perform depth-first path scoring to obtain high-risk fault chains and encapsulate them into event streams to be synchronized to the work order system and scheduling system.

[0059] S5. Decompose the high-risk fault chain by component dimension to generate a risk write-back package, write it into the feature risk labeling table of the corresponding twin through an encrypted channel, include features with importance higher than the threshold in the key upload list, increase the frequency of the next round of gradient upload, and implement gradient freezing for features not included.

[0060] S6. The federated learning scheduling service summarizes the characteristic risk labeling table of each twin to generate an upload profile. For components marked as high risk, it enables enhanced synchronization strategy, sets dedicated messages and bandwidth reservations, and enables economical synchronization strategy to extend the upload cycle for low-risk components. The synchronization cycle and bandwidth changes are written to the scheduling log and broadcast to the operation and maintenance dashboard and network management system to complete the closed loop from risk assessment to communication scheduling.

[0061] Step S1 includes the following sub-steps:

[0062] S11. Scan all electromechanical components on site one by one and input them into the database to generate set C. For each member in C, first instantiate a component-level image on the digital twin platform, and at this time, write the entity serial number and installation coordinates simultaneously. After all components in the same conveyor section or the same roadway have been imaged, call the platform aggregation interface to merge them into subsystem-level virtual entities in sequence to form a set. ; and then all Unified mounting to system-level parent instance The final result is a three-level traceable logical tree of components → subsystems → systems. A unique index table E is created on the platform side for this tree, where each row contains a record. Hierarchical identifier, entity serial number, installation coordinates This provides a quick search entry point for channel binding.

[0063] S12. Establish an independent data channel x for each component in C within its corresponding PLC or edge gateway. i Vibration acceleration, temperature rise, current waveform, and air pressure signals are written at a fixed sampling interval, and then the streaming data is sent to the component-level twin via a one-way push mechanism; simultaneously, for x i Enable a whitelist subscription strategy to allow only downlink calls and block any uplink reads; add a timestamp field and a quality identifier field to the channel header, both of which together ensure accurate indexing when aligning across devices.

[0064] S13. Compare the component-level images from step S11 with the x-images from step S12. i Binding is performed: Two-way certificate authentication is used to verify the identities of the physical and virtual terminals; a globally unique identifier k is generated according to deterministic naming rules. i And write it to index table E; record key information such as creator, release time, and hardware version, forming a complete mapping relationship between image, channel, and entity. Then write k... i Simultaneously generate chain audit entries To ensure that subsequent operation and maintenance traceability can be based on Replicate every binding action;

[0065] Based on this, any upper-level logic component only needs to call k i This allows you to locate the original data source and completely prevent unauthorized data requests across layers.

[0066] Step S2 includes the following sub-steps:

[0067] S21. Aggregate the existing operation and maintenance records of the warehousing system into a historical data table H. For each record in H, first retrieve it to the index table E by entity sequence number, and then trace back to lock the complete life cycle segment of this component. During the lockout period, extract the failure time point by comparing with the operation and maintenance logs. ,Will With the same period sensing curve S j Binding is performed within the digital twin platform to generate degradation tags. j ; put d j S j Process steps and load rating Metadata is written to accumulate and form a set of degraded references that can be directly referenced by the model. This ensures that the historical state, failure points, and field conditions of each component remain consistent within the platform.

[0068] S22. Divide all components in index table E into three categories according to the dominant sensing mode, denoted as rotational mode Z. rot Aerodynamic mode Z pne With low-power mode Z lp For each type of construction, there is a corresponding network blueprint B. u ,in The blueprint internally describes hierarchical relationships using modular methods such as convolutional stacks, loop units, or self-attention blocks.

[0069] For each blueprint B u Assign weights layer by layer, such that the s-th layer contains R. u,s There are adjustable dimensions, and the set of dimensions is denoted as . ;This is used to establish a parameter counting function:

[0070] ;

[0071] Among them, P(B) u Blueprint B u Total number of parameters; S u For Blueprint B u The number of layers; R u,s For Blueprint B u The number of parameter dimensions in the s-th layer; Let be the parameter size of the s-th layer and r-th dimension; This represents the maximum number of parameters that a single-core edge processor can handle.

