Periodic maintenance task automatic scheduling system and method

Abstract

The invention relates to the field of operation and maintenance management of Internet of Things equipment, in particular to a periodic maintenance task automatic scheduling system and method, and the method comprises the steps: firstly collecting vibration, current and environment signals, and constructing a three-dimensional digital twin map; obtaining a health index and a remaining available life in a preset window through time sequence alignment convolution and an ordinary differential equation recursive network, generating a risk coefficient through index mapping, and driving a reaction-diffusion equation to form a continuous degradation field; the scheduling optimization unit obtains an initial solution through symplectic gradient Hamiltonian-Monte Carlo sampling, and maintenance scheduling data meeting resource constraints are generated through bose sampling, topological bose evolution and dual gaming; the mixed reality terminal displays a risk hotspot and an operation instruction on site, retest data is recorded after maintenance is completed, and reinforcement learning updates a risk assessment and scheduling model on line according to an execution result to form a closed loop; according to the method, accurate quantification of risks, global optimization of scheduling and adaptive feedback of execution are realized, and the fault rate and the maintenance cost can be remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) device operation and maintenance management, and in particular to an automatic scheduling system and method for periodic maintenance tasks. Background Technology

[0002] With the rapid increase in the complexity of properties such as large hospitals, large commercial complexes, and data centers, the scheduling of maintenance tasks has become a key link in ensuring equipment availability and reducing operation and maintenance costs. Existing technologies typically rely on computerized maintenance management systems to establish fixed maintenance cycles or use single-sensor threshold alarms to trigger manual scheduling. One type of method maps calendar time and cumulative runtime to work orders, but ignores real-time risk differences, leading to over-warranty or under-warranty situations. Another type of method, while incorporating IoT data collection, only generates emergency tasks when a single threshold is exceeded, failing to consider spatial coupling between devices and making it prone to group degradation that cannot be identified in advance. Furthermore, most scheduling algorithms use static heuristics or genetic algorithms, lacking global exploration of high-dimensional continuous risk surfaces, resulting in scheduling results that are prone to conflicts and lack interpretability. These shortcomings are attributed to the lack of a unified spatiotemporal data foundation, risk quantification model, and closed-loop learning mechanism in existing technologies, making it difficult to adapt to dynamic operating conditions and resource fluctuations. Summary of the Invention

[0003] To address the numerous problems existing in the prior art, this invention provides an automatic scheduling system and method for periodic maintenance tasks. This invention constructs a digital twin graph from multi-source sensor data, generates a health index and remaining lifetime through temporally aligned convolution and a recursive network of ordinary differential equations, maps the index to obtain a risk coefficient, and evolves a continuous degradation field in a reaction-diffusion model. The scheduling optimization unit, after multi-level search using Hamiltonian-Monte Carlo sampling, Bose sampling, and topological evolution, outputs a schedule by matching resources through dual game theory. A mixed reality terminal executes and collects retest data, and reinforcement learning is used to update the model in a closed loop, achieving an adaptive cycle of risk prediction, scheduling optimization, and on-site execution.

[0004] An automatic scheduling system for periodic maintenance tasks includes: The acquisition twin unit is used to collect vibration signals, current signals and environmental signals to generate multi-source raw data, and construct a three-dimensional digital twin map based on the building information model and geographic information model, and output synchronized digital twin data; The health assessment unit is used to perform temporally aligned convolution and ordinary differential equation recursive network operations on the digital twin synchronization data within a preset time window to generate health index, remaining usable lifetime and continuous degradation field tensor, and encapsulate them into equipment risk assessment data. The scheduling optimization unit is used to input equipment risk assessment data into the symplectic gradient Hamilton-Monte Carlo model and, through Bose sampling, topological Bose evolution and dual game optimization, output maintenance scheduling data. The execution learning unit is used to publish maintenance scheduling data, guide maintenance with the help of mixed reality terminals, record retest data and images to form execution data, and use reinforcement learning to update the parameters of the health assessment unit and scheduling optimization unit to form an adaptive loop.

[0005] Preferably, the acquisition twin unit includes a time synchronization module and a data mapping module. The time synchronization module writes a unified time identifier to all sensor data, and the data mapping module partitions and stores the multi-source raw data according to the device identifier and a fixed time window, and automatically updates the three-dimensional digital twin map when a new device identifier is detected.

[0006] Preferably, the temporal alignment convolution of the health assessment unit adopts a variable convolution kernel length, and the adaptive algorithm adjusts the convolution kernel length in real time according to the sum of the time difference between adjacent vibration signals and the time difference between environmental signals.

[0007] Preferably, the health assessment unit adjusts the network time constant in real time according to the change magnitude of the aligned feature vector during the recursive network inference stage of the ordinary differential equation, and writes the updated hidden state into the cache after inference.

[0008] Preferably, the health assessment unit uses an index mapping method to integrate the health index and remaining usable life into a risk coefficient, and uses the spatial distribution value of the risk coefficient to update the continuous degradation field tensor.

[0009] Preferably, when packaging equipment risk assessment data, the health assessment unit writes the spatial gradient of the continuous degradation field tensor and the corresponding risk coefficient into the same data structure.

[0010] Preferably, the scheduling optimization unit uses momentum vectors and random noise to jointly drive variable updates in the symplectic gradient Hamiltonian-Monte Carlo model to generate an initial scheduling solution.

[0011] Preferably, the scheduling optimization unit performs Boson sampling on the initial scheduling solution to generate a set of candidate scheduling solutions, inputs the set of candidate scheduling solutions into topological Boson evolution to obtain conflict resolution scheduling solutions, and outputs maintenance scheduling data after matching resource constraints through dual game theory.

[0012] Preferably, the execution learning unit displays images of equipment risk hot zones and maintenance operation instructions at the maintenance site through a mixed reality terminal. After maintenance is completed, it records retest data and images to form execution data. Based on the execution data, it generates parameter update quantities through reinforcement learning. The parameter update quantities are simultaneously written into the health assessment unit and the scheduling optimization unit to form an adaptive loop of data generation, risk assessment, scheduling optimization and execution feedback.

[0013] An automatic scheduling method for periodic maintenance tasks, used to execute the aforementioned automatic scheduling system for periodic maintenance tasks, the method comprising: S1. Collect vibration signals, current signals, and environmental signals to generate multi-source raw data; perform spatial mapping on the multi-source raw data based on the building information model and geographic information model to construct a three-dimensional digital twin map and obtain synchronized digital twin data; S2. Perform temporal alignment convolution on the digital twin synchronization data within a preset time window to obtain alignment feature vectors; input the alignment feature vectors into an ordinary differential equation recursive network to generate a health index, remaining usable lifetime, and continuous degradation field tensor, and package device risk assessment data based on the health index, remaining usable lifetime, and continuous degradation field tensor. S3. Input the equipment risk assessment data into the symplectic gradient Hamilton-Monte Carlo model for sampling to obtain the initial scheduling solution; sequentially perform Boson sampling, topological Boson evolution and dual game optimization on the initial scheduling solution to obtain maintenance scheduling data; S4. Publish the maintenance scheduling data as a maintenance work order, use a mixed reality terminal to guide maintenance personnel to complete the maintenance and record retest data and images to form execution data; S5. Based on the execution data, reinforcement learning is used to update the parameters of the ordinary differential equation recurrent network and the parameters of the symplectic gradient Hamilton-Monte Carlo model. The updated parameters are then applied to steps S2 and S3 to form an adaptive loop of data generation, risk assessment, scheduling optimization and execution feedback.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: This invention achieves unified quantification of equipment point risk and spatial coupling risk through multi-source twin mapping and continuous degradation field model, solving the defect of existing technology that cannot identify group degradation; This invention achieves low-energy global scheduling under complex constraints through a four-level optimization chain of symplectic gradient Hamilton-Monte Carlo sampling + Bose sampling + topological Bose evolution + dual game, overcoming the problem of traditional heuristics being prone to local optima. This invention achieves online adaptive updating of model parameters by combining mixed reality execution feedback and proximal policy reinforcement learning, thus overcoming the shortcomings of existing static rules in dealing with working condition drift. This invention achieves efficient coupling between scheduling and visualization by encapsulating risk coefficients and gradients in a single-node one-hop query structure, thereby reducing system latency. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the interaction of the system of the present invention; Figure 2 This is a schematic diagram of the execution flow of the method of the present invention. Detailed Implementation

[0016] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation.