[0072] For P(B) u Apply inequality constraints and reduce the hierarchy without altering the hierarchical relationship. Real-time compression of blueprint size ensures that all three types of blueprints, when instantiated on different components, meet the requirements. Limitations and achieve a uniform output size .

[0073] The blueprint with controlled parameters is used to generate a model script, which is then packaged into an OCI container image. (i is the component index), written in the mirror label Entity serial number, blueprint type, blueprint version number ; Invoke platform-level template scheduling to The application is distributed to the corresponding component twin and a unique runtime instance identifier is generated on the target device side. Based on this, all components under different sensing modes acquire lightweight, parameter-consistent, and unified degradation modeling capabilities.

[0074] S23. Transfer the image sent in step S22. Local training is completed by calling the degenerate reference set on the component side. After training, a normalized mapping is applied to the model output channel to generate a health scale h. i Where 0 represents complete failure, 1 represents a completely new state, and the minimum resolution is 0.01; for h i Bind it to a predefined color block, and set h i Write component metadata channel .

[0075] S24. Indicator channel using read-only upward throw. h i Synchronize to the subsystem-level scheduler and align the push cycle with the physical device's cycle time by default; Write the metadata area to the component public key K pub,i At the same time Enable lifecycle monitoring; when a component enters maintenance or replacement mode and triggers a flag. Immediately freeze this channel to prevent incorrect health status from interfering with subsequent system decisions;

[0076] The entire synchronization process automatically generates an audit chain. .

[0077] Among them, such as Figure 2 Step S3 includes the following sub-steps:

[0078] S31. Deploy the federated learning scheduling service F on the warehouse execution system server or cloud mirror, and set the on-site shift beat rate T. shift For each cycle, the training round index r is generated by F at the start of each cycle. k A unique participation identifier is issued to each registered component twin in index table E. Through two-way key negotiation in the encrypted channel Internal writing of gradient upload quota q i Ensure all electromechanical nodes are in a unified communication topology view before aggregation starts. Upon completion of distribution, F immediately records the round version number v. k And push to the platform log stack .

[0079] S32. After updating the local weights of each component twin in this round, obtain the gradient vector g. i Encapsulated into a message ;exist Twin identifiers are written sequentially in the middle. Subsystem Number With the generation timestamp via the edge gateway Pushing data to F ensures that only parameter differences are transmitted without leaking original vibration, temperature rise, and current data. The gateway side synchronously sets the upload rate threshold. ,like Queue length reached The queue will be automatically delayed to prioritize the bandwidth of sorting and stacking control messages.

[0080] S33. Perform gradient integration using the built-in hierarchical scenario weighted aggregation operator A of the federated learning scheduling service in a two-level order:

[0081] 1. A represents the gradient vector g from the same conveyor section, sorter, or aisle stacking unit. i Establish subsystem B p Subsequently, the digital twin platform was invoked in real time to obtain three field indicators for the equipment serial connection depth. The volume of orders processed in the last batch Historical downtime costs Generate local coefficients for each device After weighting is completed at the subsystem level, A then drafts q for each subsystem. p Read task priority f p Spare parts availability g p Difficulty of manual entry m p Calculate the global coefficient k p Finally, output the system-level degradation parameter set G and write all coefficient sources into the auditable table. This ensures that causal reasoning can be traced back to the aggregation path, and that only the indicator dictionary needs to be updated when the warehouse network is rerouted, without rewriting the internal logic of A:

[0082] ;

[0083] Where A is the hierarchical scene weighted aggregation operator; g i B is the gradient vector of the i-th device; p For the p-th subsystem; This is the device serial connection depth coefficient (the deeper the connection, the greater the weight). This is the order throughput coefficient for the most recent shift (the busier the shift, the greater the weight). This is the historical downtime cost coefficient (the higher the cost, the greater the weight). For the first The overall coefficient of the equipment at the subsystem level (i.e., in the formula) ); f p g represents the priority coefficient for subsystem tasks. p For the availability coefficient of spare parts for the subsystem (the scarcer the part, the greater the weight); m p k represents the difficulty coefficient of manual intervention in the subsystem (the more difficult, the higher the weight); p The comprehensive coefficient of the p-th subsystem at the global level (i.e., in the formula) P represents the total number of subsystems; q p G is the degenerate draft of the p-th subsystem (implied in the weighted summation of the formula, so it is not listed separately); G is the system-level degenerate parameter set; This is an audit table for aggregated metadata.

[0084] S34, All q p The data is then aggregated again, and the logistics operation priority matrix J, order throughput matrix O, and spare parts consumption sensitivity matrix Y are read to form a weight tensor z; for each q p After applying z, normalized weighting is performed to generate the system-level degradation parameter set G. * F is G * Assign a unique version number It is also broadcast to the monitoring and planning system via the message bus MB to ensure that core lines and high-concurrency nodes receive priority attention and that upstream and downstream decision-making sequences are consistent.

[0085] Where G in step S33 is the same as G here. * This is not a repetition of the same symbol, but rather the result of the same type of object at different computational stages: G in step S33 is the system-level degradation parameter set obtained by the hierarchical scenario weighted aggregation operator after completing two levels of weighting from component to subsystem and from subsystem to system; Step S34, based on this, introduces a weight tensor z composed of the logistics operation priority matrix J, the order throughput matrix O, and the spare parts consumption sensitivity matrix Y, for each q p After further normalization and weighting, the release result G is obtained for external broadcasting and version control. * Therefore, the version number Binding G * G * This is the system-level degenerate parameter set obtained by weighting G twice with z, and G will be used uniformly in subsequent deployments and causal inferences. * As the final standard.

[0086] S35, G * via the original channel The broadcast is sent to all component twins, and each twin receives the version number. The local model weights are then immediately replaced and used in the next sampling period, enabling each device to share global degradation cognition with a single lightweight synchronization without leaking sensitive information. Once the weight replacement is complete, the twin returns a confirmation flag. F based on all Refresh the online participation list After weight replacement, h is generated and uploaded. i H is generated by aggregation on the platform side.

[0087] S36, Version number Weight vectors of each subsystem With gradient summary A read-only audit table is written through the digital twin platform, and a read-only interface (RO) is provided for the causal reasoning module to obtain the latest health priors. Simultaneously, a version change signal is sent to the warehouse execution system. This triggers a fine-tuning of the scheduling strategy.

[0088] Among them, such as Figure 3 Step S4 includes the following sub-steps:

[0089] S41. Denote all electromechanical components of the warehousing center as the bottom-level node set D, each conveyor section, cross belt, or aisle stacking unit as the middle-level node set S, and business consequences such as delivery delays, throughput reductions, and sudden increases in energy consumption as the top-level node set. Subsequently, D, S, Instantiated sequentially into a three-level directed causal graph The nodes. The physical connections between nodes and the flow of goods are extracted as a priori edge set. And solidify it within the digital twin platform, making The structure corresponds to the site topology. After completion, write the metadata area. Node number, hierarchical level, installation coordinates .

[0090] S42. Split the global health vector H that is aggregated after each round of federated learning update, and extract the corresponding component health scale h. i Write to the prior distribution of the underlying nodes; along Perform a Bayesian probability push from bottom to top, first calculating the posterior of the middle-level nodes, and then continuing to push up to the top level. Ultimately, this forms a synchronous probability surface covering the entire portfolio. Ensure that causal reasoning remains consistent with the latest degenerate cognition.

[0091] The global health vector H is the set of system-level degradation parameters completed in each round of federated learning (or, if using the aforementioned approach, it is...). After the broadcast replacement, each component twin performs a health scale calculation on the current data window locally based on the latest parameters and obtains h. i Then, it is synchronized to the platform side via a read-only indicator channel for aggregation. The platform will then list all components according to the registration order of index table E. i spliced ​​together And write it to the read-only audit table for use in this step S42 to update the underlying prior.

[0092] S43. Use a fixed-width sliding window W to maintain the node activation timestamps within the most recent shift; for Edges falling outside the window are set to zero weight. Then, a service occupancy threshold is set for each subsystem node. Only when the number of orders in the window is consistently greater than Its incident edge is activated only when necessary. By compressing the probability flow to the actual production hotspot through double pruning, the computational load is reduced and idle links are prevented from interfering with the inference results.