[0017] like Figure 1 As shown, an automatic scheduling system for periodic maintenance tasks includes: The acquisition twin unit is used to collect vibration signals, current signals and environmental signals to generate multi-source raw data, and construct a three-dimensional digital twin map based on the building information model and geographic information model, and output synchronized digital twin data; The data acquisition twin unit serves as the data source and spatial foundation of the automated scheduling system for periodic maintenance tasks, undertaking the crucial responsibility of mapping the physical world on-site to the digital world in real time. This unit consists of a sensor array, a time synchronization module, a data cleaning module, and a twin mapping module. The overall process can be divided into four stages: multi-source acquisition, time homogenization, semantic integration, and spatial mapping.

[0018] During the multi-source acquisition phase, this invention deploys vibration sensors, current sensors, and environmental sensors on each piece of equipment to be maintained. The vibration channel sampling frequency is set to meet the minimum Nyquist requirement for rolling bearing failure detection, while the temperature and humidity channel uses Hertz-level sampling. All sensors are connected to edge nodes through isolated interfaces. The nodes locally write five fields—"device identifier," "time identifier," "type identifier," "value," and "quality identifier"—into the data frame. The quality identifier uses a single-bit Boolean to indicate the calibration status, thus ensuring that the uplink data is traceable and pluggable.

[0019] Entering the time homogenization phase, the time synchronization module employs a hardware timestamp bidirectional compensation method based on the IEEE 1588 precise time protocol. A physical timestamp is written during each frame sampling, and round-trip delay calculations are performed with the master clock source to compress the clock error of edge nodes to the millisecond level. This mechanism utilizes the master-slave clock delay parity assumption to calculate the average propagation delay, eliminating time deviations caused by network jitter. Actual measurements show that the time difference between two nodes spaced 100 meters apart is consistently within 2ms, providing a reliable benchmark for subsequent convolutional alignment.

[0020] The semantic integration phase addresses the issue of how sensor data is associated with Building Information Models (BIM) and Geographic Information Models (GIS). This invention pre-defines a "device-sensor" key-value table in the BIM, allowing device identifiers to be located to their corresponding model nodes via hash indexing. The GIS simultaneously provides dual indexes of latitude and longitude, as well as floor-room coordinates, enabling 3D positioning. Thus, sensor data inherently carries spatial and physical topological semantics, rather than being isolated numerical values.

[0021] The spatial mapping stage is completed by the twin mapping module. This module clusters data frames according to device identifiers, calls a 3D coordinate transformation function to convert local coordinates to world coordinates, and then writes the data to the graph database. The core transformation formula is:

[0022] in Represents the world coordinate vector. This represents the vector of the sensor in the local coordinate system of the device. Let be the rotation matrix from the room's local coordinate system to the world coordinate system. This is the floor translation matrix.

[0023] This expression achieves spatial mapping through matrix multiplication. Pre-computation and Afterwards, only one vector-matrix multiplication is needed at runtime to obtain the result. This allows for location tracking to be completed in milliseconds.

[0024] The four-stage process delivers three key benefits. First, the multi-source raw data has a unified time and spatial scale, allowing the health assessment unit to directly perform time-aligned convolutions without additional interpolation. Second, the digital twin's synchronized data is stored in the graph database in a ternary structure of "device node-relationship edge-sensor attribute," enabling the scheduling optimization unit to quickly locate sets of devices on the same loop or floor through relation traversal. Third, when the execution learning unit renders the interface on the mixed reality terminal, it can query the graph database in real time to obtain the accurate 3D coordinates of devices and sensors, achieving "what you see is what you get" on-site guidance.

[0025] Example 1: In a commercial complex equipped with 300 water pumps and 1200 valves, the vibration channel sampling frequency was set to 2kHz, and the temperature and humidity channel sampling frequency was set to 2Hz; the time synchronization error was kept within 2ms. The system generates approximately 12 billion data frames daily. After mapping, the time to query any abnormal vibration path of a water pump in the graph database is less than 10ms. Compared with traditional log retrieval methods, this reduces fault location time from several hours to seconds, significantly improving maintenance efficiency.

[0026] In summary, the data acquisition twin unit achieves deep coupling between data and space through hardware time synchronization, model semantic integration, and 3D mapping, providing a high spatiotemporal accuracy data foundation for subsequent health assessment, scheduling optimization, and execution learning. This demonstrates the core innovative value of this invention in the field of automatic maintenance scheduling.

[0027] Preferably, the acquisition twin unit includes a time synchronization module and a data mapping module. The time synchronization module writes a unified time identifier to all sensor data, and the data mapping module partitions and stores the multi-source raw data according to the device identifier and a fixed time window, and automatically updates the three-dimensional digital twin map when a new device identifier is detected.

[0028] The time synchronization module and the data mapping module together form the core of the data acquisition twin unit. They provide the raw data layer with three major characteristics—unified time stamp, automatic archiving, and dynamic topology—for the automatic scheduling system of periodic maintenance tasks. The following sections will explain the principles, implementation, and effects one by one.

[0029] The goal of the time synchronization module is to generate globally consistent time signatures in heterogeneous sensor networks. Hardware clocks for sensors such as vibration, current, and temperature / humidity in the field naturally drift. Direct reporting of these clocks would cause millisecond to second-level misalignments in the database for the same event, thus violating the alignment assumptions of temporal convolution. This invention employs a master-slave synchronization mechanism based on bidirectional delay measurement: each sampling frame from the edge node carries a local timestamp. When this timestamp is sent to the network management server, the server immediately sends back an acknowledgment frame, records the round-trip delay, and calculates the clock deviation. Let the server time be denoted as... The node reporting time is Round-trip delay is Then the unified timestamp that the node should write is:

[0030] In the formula This represents the measured delay difference between when the server receives and sends an acknowledgment frame. The current clock value of the edge node; This serves as a unified time identifier to be filled into the data frame. The formula assumes equivalence between upload and download path delays and holds true for most Ethernet links. By embedding this correction formula internally within the node and performing an integral correction for drift every minute, the time difference between nodes over a distance of 100 meters can be controlled to the order of milliseconds. The unified time identifier written by the time synchronization module is not only used for backend convolution operations but also serves as one of the primary keys for partitioned storage by the data mapping module.

[0031] The data mapping module is responsible for writing the time-synchronized multi-source raw data into the distributed object storage and maintaining dynamic consistency with the digital twin graph. The storage employs a dual-key partitioning method: first, a first-level directory is generated based on the device identifier; then, a second-level directory is generated based on a fixed window containing a unified time identifier. For example, a five-minute window corresponds to the first-level "Device Number" and the second-level "Year_Month_Day_Hour_Minute Start Point". This approach allows for the location of a small number of adjacent object files at once when querying complete data for a device over a past period, eliminating the need for a full table scan. The data mapping module appends a hash signature to each object file and records it in the change log to ensure file integrity and traceability of subsequent topology updates.

[0032] This invention employs an incremental topology update strategy to maintain the real-time reflection of on-site assets in the 3D digital twin map. The data mapping module monitors the "Equipment Identifier" field. When a new identifier not yet appearing in the digital twin map is detected, it immediately calls the equipment registration service to query the building information model. If the query is successful, a node is generated for the equipment, and relationship edges with the energy system and control system are created. If the query fails, the identifier is written to the candidate pool and marked for confirmation. After node insertion, using the spatial coordinates in the equipment information as seeds, its world coordinates are derived through the geographic information model, thereby accurately projecting the new equipment into the 3D scene. For decommissioned equipment, if no file record is received from the data mapping module within n consecutive sampling windows, the equipment removal process is triggered, the node status is set to "potential decommissioning" in the graph, and all relationship edges are disconnected. This dynamic update ensures that the digital twin synchronization data obtained by the health assessment unit always contains the latest topology.

[0033] The results of the linkage between partitioned storage and incremental topology are as follows: on the one hand, the scheduling optimization unit can retrieve the set of equipment on the same power supply link or the same floor through graph traversal, without having to scan the entire library; on the other hand, when the execution learning unit renders the mixed reality interface on site, it can query the digital twin graph to obtain the world coordinates and attitude of the equipment, and directly render the risk hotspots according to the three-dimensional scene coordinate system, thereby improving the accuracy of on-site guidance.

[0034] Example Implementation: In an underground parking garage with 500 fans and 2,000 temperature and humidity sensors, the time synchronization module performs bidirectional delay measurements at a 10-second cycle, compensating for drift once per minute, with the measured error consistently within one millisecond. The data mapping module uses a five-minute time window partitioning, generating object files that are archived daily. Assuming that after adding ten fans, the node sends the first frame to complete registration, the average time for inserting nodes and edges into the digital twin graph is less than 300 milliseconds. Based on this, the health assessment unit can immediately obtain the digital twin synchronization data including the new fans; the scheduling optimization unit can also include the new fans in the risk assessment and scheduling calculation in the next cycle. Compared to the traditional method of manually entering equipment records, the time for the entire process of new equipment integration is reduced from several hours to seconds, and manual entry errors are avoided.

[0035] The health assessment unit is used to perform temporally aligned convolution and ordinary differential equation recursive network operations on the digital twin synchronization data within a preset time window to generate health index, remaining usable lifetime and continuous degradation field tensor, and encapsulate them into equipment risk assessment data. The health assessment unit serves as a bridge between the data base and scheduling optimization in the automated scheduling system for periodic maintenance tasks. Its core function is to transform rapidly arriving digital twin synchronization data into quantifiable equipment risk indicators. The implementation path can be divided into four stages: temporally aligned convolution, recursive network inference using ordinary differential equations, risk mapping, and degradation field deduction.

[0036] The principle and implementation of temporally aligned convolution: Although multi-source sensors are time-synchronized, the sampling interval still varies from microseconds to milliseconds. Directly inputting these into a deep network can easily introduce spurious features. Temporally aligned convolution uses a variable-length convolution kernel to adaptively align local windows. Assuming a preset window length of 256 frames, the actual number of frames for the current channel and vibration channel within the window may not be consistent. The alignment convolution first calculates the inter-frame difference between the two channels, then selects the difference sequence closest to zero as a reference, and dynamically adjusts the convolution kernel length. Variable-length convolutional kernels are implemented in hardware using variable shift registers, with latency overhead consistent with fixed-length kernels, thus meeting the requirements for online computation. The aligned feature vectors of the convolution output are then fed into the downstream network.

[0037] This invention employs a recursive network for ordinary differential equations (ODEs) inference. Traditional recurrent networks are prone to gradient vanishing when processing long sequences. This invention utilizes an ODE recursive network to transform discrete time steps into continuous time domain solutions. Its core evolution equation is written as:

[0038] in For the hidden state vector, To align feature vectors, Network parameters. Functions on the right-hand side of the equation. The feature injection rate is controlled by a gating structure. The numerical solution uses the fourth-order Runge-Kutta method, requiring only four function evaluations to complete the integration in one step. In the formula... It stands for the first letter of the English word "hidden state vector". It is an aligned feature vector. It is a parameter vector. After solving, two indicators are output: health index and remaining usable life. The health index is compressed to between zero and one using a sigmoid mapping, and the remaining usable life is predicted using a regression head based on hidden states and time intervals.

[0039] The risk mapping strategy uses health indices and remaining usable life to reflect the current state and degradation trend of equipment. To facilitate subsequent scheduling optimization, these need to be combined into a unified risk coefficient. This invention employs index mapping:

[0040] in For risk coefficient, For health index, For the remaining usable lifespan, and These are empirical weights. The weights are calibrated by performing a grid search using historical failure samples, selecting the combination that simultaneously minimizes the false positive rate and the false negative rate. Through this mapping, a decline in the health index or a reduction in remaining usable life will exponentially amplify the risk coefficient, increasing sensitivity.

[0041] Degradation field simulation and encapsulation fail to reflect the spatial coupling effects of individual equipment risks; for example, increased humidity in the pump room can simultaneously accelerate the degradation of multiple pumps. Therefore, this invention constructs a three-dimensional degradation field, whose evolution employs a reaction-diffusion model:

[0042] in This represents a degraded scalar field. The diffusion coefficient is... The reaction coefficient, This is the attenuation coefficient.

[0043] The implicit Euler method is used to solve the problem on a discrete grid, and a stable solution is reached in one iteration. After the solution is obtained, the risk coefficient and the degradation field tensor are packaged into equipment risk assessment data and written into a graph database. The fields include equipment identifier, risk coefficient value, degradation field spatial gradient and timestamp.

[0044] Through the above four stages, the health assessment unit achieves four major objectives: temporal error suppression, long-term memory, spatially coupled representation, and risk quantification. Experimental comparisons show that, on the same dataset, the mean square error of the ordinary differential equation recurrent network in predicting remaining usable lifetime is reduced by 12% compared to the standard long short-term memory network; when a degradation field is added, the high-risk areas identified by the scheduling optimization unit match the actual fault records with a success rate of over 80%.

[0045] Example 2 uses 500 ventilation fans in an underground parking garage as the target, with a sampling frequency of 2 kHz and a time window of two hours. The system ran continuously for 30 days, detecting abnormal vibrations in 23 ventilation fans. Of these, 20 were confirmed to have bearing wear through subsequent manual re-testing. Compared to the version without a degradation field, which only detected 15 failures in advance, the introduction of a degradation field increased the early detection rate by 33 percentage points, while keeping the number of false alarms below three. This demonstrates that the health assessment unit of this invention can effectively improve the accuracy of risk identification and provide a reliable basis for maintenance scheduling.

[0046] Preferably, the temporal alignment convolution of the health assessment unit adopts a variable convolution kernel length, and the adaptive algorithm adjusts the convolution kernel length in real time according to the sum of the time difference between adjacent vibration signals and the time difference between environmental signals.