[0093] S44. Fix the cause-effect graph of the trimmed sliding window as follows: And hosted in memory, where A set of nodes representing a cause-effect graph, containing steps Instantiated component nodes, subsystem nodes, and business consequence nodes; This represents the set of directed edges that remain active after being pruned by the sliding window W and the service occupancy threshold, with the corresponding conditional probability edge weights retained on each edge. Edges outside the window or those not activated are set to zero or excluded from this round of inference. ,therefore This is the causal graph used for depth-first path scoring in this round after the sliding window has been clipped. Subsequently, the business-driven depth-first path scoring operator proposed in this application is deployed within the federated scheduling node. This operator first initializes the explicit stack. Save the depth-first search state and then initialize the min-heap. Used as a risk outcome priority queue; the algorithm is divided into three segments, each of which can access the real-time fields of the digital twin, thereby continuously maintaining consistency with the on-site working conditions.

[0094] First, path enumeration.

[0095] For each underlying component node Will Push the node onto the stack, with the four-tuple consisting of the current node, cumulative level, cumulative probability, and minimum health. D reads the cargo flow carrying capacity from the digital twin when expanding child nodes. and the level of the hub If the product of the two is lower than the threshold Then immediately backtrack to eliminate non-critical branches during the enumeration phase.

[0096] Second, prune in real time.

[0097] When the search depth reaches the top-level business node and the path is still preserved At that time, the operator immediately calculates its conditional probability product. Minimum health of the path If either of them is below the scheduling threshold All paths are directly discarded, enabling immediate pruning of low-impact links.

[0098] The third section is risk assessment.

[0099] For each item that passes the first two filters The operator continues to read the spare parts accessibility coefficient from the digital twin. Difficulty level of manual entry Then introduce a unified weight constant for the business layer. , The risk score is finally output based on the following formula. And write :

[0100] ;

[0101] Among them, D m For path The set of component nodes on h; d The health scale for node d (output by federated learning, ranging from [0,1]); For path The set of directed edges on p; e u1 is the conditional probability of edge e (given a priori by the causal graph); u2 is the global weight constant for spare parts accessibility (the more scarce the spare parts, the larger the value); u3 is the global weight constant for the difficulty of manual entry (the more difficult the entry, the larger the value). For path Spare parts availability coefficient; For path The difficulty level of manual entry; For path The final risk score.

[0102] Based on the above, the operator only references three types of information: health status, conditional probability, and business coefficients, and does not read any raw vibration, temperature rise, or current data. After automatic sorting using a min-heap, a high-to-low risk list is formed and returned for inference, enabling interpretable priority output for the entire portfolio's failure chain.

[0103] S45, High-risk path sets A business impact mapping table is used to translate consequences such as cargo flow congestion, roadway downtime, and delayed delivery into scheduling instruction priorities. At the same time, for each The health differences of each internal component are color-coded to generate a visual link summary V. k And push it to the operations and maintenance panel.

[0104] S46. Encapsulate high-risk link summaries into event streams and push them synchronously to the work order system. With scheduling system ; Automatically generate maintenance tasks and bind fault chain numbers. , Adjust picking routes or stacking schedules in real time to prevent potential disruptions from spreading; simultaneously, write the identified measures taken to the corresponding nodes in the cause-effect graph. .

[0105] Among them, such as Figure 4 Step S5 includes the following sub-steps:

[0106] S51. The set of high-risk paths returned by the causal reasoning module. Break it down to a single component dimension: for each line Extracting component identifiers d node by node j Trigger Feature Name feat j and the importance scale obtained by normalizing path contribution. Then generate a risk write-back package R for each component. j Including the inference timestamp With shift number Create a write-back queue within the federated learning scheduling service. And cache all R in the order of arrival. j .

[0107] S52, to Each R in j Employs an encrypted channel shared with the gradient path. Pushed to the corresponding component twin. The twin receives the R for the first time. j Create a local feature risk label table at the time and will Write to the table; simultaneously call the UI adaptation layer to... Mapped to the operations and maintenance visual interface.

[0108] S53, Completed in twins After the update, the local training scheduler is invoked to adjust the importance scale in the table. Above the threshold Features added to the key upload list ; The frequency of gradient upload for the next round of federated training Increased to twice the original amount, only for The system utilizes open gradient uplink for features. Healthy components not included in the risk list maintain their original frequency, while gradients outside the list are frozen to prevent non-critical data from diluting global degradation cognition and to achieve on-demand allocation of bandwidth and computing power.

[0109] S54. Generate audit entries for each risk write-back action on the digital twin platform. The entry contains And write it to the read-only audit table; at the same time, send it to the work order system. Push maintenance suggestions , which includes component identifier dj Affected features and importance scale. Automatically generate maintenance task orders by combining inventory and scheduling strategies. And bind the causal graph index.

[0110] Among them, such as Figure 5 Step S6 includes the following sub-steps:

[0111] S61. The feature risk labeling table generated by the causal reasoning module in each twin body. The data is compiled and returned to the Federated Learning Scheduling Service F according to component number; for each component... In F, scale according to the highest importance level in the table. Calculate risk level And regenerate the uploaded image . Include ,in As the feature importance threshold, The target synchronization period. A new column is added to the digital twin index table E, enabling the field gateway to read the latest profile and load the corresponding upload strategy before the next round of training, ensuring that risk assessment and communication plan are consistent.

[0112] S62, to Components marked as high-risk will be subject to an enhanced synchronization strategy: all components with an importance no lower than [a certain level] will be synchronized. Feature gradients are packaged into dedicated messages And adjust the upload cycle to The gateway side is the channel. Enable bandwidth reservation This ensures that the uplink can be completed on time even during peak periods, and avoids delays in critical health information that could affect the system's cognitive updates.

[0113] S63. For components that are not marked or have been unmarked, adopt a synchronization-saving strategy: extend the upload cycle to... Furthermore, the gradient upload scope is limited to the core structure layer, generating simplified messages. ;

[0114] S64. Write all changes to the synchronization cycle of new and old components and bandwidth reservation records into the scheduling log L, and broadcast them to the maintenance dashboard via the message bus BUS. With network management system . The link priority is dynamically adjusted based on the broadcast results. It displays the synchronization status and risk distribution of each device in real time.

[0115] like Figure 6As shown, compared with the common approach of centralized monitoring plus cloud analysis, the solution in this application establishes a unidirectional write-only channel on each electromechanical component side and generates an independent image for it on a digital twin platform, achieving complete local archiving of the original vibration, temperature rise, and current waveforms. This structure avoids the bandwidth occupation and leakage risks caused by the large-scale cross-network aggregation of raw data, while still being able to complete data traceability in milliseconds through unified identification and chain indexing. Mechanistically, it simultaneously meets the dual requirements of information isolation and real-time perception, making up for the gap in existing technologies where these two aspects are difficult to balance.

[0116] In scenarios involving the large-scale deployment of heterogeneous equipment, traditional FedAvg solutions often aggregate data uniformly based on equipment type or sample size, resulting in outputs typically limited to single-machine health scores. This application introduces a three-dimensional scenario weighting approach—sequential depth, order throughput, and downtime cost—in the federated learning scheduling phase, and performs segmented weighting based on the hierarchical relationship between components, subsystems, and the system, thereby obtaining a system-level degradation vector covering the entire warehouse topology and business load. This application breaks through the single-layer granularity limitation in its algorithm flow, enabling, for the first time, dispersed health information to be mapped into executable parameters for production line scheduling without exposing the original data.

[0117] For fault chain identification, existing 3D visualization and rule engines focus on static threshold alarms, lacking quantitative reasoning about fault propagation paths. This application uses a three-layer causal Bayesian network to associate component health, subsystem status, and business consequences into a directed graph. After real-time edge pruning in a sliding window, depth-first scoring is performed to rank potential downtime chains within the probability domain. Combined with direct event stream push to the work order and scheduling systems, and a write-back mechanism that automatically increases the gradient synchronization frequency for high-risk components, the solution forms an integrated operation and maintenance closed loop encompassing health fusion, causal reasoning, and adaptive communication strategies, distinguishing it from existing methods that rely on manual review or static thresholds.