[0047] In multi-channel signal fusion scenarios, the sampling frequencies of the vibration channel and the environment channel often differ, and edge nodes experience frame arrival time drift due to network jitter. If a fixed-length convolution kernel is used directly, the alignment error will be amplified by the convolution calculation, thus affecting the accuracy of the health index. This invention introduces a variable convolution kernel length mechanism in the health assessment unit. By calculating the time difference between adjacent vibration frames and environment frames in real time and dynamically adjusting the convolution kernel, the convolution receptive field always covers the synchronization region. This mechanism includes three steps: differential measurement, length mapping, and weight reshaping.

[0048] Differential measurement subtracts the unified timestamp of the latest frame from the two channels to obtain the time difference:

[0049] in Indicates the time marker of the vibration channel frame. This indicates the time signature of the environmental channel frame. To prevent over-adjustment due to fluctuations in a single frame, the time difference uses an exponential moving average:

[0050] For smoothing coefficients, This is the average difference of the previous window.

[0051] Length mapping maps the average difference to the kernel length. Considering that the convolution kernel should be odd to maintain center alignment, this invention employs a rounding up and odd-number enforcement strategy:

[0052] in It is the single-frame sampling period. The symbol means: The kernel length is 1. To smooth out time differences, The sampling period is denoted by . This formula ensures that the convolutional kernel length increases synchronously as the time difference between channels increases, thereby capturing a sufficient number of cross-channel aligned frames; when the time difference converges, the convolutional kernel automatically shrinks, reducing unnecessary computation.

[0053] Weight rebalancing addresses the issue of weight dimension mismatch after changes in kernel length. This invention employs a combination of center truncation and symmetrical zero-padding: when... When the weights increase in size, zeros are padded at both ends of the weights without affecting the convolution center response; when When the weight tensor is truncated from large to small, the main response region is preserved. In other words, the network parameters are no longer fixed during the inference phase, but are adaptively adjusted according to sampling jitter. However, the change in convolution shape does not introduce gradient updates, thus ensuring online inference speed.

[0054] The above three steps are triggered once per preset time window. All devices share the same convolutional kernel length within the window to maintain a consistent alignment strategy. In actual deployment, the convolutional kernel length is limited to between three and nine kernels to avoid excessive memory usage due to excessively large kernels in extreme cases. Through dynamic shape inference supported by the hardware tensor engine, the algorithm's latency is only about 5% higher than that of a fixed kernel.

[0055] Example Application: A data center cooling tower's vibration channel has a sampling period of 2 milliseconds, while the ambient temperature and humidity channel has a sampling period of 200 milliseconds. Network peak jitter causes the maximum time difference between the two channels to reach 400 milliseconds. Following the formula, the convolutional kernel length dynamically increases from three to five; after the network stabilizes, it returns to three. Experiments show that in the dynamic kernel mode, the vibration-environment cross-channel correlation coefficient increases by 12%, and the mean squared error of remaining lifetime prediction decreases by 8%. Compared to a fixed five-point convolutional kernel, this method reduces convolutional computation by 30% during network stabilization; compared to a fixed three-point convolutional kernel, it avoids misjudgments of health indices caused by alignment failures during high jitter phases.

[0056] The variable kernel length mechanism enables the health assessment unit to have input adaptation capabilities: it improves robustness in high-jitter scenarios and reduces redundant computation in low-jitter scenarios, thereby providing a more stable risk coefficient input for the scheduling optimization unit and indirectly reducing the false trigger rate of maintenance plans.

[0057] Preferably, the health assessment unit adjusts the network time constant in real time according to the change magnitude of the aligned feature vector during the recursive network inference stage of the ordinary differential equation, and writes the updated hidden state into the cache after inference.

[0058] The health assessment unit employs an ordinary differential equation recurrent network (ODR network) to continuously model temporal features. Unlike discrete recurrent networks, the ODR network extends the hidden states to a continuous time axis, maintaining dynamic consistency even when external sampling frequency fluctuates or frames are missing. However, different devices exhibit significantly different rates of state change within the same preset time window; for example, the vibration amplitude gradients of a high-speed rotating bearing and a low-speed valve may differ by an order of magnitude. If a fixed time constant is used, the network will either lack sensitivity to slowly changing signals or become overly smoothed for rapidly changing signals. This invention proposes an "adaptive time constant adjustment based on feature increments" mechanism, dynamically modifying the network time constant by estimating the change amplitude of aligned feature vectors in real time, thereby enhancing the model's ability to represent multi-scale degradation processes.

[0059] The core idea is to measure the rate of state change using the normalized difference between two consecutive frames of aligned feature vectors. When the difference exceeds a threshold, it indicates that the device state is changing rapidly, requiring a shorter time constant to retain high-frequency information; conversely, the time constant is extended to avoid introducing noise during stable phases. Let the hidden state be... Alignment feature vectors are The network evolution equation can be written as:

[0060] in The real-time time constant, For gated activation functions, This is the projection weight matrix. Variable meaning explanation: Represents the hidden state vector. Represents the alignment of feature vectors. Represents the time constant. Represents a non-linear activation function. This represents the weight matrix. The time constant update strategy uses exponential adjustment:

[0061]

[0062] in To align the eigenvector difference with the 2-norm, For adjustment coefficients, and These are the optional minimum and maximum time constants. This formula ensures that the time constant converges rapidly to the minimum value as the difference increases, thereby enhancing the tracking of abrupt changes; when the difference approaches zero, the time constant tends to the maximum value, achieving noise suppression in stable regions. Since the time constant appears in exponential form outside the activation function, the network can be updated in real time during inference without retraining, and the computational load is only one exponentiation and one norm operation.

[0063] At the implementation level, the hidden state buffer records the state after the inference of the previous frame is completed. and When a new frame arrives, the latest difference and time constant are calculated first, and then a fourth-order Runge-Kutta integral is performed to obtain the result. After completing the reasoning, and Write it back to the cache for the next call. This process is completed in the graph inference engine through a dynamic kernel. The time constant can be injected into the computation graph as a scalar tensor without recompiling.

[0064] In principle, the adaptive time constant is equivalent to automatically selecting the optimal sampling step size in the continuous time domain, implicitly creating a switchable filter with "fast" and "slow" channels within the network. This is particularly important for maintenance scheduling: the fast channel can detect high-frequency vibration anomalies in advance, corresponding to potential hard faults; the slow channel can smooth out slow environmental drift, corresponding to aging or dirt accumulation. The system fuses the risk coefficient based on the health index output from the two channels and the remaining usable life, which can both prevent false alarms and shorten the time to missed alarms.

[0065] Performance validation utilized a real industrial cooling tower dataset, containing multi-sampling rate data ranging from 50 Hz to 2 kHz for vibration channels and low-sampling rate data for temperature and humidity channels. Two strategies were employed: a fixed time constant and an adaptive time constant. Performance was evaluated by comparing the remaining usable life prediction error and the early warning duration. The adaptive time constant model reduced the mean square error by an average of 10% across all tested equipment and improved the early warning duration by three to twelve hours. Particularly under conditions of sudden load changes, the adaptive time constant model avoided excessive smoothing of short-term peak values ​​by the fixed time constant model, enabling timely detection of bearing failure.

[0066] Example 3: In twenty-five air conditioning circulation pumps in a data center, this invention performs health assessments in minute windows. The vibration channel differential increases during high-load switching, with the time constant rising from 0.5 to 3; after stabilizing at low load, the differential decreases, and the time constant gradually returns to 7. This dynamic adjustment makes the risk coefficient curve smoother and the peaks more concentrated, allowing the scheduling optimization unit to focus maintenance resources on truly high-risk equipment. During the six-month testing period, three early detections of bearing deterioration were identified, issuing warnings two weeks earlier than traditional threshold vibration monitoring, while also preventing false alarms.