[0118] This application aims to establish a predictive maintenance system for highly automated warehousing and distribution centers, enabling health integration, fault chain reasoning, and scheduling linkage without leaking component-level raw data. The core technical features are threefold: First, a strict mapping is established using digital twins at the component, subsystem, and system levels, with write-only channels ensuring the original waveforms are permanently preserved at the edges; second, scenario weights such as topology depth, order load, and downtime cost are introduced into federated learning to generate a system-level degradation vector; third, a three-layer causal Bayesian network is used to transform this degradation vector into a high-risk fault chain that can trigger work orders and rescheduling, and the reasoning results are written back to the communication strategy, forming a real-time closed loop. The system covers key nodes such as mapping hierarchy relationships, gradient weighting methods, and risk-communication coupling, with the overall protection scope revolving around global cognition and closed-loop operation and maintenance under data isolation conditions.

[0119] Compared with existing technologies, the technology in this application adopts a combination strategy of layered twin isolation of raw data, scenario-weighted federated aggregation, and causal fault chain closed loop, which ensures security and compliance while taking into account real-time inference and scheduling linkage.

[0120] The commercial value of this patent lies primarily in its ability to provide system-level, executable degradation awareness for large-scale distributed equipment without touching sensitive raw data, directly linking work orders, scheduling, and spare parts management. This capability enables equipment owners to elevate predictive maintenance from single-machine health scoring to business continuity assurance. In continuously high-load warehousing and distribution scenarios, unplanned downtime can be reduced to a lower range of previous levels, significantly decreasing standby capacity and redundant inventory. Because the solution incorporates operational variables such as topology depth, order load, and downtime costs, maintenance and scheduling decisions are quantified to the same dimension as the profit model. Enterprises can clearly estimate potential cost savings and improved delivery reliability when assessing return on investment, thus forming a verifiable business closed loop. Simultaneously, all analysis processes perform data minimization at the component level, helping to comply with ISO27001, IEC62443, and local data export requirements, reducing compliance costs and enhancing customers' competitiveness in bidding.

[0121] From a supply chain collaboration perspective, the unified format of degradation vectors and causal failure chains can serve as an ecosystem interface, providing detailed equipment status information to upstream spare parts suppliers, maintenance outsourcing providers, or insurance institutions without exposing confidential equipment processes. Spare parts suppliers can dynamically adjust their safety stock accordingly, outsourcing providers can establish tiered warranty terms based on risk, and insurance institutions can offer differentiated premiums based on quantified degradation levels. Service providers thus gain new sources of value-added data, while equipment owners can further optimize overall maintenance costs through risk spillover and transfer, forming a win-win business network for all parties.

[0122] In terms of scalability, the multi-level digital twin and federated aggregation mechanism of this patent is not dependent on the type of components, thus allowing for smooth migration to discrete manufacturing, process industries, and energy operation and maintenance. In automotive and electronics assembly lines, the workstation density and automation level are similar to those in warehousing scenarios, and the same causal network can be reused by introducing production line cycle time and quality control indicators. In process industries such as refining, metallurgy, or pharmaceuticals, although the equipment forms are different, pipelines, pump sets, and reactors also present a hierarchical structure of components-units-devices, which can map process variables such as temperature, pressure, and proportions to degradation scales to achieve real-time decision-making for safe production. In the energy sector, such as wind farms and distributed photovoltaics, where units are dispersed and network bandwidth is limited, the local training and global decision-making mechanism of this application can compensate for the scalability bottleneck of centralized monitoring, while complying with the regulatory requirements of various countries for power data security.