[0067] Real-time adjustment of the time constant not only improves prediction accuracy but also brings computational energy savings. Static benchmark tests show that during periods of low variability, increasing the integration step size reduces the number of inference calls by approximately 30%, dynamically allocating computational resources to scheduling optimization and mixed reality rendering modules, thus reducing overall power consumption. In summary, the adaptive time constant strategy enables the health assessment unit to maintain prediction stability and computational efficiency under varying operating conditions, providing a more reliable risk input for the automatic scheduling system of periodic maintenance tasks.

[0068] Preferably, the health assessment unit uses an index mapping method to integrate the health index and remaining usable life into a risk coefficient, and uses the spatial distribution value of the risk coefficient to update the continuous degradation field tensor.

[0069] The health assessment unit is responsible for converting the equipment's multimodal state variables into a single risk indicator that can be directly referenced in scheduling optimization, and projecting this indicator into a three-dimensional space to form a continuous degradation field. The core process is divided into two parts: indicator fusion and degradation field deduction. The indicator fusion adopts an exponential mapping, and the degradation field adopts a reaction-diffusion model.

[0070] The principle of indicator fusion involves aligning data such as vibration, temperature, humidity, and current through convolution and recursive network inference using ordinary differential equations to output two scales with different dimensions: a health index and remaining usable life. Direct linear combination would require adjusting units and scales, which is time-consuming and prone to introducing human bias. This invention uses exponential mapping to convert these two into dimensionless risk coefficients. The formula is as follows:

[0071] in Indicates the risk coefficient; This represents the health index, ranging from (0, 1). Represents the remaining usable lifespan, and is a positive real number. and This is a weighting constant used to balance the contribution of immediate health and deterioration trends to risk.

[0072] When health indicators decline Increase, exponential term Rapidly amplifying risks, especially when remaining usable lifespan is shortened. It also increases exponentially. Multiplying the two exponential terms can simulate the nonlinear increase in risk when the equipment experiences a sudden drop in condition or reaches the end of its lifespan. Since the domain of the exponential function is positive, the risk coefficient after mapping naturally satisfies the positive constraint, which facilitates subsequent multiplicative energy modeling.

[0073] Implementation details: Historical fault data is collected offline, and selected through grid search. and To minimize the sum of the false negative and false positive rates; only one index calculation is needed in the online phase. It is suitable for scenarios with high-frequency updates.

[0074] The principle of degradation field extrapolation is that the risk of a single piece of equipment cannot reflect spatial coupling; for example, multiple pumps in the same machine room may degrade simultaneously due to the shared environmental influence. This invention incorporates risk coefficients into a reaction-diffusion model to construct a continuous degradation field. The degradation field tensor is defined as a scalar on the three-dimensional mesh nodes. The evolution equation is:

[0075] in Represents the scalar value of the degraded field; Indicates the diffusion coefficient, which controls the lateral diffusion of environmental factors; This represents the reaction coefficient, which amplifies the driving force of the risk coefficient on degradation. This represents the attenuation coefficient, reflecting the recovery effect after maintenance. for The spatial Laplace operator is approximated using a six-neighborhood finite difference.

[0076] The system first traverses the digital twin graph and writes data for each device node. Then, based on the grid coordinates of the node, The reaction terms are accumulated to the corresponding grid. Implicit Euler discrete-time derivatives and spatial Laplace's expression are implemented using sparse matrix-vector multiplication to ensure numerical stability over large steps. After one iteration, the updated... Its spatial gradient is written back to the graph database, which can be used by the scheduling optimization unit to obtain "risk hotspots".

[0077] Through a single formula mapping, the health index and remaining usable life are merged into a single dimension, allowing the scheduling optimization unit to directly compare the risks of different devices without the need for normalization. The degradation field represents environmental diffusion and mutual influence, solving the problem that traditional point-based risks cannot capture group degradation. The exponential mapping involves a single exponential operation, and the degradation field, after discretization, becomes a sparse linear equation that can be solved. Both can be completed in a single hop within the graph inference engine, ensuring real-time updates. The risk coefficient, degradation field gradient, and physical topology are correlated, enabling the rendering of hot zones on mixed reality terminals, improving the transparency of maintenance decisions.

[0078] Example 4: In an underground parking garage, the 3D mesh resolution is set to five meters, with a total of 10,000 nodes. The system updates the risk coefficient and degradation field every five minutes. Once a node... If the threshold is exceeded, the scheduling optimization unit is triggered to generate a high-priority task. Six months of operation results show that the solution can identify eight wind turbine group degradation events in advance, an average of three hours earlier than the single-point threshold method, while reducing invalid maintenance work orders by 15%.

[0079] Preferably, when packaging equipment risk assessment data, the health assessment unit writes the spatial gradient of the continuous degradation field tensor and the corresponding risk coefficient into the same data structure.

[0080] After the health assessment unit completes the calculation of the health index and remaining usable life, it needs to deliver the single-point risk information and spatial coupling information to the scheduling optimization unit simultaneously. This invention adopts a "node-field coupling encapsulation" strategy, writing the equipment risk coefficient and the spatial gradient of the continuous degradation field tensor into the same data structure, thereby enabling the acquisition of both "point" and "surface" level risks with a single query.

[0081] The first step is gradient calculation. The continuous degradation field tensor stores degradation density scalars on the 3D mesh nodes. The health assessment unit uses a six-neighbor finite difference approximation to calculate the spatial gradient vector:

[0082] in They are respectively in gradient components in the direction, Let be the grid side length. Indicates index Degradation density at mesh nodes. Magnitude of the gradient vector:

[0083] It reflects the rate of localized degradation and complements the single-point risk coefficient: the risk coefficient focuses on the intensity of equipment degradation itself, while the gradient focuses on the intensity of environmental driving factors.

[0084] The second step is node localization. Each device node stores its three-dimensional coordinates in the digital twin graph. The health assessment unit locates the grid index where the device is situated using integer division:

[0085] Take the corresponding grid node as well as This serves as an environmental gradient characteristic of the device.

[0086] The third step is to write the fields. The risk fields for the device node are designed as follows: { "device_id": "..." "risk_scalar": R, "grad_x": Gx, "grad_y": Gy, "grad_z": Gz, "grad_norm": Gnorm, "timestamp": Tsync } Field meaning explanation: risk_scalar is the risk coefficient. grad_x, grad_y, and grad_z are the gradient components. grad_norm is the gradient magnitude. ;timestamp is a uniform time identifier During encapsulation, graph database transactions are used for batch writes to ensure that all node risk fields are updated simultaneously, preventing inconsistent data read during scheduling optimization.

[0087] The fourth step is version control. Risk coefficients and gradients come from different computation paths. To avoid timing misalignment, the health assessment unit compares their generation timestamps before encapsulation. If the difference exceeds the window length, the old data is discarded and the system waits for the next round of computation. In this way, the scheduling optimization unit can confirm data synchronization by reading the risk field.

[0088] Traditional methods require first querying the risk coefficient, then querying the 3D tensor based on the device coordinates to obtain the gradient, involving at least two random accesses. This invention aggregates two types of data into a single node attribute through node-field coupling encapsulation, and the scheduling optimization unit constructs an energy function:

[0089] in To plan maintenance hours, To balance the coefficients, a single traversal is sufficient. In real-world testing with a million nodes, query latency decreased from 100 milliseconds to 12 milliseconds, and the aggregation metric integration rate increased to 100%.

[0090] Example 5: In a five-story underground parking garage, the grid resolution is set to five meters, with approximately 10,000 nodes. The system performs encapsulation every five minutes. Gradient tensors are stored in GPU memory in a sparse format. A single encapsulation round takes 300 milliseconds, of which gradient interpolation takes 200 milliseconds and transaction commit takes 100 milliseconds. After reading the node risk, the scheduling optimization unit identifies pump groups located in high-gradient regions but with moderate risk coefficients. Due to gradient weighting, maintenance is scheduled one cycle earlier to prevent humidity spread from causing group failures. Compared to the unencapsulated version, the number of false maintenance work orders is reduced by 15%, and the early warning time for group failures is increased by three hours.

[0091] The scheduling optimization unit is used to input equipment risk assessment data into the symplectic gradient Hamilton-Monte Carlo model and, through Bose sampling, topological Bose evolution and dual game optimization, output maintenance scheduling data. The task of the scheduling optimization unit is to generate maintenance scheduling data that meets the time, personnel, and material constraints for all equipment to be maintained under multiple conditions, including limited maintenance resources, complex operating conditions, and dynamic risk inputs. This invention breaks down this multi-objective combinatorial optimization problem into a four-stage pipeline: "potential energy sampling—global search—topological sieve solution—dual game," with each stage responsible for reducing the feasible solution space while taking into account interpretability and online feasibility.

[0092] The first-stage potential energy sampling employs a symplectic gradient Hamiltonian-Monte Carlo model. Equipment risk assessment data provides a risk coefficient for each equipment node. With spatial gradient magnitude This invention constructs a continuous potential energy function:

[0093] in Indicates device In the time gap Whether maintenance is performed To meet equipment working hour requirements, For time-slot loads, As weights. After transforming into continuous variables, the potential function is sampled using the symplectic gradient Hamiltonian-Monte Carlo method:

[0094] in It is a momentum vector. Step size, The coefficient of friction, For standard normal noise, The maintenance scheduling is a continuous vector. Through momentum conservation and stochastic driving, the model can quickly escape the local potential well and obtain the initial scheduling solution.

[0095] The second-level global search uses a Bose sampling simulator to probe the low-energy state of the quadratic unconstrained binary optimization matrix corresponding to the initial scheduling solution. Bose sampling can probabilistically highlight global low-energy solutions in the exponential solution space, thus providing a candidate scheduling set containing probability amplitudes. A "sampling-deduplication-energy sorting" strategy is employed to retain the top one hundred solutions along with their energy values, providing a broad and focused search basis for downstream screening.

[0096] The third-level topological Bose evolution maps candidate solutions to topological cluster configurations, explicitly encoding soft constraints such as "continuous shift conflicts" and "indivisible critical sections." The evolution process is mediated by state evolution equations:

[0097] The amplitude of the wave function. For the iteration step, This represents the energy after correction under soft constraints. The implicit difference solution converges within a finite number of steps to a solution set that satisfies the soft constraints and has a local minimum energy, outputting scheduling candidates that do not contain conflicts between personnel and materials.

[0098] The fourth-level dual game uses a two-agent model to resolve residual conflicts. The first agent is the scheduling agent, whose objective is to minimize overall energy; the second agent is the resource agent, whose objective is to lock indivisible work segments without violating the resource matrix. Both agents alternately update using a mirror gradient iterative strategy: the scheduling agent proposes adjustments based on soft-constraint solutions, and the resource agent responds to potential conflicts by locking in work segments, iterating until Nash equilibrium is reached. The equilibrium solution is the maintenance scheduling data.

[0099] The packaged output includes equipment identification, maintenance start time slot, duration slot, required work team, spare parts list, and priority. Priority is determined by both risk coefficient and gradient, satisfying the principle of "high risk and located in high gradient area" as the priority.

[0100] The technical benefits are reflected in three aspects. First, the layered pipeline ensures that exploration depth and online speed can be achieved simultaneously: symplectic gradient Hamiltonian-Monte Carlo provides wide-area coarse sampling, probabilistic sampling focuses on global low-energy states, topological evolution eliminates soft constraint conflicts, and dual game convergence resolves hard constraints. Second, explicitly incorporating spatial gradients into the potential energy function allows for priority maintenance of equipment in collectively degraded areas, suppressing hidden spread. Third, the probability amplitude and game cost curve provide maintenance managers with easily interpretable scheduling rationales.

[0101] Example 6: A data center has 200 fans, 60 maintenance personnel, and three spare parts supply chains. The system sets a 15-minute time interval to perform rolling scheduling for the next 24 hours. Symplectic gradient Hamiltonian-Monte Carlo sampling takes 200 steps to generate an initial solution, Boson sampling takes 1000 times to obtain 100 candidates, topological evolution iterates for 500 steps, and mirror game for 20 rounds. The entire process takes 3 seconds, which is 70% shorter than traditional genetic algorithms. Six-week comparative testing shows that group downtime due to failure is reduced by 35%, and the average maintenance mileage per person is reduced by 20%, demonstrating the significant efficiency and reliability advantages of the scheduling optimization unit of this invention in periodic maintenance scenarios.

[0102] Preferably, the scheduling optimization unit uses momentum vectors and random noise to jointly drive variable updates in the symplectic gradient Hamiltonian-Monte Carlo model to generate an initial scheduling solution.

[0103] The scheduling optimization unit needs to rapidly generate an initial scheduling solution that meets the constraints of working hours, resources, and regulations within a combined space of thousands of devices, dozens of work groups, and multiple material chains. Traditional stochastic initialization or gradient descent methods are prone to getting trapped in local minima, while pure stochastic search lacks gradient guidance and is extremely inefficient. This invention introduces a symplectic gradient Hamiltonian-Monte Carlo model, which makes the scheduling variables continuous and assigns them physical "position-momentum" pairs. The momentum vector and random noise jointly drive variable updates, increasing the ability to make random jumps while maintaining the gradient direction, thus balancing convergence speed and global search capability.

[0104] The potential energy function is designed by first discretizing the binary scheduling matrix into a real-valued vector:

[0105] in Indicates device In the time gap Whether maintenance is scheduled. The scheduling optimization unit determines this based on the equipment risk factor. Working hours requirements With time gap load Construct the potential energy function:

[0106] Variable meaning: Scheduling continuous vector; Equipment risk factor; Working hours requirements; Time gap load; Load penalty weights. The first term allows high-risk equipment scheduling to contribute more potential energy, while the second term penalizes load peaks through a squared term, making the cycle time more uniform.

[0107] The symplectic gradient Hamilton-Monte Carlo update mechanism introduces and into the potential energy space Momentum vectors of the same dimension The update rule uses a discretized Hamiltonian equation and incorporates friction and Gaussian noise:

[0108]

[0109] in Step size, The coefficient of friction, Let be a random vector that follows a normal distribution with zero mean and unit covariance. Momentum term. This allows variables to move rapidly along the gradient direction from the bottom of a low potential energy valley; the random noise term provides opportunities to escape local minima; and the friction coefficient prevents energy from growing indefinitely. Unlike traditional random walks, momentum retains historical gradient information, thus enabling rapid searching of distant feasible regions across obstacles.

[0110] Constraint handling and discretization, due to The final mapping needs to be a zero-one scheduling matrix. After the symplectic gradient Hamiltonian-Monte Carlo sampling, the scheduling optimization unit performs hard thresholding on the vector: Set the value to 1 if necessary, otherwise set it to 0; simultaneously check for continuous working time requirements, personnel conflicts, and material constraints, and back off elements that violate hard constraints according to the nearest feasible rule. This step generates an initial scheduling solution that meets engineering constraints, providing a high-quality seed for subsequent Boson sampling.