[0123] Furthermore, infrastructure such as urban rail transit, data centers, and smart buildings also possess multi-layered electromechanical structures, and their operational safety and service continuity are directly linked to commercial revenue. Taking rail transit systems as an example, station air conditioning, escalators, signaling, and traction power supply are different subsystems, but they collectively determine transport capacity. Based on the technical framework of this application, the health status of each subsystem can be causally fused and linked with the train timetable to assist in formulating graded speed limits or temporary track rerouting schemes. For data centers, IT load, cooling units, and power distribution cabinets can be mapped as three-layer nodes in a causal network, achieving unified measurement of energy consumption, risk, and SLA, and supporting energy consumption scheduling strategies under dynamic electricity pricing. Through parameterized adaptation in different fields, this patent can form a reusable commercial template in SaaS or edge-cloud hybrid deployment models, increasing continuous revenue from software subscriptions and operation and maintenance hosting.

[0124] In summary, this patent upgrades traditional equipment monitoring into a predictive maintenance service that is quantifiable, accountable, and collaborative through a secure data isolation mechanism, a federated learning weight design coupled with business metrics, and a real-time closed-loop causal reasoning framework. This forms a comprehensive value proposition that balances compliance, performance, and economy, and has the potential to be replicated across industries.

[0125] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A predictive maintenance method for an intelligent logistics system, characterized by, Comprise the following steps: S1, all electromechanical components in the warehouse site are coded and scanned, the electromechanical components coded and scanned are sequentially instantiated component level image, subsystem level virtual body and system level parent instance on the digital twin platform, a one-way write-only data channel is opened for each component on the corresponding PLC or edge gateway, and the original sampling information including vibration, temperature rise and current is written, and the entity serial number, installation coordinates and unique binding identifier of the component level image-data channel are recorded in the index table established on the platform side; S2, the historical operation and maintenance records of the warehouse system are aggregated to generate a degradation reference set, all components in the index table are divided into multiple categories according to the dominant sensing mode, a parameter-constrained lightweight model blueprint is constructed for each category, the blueprint is packaged as a container image and distributed to each component twin, the image is trained locally using the degradation reference set and outputs a health degree scale, which is then synchronized to the subsystem scheduling layer through the read-only index channel; S3, deploy a federated learning scheduling service on the execution system or cloud side, distribute participation identifiers and gradient upload quotas to all registered component twins in the index table according to the warehouse site shift rhythm, and upload local gradients after each component twin completes local weight update; the scheduling service calls the built-in hierarchical scene weighted aggregation operator to obtain a system-level degradation parameter set according to the local gradients uploaded by each component twin, and broadcasts the system-level degradation parameter set with version number for each component to replace it once; S4, instantiate the component nodes, subsystem nodes and business consequence nodes as a three-layer causal Bayesian graph, update the bottom-layer priori using the global health vector obtained by the federated learning scheduling service in step S3, infer the synchronous probability surface from bottom to top, perform depth-first path scoring on the causal graph after sliding window and business occupancy threshold clipping, obtain high-risk fault chains and package them as event streams to synchronize to the work order system and the scheduling system to automatically generate maintenance work orders and adjust picking or stacking scheduling.

2. The predictive maintenance method of claim 1, wherein, In step S1, sequentially instantiating component level image, subsystem level virtual body and system level parent instance for electromechanical components coded and scanned on the digital twin platform comprises: instantiating component level image for each electromechanical component on the digital twin platform, and synchronously writing entity serial number and installation coordinates at this time; after instantiating component level image for all electromechanical components, calling the platform aggregation interface to combine the instantiated component level images of all electromechanical components into multiple subsystem level virtual bodies in turn; then, mounting all subsystem level virtual bodies under the system level parent instance to form the system level parent instance.

3. The predictive maintenance method of claim 1, wherein, In step S2, the historical operation and maintenance records of the aggregation warehouse system generate a degradation reference set, including: aggregating existing operation and maintenance records of the warehouse system into a historical data table, retrieving each record in the historical data table to an index table according to the entity serial number, and backtracking to lock the complete life cycle segment of the component; extracting the failure time point by comparing the operation and maintenance log during the locking period; binding the failure time point with the same period sensing curve in the digital twin platform to generate a degradation label; writing the degradation label, the same period sensing curve, the process link and the load into the metadata to accumulate and form a degradation reference set that can be directly referenced by the model.

4. The predictive maintenance method of claim 1, wherein, In step S2, the read-only index channel is opened to monitor the life cycle, and when the component enters the maintenance or replacement state and triggers the marker, the channel is immediately frozen.