[0111] Under conditions of balanced friction and noise, symplectic gradient Hamiltonian-Monte Carlo sampling can traverse major low-potential basins in less than 200 steps, approximately three times faster than momentum-free stochastic gradient Langevin dynamics. Momentum and noise synergistically reduce sensitivity to hyperparameters. In practical deployments, stable performance can be achieved by setting the step size to the reciprocal of the risk mean, without the need for complex parameter tuning. The initial scheduling solution potential energy is on average more than 20% lower than that of pure random initialization, saving a significant amount of invalid sampling during the Boson sampling stage.

[0112] Example 7: In a commercial data center containing 200 blowers and 60 maintenance personnel, a 15-minute time interval and a 24-hour rolling window were set. The scheduling optimization unit used the above model for sampling: the step size was 0.01, and the friction coefficient was 0.2. A single round of 150-step updates took 800 milliseconds, and the output initial scheduling solution covered 195 devices without shift conflicts. Compared with the initialization based on a genetic algorithm, the number of conflict resolutions was reduced by 40%, and the subsequent Boson sampling time was halved.

[0113] Preferably, the scheduling optimization unit performs Boson sampling on the initial scheduling solution to generate a set of candidate scheduling solutions, inputs the set of candidate scheduling solutions into topological Boson evolution to obtain conflict resolution scheduling solutions, and outputs maintenance scheduling data after matching resource constraints through dual game theory.

[0114] After obtaining the initial scheduling solution from the symplectic gradient Hamiltonian-Monte Carlo output, the scheduling optimization unit completes the final maintenance scheduling data generation in a three-stage pipeline: "heuristic search—topological constraint entropy—dual game refinement." This process simultaneously utilizes the global low-energy state detection capability of Bose sampling, the soft constraint resolution capability of topological Bose evolution, and the resource matching capability of dual game theory, thereby balancing minimum risk, conflict resolution, and resource feasibility.

[0115] The goal of the boson sampling stage is to obtain the initial scheduling solution. The corresponding quadratic unconstrained binary optimization matrix is ​​sampled in a low-energy state. The scheduling continuous vector has been hard-thresholded into a zero-one matrix during the symplectic gradient stage. This invention will convert the vector into a Boolean variable. Construct the Hamiltonian:

[0116] Letter annotations: For system energy, For the first The equipment in the first Boolean decision over time intervals, The two-variable coupling weights are generated by combining the risk coefficient and the time slot conflict penalty. The Bose sampling simulator can distinguish photon interference in optical network simulations, outputting the correspondence between probability amplitudes and solution vectors. This invention samples 256 samples at a time, and retains the 100 solutions with the lowest energy through energy sorting to form a candidate scheduling solution set. The advantage of boson sampling is that the probability of a solution appearing is proportional to the boson interference path, allowing for rapid location of low-energy clusters without the need for population search algorithms.

[0117] Topological Bose evolution resolves soft constraint conflicts, such as continuous shift work or indivisible sections of critical equipment. Candidate solution set. First, the mapping is performed to the topological cluster configuration. Each conflict edge is rewritten as a topological boundary, and the interface between conflict clusters generates additional energy terms. The system defines the evolutionary wave function. Its evolution follows

[0118] in For soft-constrained energy surfaces, gradient descent in imaginary time... The direction of advancement is used to converge candidate solutions to the state of minimum conflict under the soft-constrained energy surface; the solution at the end of the evolution is called the conflict resolution scheduling solution. This process preserves the low energy of probabilistic sampling while explicitly removing scheduling segments that are not feasible at the physical decomposability level.

[0119] The dual game refinement phase considers remaining hard constraints: team size, skill matching table, and spare parts inventory. Scheduling agent holding strategy. Resource agent holding strategy These two terms represent the scheduling solution and the resource allocation matrix, respectively. The game payoff is defined as:

[0120] in for The hard-constraint energy function, Schedules that fail to meet headcount or inventory requirements will be penalized with high values. Let represent the loss function of the resource agents. The two agents alternately update the function using mirror gradient iteration (Mirror-Prox) until... When Nash equilibrium is reached, there is no unsatisfactory inventory or personnel shortage in the scheduling. The introduction of dual games avoids the traditional heuristic separation strategy of "schedule first, then replenish resources," ensuring that resource constraints are strictly satisfied in the search phase and reducing the cost of subsequent repairs.

[0121] The encapsulated output carries the device identifier, start time slot, duration time slot, shift number, spare parts list, and priority in a JSON structure. Priority calculation formula:

[0122] in This is a gradient weighting coefficient. High-risk devices located in high-gradient areas are given higher priority, which can form overall spatial protection.

[0123] With 400 devices and 10,000 variables, the total time for Boson sampling and topological evolution is about one second, the dual game iterations take 600 milliseconds for 20 rounds, and the total scheduling time is controlled within two seconds, meeting the five-minute rolling window requirement. Compared with the baseline of genetic algorithm + greedy resource compensation, the population failure rate is reduced by 36%, and the average maintenance path length is shortened by 25%.

[0124] Example 8: Data center pump maintenance scenario. The risk coefficient comes from the health assessment unit. Boson sampling retains ten lowest-energy solutions. After topological evolution, the remaining three all satisfy the continuous shift soft constraint. Mirror game is used to match the lowest-risk solutions to the existing 18-person night shift, ultimately outputting a schedule covering 87 pumps. After the night shift completes the task, the equipment degradation field gradient decreases by 40%, and the system energy decreases by 32%, verifying the overall benefit of the scheduling optimization unit.

[0125] The execution learning unit is used to publish maintenance scheduling data, guide maintenance with the help of mixed reality terminals, record retest data and images to form execution data, and use reinforcement learning to update the parameters of the health assessment unit and scheduling optimization unit to form an adaptive loop.

[0126] The execution learning unit is responsible for transforming maintenance scheduling data into actionable instructions on-site, and then feeding back on-site execution feedback to the health assessment unit and scheduling optimization unit, thus closing the adaptive loop of "data generation - risk assessment - scheduling optimization - execution feedback". Its internal process can be divided into four stages: work order issuance, mixed reality guidance, execution data collection, and reinforcement learning update.

[0127] During the work order issuance phase, maintenance scheduling data is first written to the work order management module. Each scheduling record includes equipment identifier, start time slot, duration time slot, work group number, spare parts list, and priority. The execution learning unit sorts work orders according to priority, generating a queue to be dispatched. When a work group's mobile terminal connects to the network, the system pushes the work order list of the work group it is responsible for, along with the world coordinates of the equipment in the 3D digital twin map. This ensures that maintenance personnel know the exact location and spare parts requirements before arriving on site, reducing travel time.

[0128] The mixed reality guidance phase utilizes a head-mounted mixed reality terminal to align the coordinates of the digital twin image to the real scene. The system calculates the rigid transformation matrix from the equipment model to the on-site equipment through planar detection and QR code-assisted positioning; subsequently, it overlays risk hotspot textures and step-by-step disassembly / assembly animations onto the field of view. Maintenance personnel complete disassembly, inspection, lubrication, and component replacement operations according to instructions. After each key step is completed, the terminal automatically captures before-and-after comparison images and calls the edge inference module to read real-time vibration and instantaneous current values, forming a retest sample. All retest data is locally signed and then uploaded to the execution data interface.