5. The predictive maintenance method of claim 1, wherein, In step S3, after the local weight update of each component twin, the local gradient uploaded by each component twin includes: the gradient vector obtained after each component twin completes the local weight update is packaged into a message; the twin identification, the subsystem number to which it belongs and the generation timestamp are written in the message in turn, and the edge gateway pushes it to the federated learning scheduling service through an encrypted channel between the edge gateway and the federated learning scheduling service; the edge gateway side synchronously sets an upload rate threshold, and if the message queue length reaches the upload rate threshold, the queue is automatically delayed.

6. The predictive maintenance method of claim 1, wherein, In step S3, the scheduling service calls the built-in hierarchical scene weighted aggregation operator to obtain a system-level degradation parameter set according to the local gradient uploaded by each component twin, including: the weighted aggregation operator establishes a subsystem for gradient vectors from the same conveying section, sorting machine or lane stacking unit; then, real-time call the digital twin platform to obtain three on-site indexes of device series depth, near one-shift order throughput and historical downtime cost to generate a local coefficient for each device; after weighting at the subsystem level, the weighted aggregation operator reads the task priority, spare parts availability and manual access difficulty for each subsystem draft to calculate the global coefficient; finally, output the system-level degradation parameter set.

7. The predictive maintenance method of claim 6, wherein, The calculation formula for obtaining the system-level degradation parameter set is: ； where g i is the gradient vector of the ith equipment; B p is the pth subsystem; is the equipment series depth coefficient, the deeper the weight is larger; is the near-order flow coefficient, the more busy the weight is larger; is the historical downtime cost coefficient, the higher the cost the weight is larger;f p is the subsystem task priority coefficient; g p is the subsystem spare parts availability coefficient, the more scarce the weight is larger; m p is the subsystem manual approach difficulty coefficient, the more difficult the weight is larger; P is the total number of subsystems; G is the system-level degradation parameter set.

8. The predictive maintenance method of claim 1, wherein, In step S4, the depth-first path scoring algorithm is divided into three segments: The first segment, path enumeration: for each bottom component node, a four-tuple of the current node, the cumulative layer number, the cumulative probability, and the minimum health degree is stacked; when expanding the child node, read the cargo flow carrying value and the machine pivot level in the digital twin, and if the product of the two is lower than the threshold, backtrack immediately; The second segment, real-time pruning: when the search depth reaches the top business node and the path is still retained, the conditional probability product and the path minimum health degree are calculated immediately, and if either of the two is lower than the scheduling threshold, the path is directly discarded; The third segment, risk grading: for each path filtered through the first two segments, continue to read the spare parts accessibility coefficient and the manual access difficulty coefficient from the digital twin, and then introduce a business layer uniform weight constant to finally determine the risk score.

9. The predictive maintenance method of claim 8, wherein, The risk score is determined by the following formula: ； where D m is the set of component nodes on path h d is the health scale of node d, output by federated learning, ranging [0, 1]; is the set of directed edges on path p e is the conditional probability of edge e, given by the causal graph prior; u1 is the global weight constant for spare parts accessibility, the more scarce the value is larger; u2 is the global weight constant for manual approach difficulty, the more difficult the value is larger; is the spare parts accessibility coefficient of path ; is the manual approach difficulty coefficient of path ; is the risk score of path .

10. The predictive maintenance method of claim 1, wherein, It also includes: S5, disassemble the high-risk fault chain by component dimension to generate risk write-back package, write into the feature risk label table of the corresponding twin body through the encryption channel, list the features with importance higher than the threshold in the key upload list, improve the frequency of the next round of gradient upload, and implement gradient freezing on the features not listed; S6, the federal dispatch service aggregates the feature risk label table of each twin body to generate an upload image, enables a strengthened synchronization strategy for components marked as high-risk, sets up a special message and bandwidth reservation, enables a saving synchronization strategy for low-risk components to extend the upload period, and writes the synchronization period and bandwidth changes into the dispatch log before broadcasting to the operation and maintenance board and network management system to complete the closed loop from risk judgment to communication dispatch.

Images