[0129] During the data acquisition phase, image hashes, retested sensor vectors, and operation timestamps are aggregated into execution data. The execution learning unit defines three evaluation metrics: failure rate, overload rate, and regulatory penalty rate, denoted as follows: , , After completing all work orders for the shift, the system calculates these three metrics displayed in the scrolling window and constructs a reward function:

[0130] in The weighting constant is calibrated offline using historical data. Indicates the reward value. Indicates the failure rate. Indicates the overload rate. This indicates the penalty rate. The negative sign ensures that the worse the indicator, the lower the reward.

[0131] The reinforcement learning update phase employs a proximate policy optimization algorithm. (State vector) It consists of five dimensions: failure rate, overload rate, penalty rate, average risk reduction, and minimum energy value; action vector. This refers to the adjustment of the time constant for the health assessment unit and the adjustment of the potential weight for the scheduling optimization unit. The policy network and value network enter offline updates after the execution data arrives, performing one round of optimization by sampling 10,000 historical trajectories each time. If the new strategy increases the cumulative discount reward by more than a threshold, it takes effect, and the updated parameters are injected back into the health assessment unit and scheduling optimization unit via remote procedure call, ensuring that the risk assessment and scheduling weights for the next cycle are aligned with the actual situation in real time.

[0132] The above four stages complete a closed loop within a rolling cycle, producing the following effects: First, mixed reality guidance precisely aligns the digital twin map with the real equipment, making maintenance paths more intuitive and reducing the average time per operation by 25%. Second, real-time retest data and images provide post-event correction samples for the health assessment unit, reducing model drift. Third, reinforcement learning automatically adjusts the time constant and potential weights based on execution quality, preventing over-concentration or over-sparse scheduling, and achieving a balance between resource consumption and risk reduction in the long run.

[0133] Example 9: In the air conditioning water pump system of a large commercial complex, the execution learning unit dispatches 100 work orders daily. After overlaying risk hot zones on the mixed reality terminal, the average time for maintenance personnel to locate target equipment decreased from five minutes to two minutes; the failure rate dropped from 2.5% to 1.6% within six months. The reward function weights were calibrated offline for two weeks before going online. Reinforcement learning underwent three rounds of offline training every night at midnight. After two weeks, the average time constant increased by nine percentage points, and the potential weight decreased by five percentage points; the scheduling concentration reached equilibrium under the constraints of team resources, and the overtime rate per shift decreased by 18%. The results show that the execution learning unit can transform the theoretical benefits of scheduling optimization into actual operational benefits, and continuously improve the parameters of the health assessment unit and the scheduling optimization unit through reinforcement learning, achieving a true adaptive maintenance closed loop.

[0134] Preferably, the execution learning unit displays images of equipment risk hot zones and maintenance operation instructions at the maintenance site through a mixed reality terminal. After maintenance is completed, it records retest data and images to form execution data. Based on the execution data, it generates parameter update quantities through reinforcement learning. The parameter update quantities are simultaneously written into the health assessment unit and the scheduling optimization unit to form an adaptive loop of data generation, risk assessment, scheduling optimization and execution feedback.

[0135] The execution learning unit is the closed-loop hub of the periodic maintenance task automatic scheduling system. Its core objective is to transform the maintenance scheduling data output by the scheduling optimization unit into executable instructions for the field, collect retest data after maintenance is completed, and use reinforcement learning mechanisms to update the key parameters of the health assessment unit and the scheduling optimization unit. The workflow of this unit can be divided into three stages: mixed reality presentation, execution data generation, and reinforcement learning update, forming a closed loop of "perception-decision-execution-learning".

[0136] In the mixed reality presentation stage, a head-mounted mixed reality terminal aligns the equipment model in the digital twin with the actual physical entity on site. The alignment process consists of three steps: planar detection, QR code positioning, and world coordinate binding. The system first identifies the site plane in the terminal's camera stream, establishing a local spatial reference. Then, it scans the QR code on the equipment nameplate to obtain the equipment identification and queries the equipment's world coordinates through the digital twin. Finally, it calculates the rigid transformation matrix, overlaying the equipment's 3D model and risk hotspot texture onto the field of view. Risk hotspots are encoded with color gradients representing the product of the risk coefficient and the degradation field gradient, allowing maintenance personnel to easily identify high-risk components. Maintenance operation instructions are fixed next to the model in the form of semi-transparent text patches, including steps such as power off, disassembly, lubrication, reassembly, and retesting. Clicking on these patches reveals a decomposed animation. This method replaces traditional paper or 2D work orders, reducing cognitive burden and operational errors.

[0137] The data generation phase is automatically triggered after maintenance operations are completed. The mixed reality terminal calls the local edge inference framework to read the instantaneous values ​​of the re-measured sensors, such as the effective value of vibration, the average value of current, temperature, and humidity, and performs differential encoding on the images before and after maintenance to generate three comparison images. All re-measured indicators are combined into a vector. The data, along with the image fingerprint, is encapsulated into an execution data object. The object structure includes a device identifier, a maintenance completion timestamp, a retest vector, an image hash, and maintenance time. The execution data is signed on the device side and then reported to the execution learning unit.

[0138] The reinforcement learning update phase uses a continuously scrolling window as the time granularity, summarizing all execution data within the window and calculating the failure rate. Overload rate and penalty rate The failure rate is the percentage of undetected faults within the window; the overload rate is the percentage of overtime hours worked by the shift; and the penalty rate is the ratio of penalties for exceeding regulatory deadlines to the total number of work orders. The execution learning unit is designed with a linear negative reward function.

[0139] in The non-negative weights selected for offline cross-validation are used to balance the contributions of the three metrics to rewards and penalties. Indicates the reward value. Indicates the failure rate. Indicates the overload rate. Indicates the penalty rate. This is the weight constant. State vector. Take the failure rate, overload rate, penalty rate, and average risk reduction of the past window. and minimum potential energy value Action vector Includes adjustment of the time constant of the health assessment unit Adjustment of potential energy weights with scheduling optimization unit The proximal policy optimization algorithm is used to perform optimization on the policy network and value network. Offline updates are performed in cycles; if the cumulative reward increase exceeds a set threshold, the parameter update amount is generated. The health assessment unit and scheduling optimization unit are written through the remote procedure call interface, so that the next cycle assessment and scheduling can adapt to the on-site execution quality at the parameter level.

[0140] This adaptive loop delivers multiple technical benefits. First, the 3D alignment of the mixed reality terminal reduces device positioning time from an average of five minutes to two minutes, shortening the total maintenance time by 25%. Second, the retest vector provides closed-loop correction samples for the health assessment unit, reducing model drift and lowering the prediction mean square error by approximately 8%. Third, after reinforcement learning dynamically adjusts the time constant and weights, scheduling is more risk-reduction-oriented during high-load phases and more load-balanced during low-load phases, resulting in a 30% reduction in failure rate and a 20% reduction in overload over long periods. Finally, image hashes and load metrics in the execution data can be directly used for compliance audits, improving operational transparency.

[0141] Example 10: In an industrial park with 250 cooling tower fans, six months after the execution learning unit went online, the system generated a total of 17,000 execution data points. The policy network underwent three rounds of offline training every night, resulting in a 10% improvement in the average time constant and a 5% reduction in potential weights. Compared to before the deployment, the total downtime due to group failures decreased by 30 hours, and the number of penalties decreased from seven to zero. Maintenance personnel adopted mixed reality guidance at a rate of 95%, and a survey revealed a significant decrease in misassembly incidents. This demonstrates that the present invention can effectively improve on-site execution quality and continuously optimize overall scheduling performance through a reinforcement learning closed loop.

[0142] like Figure 2 As shown, an automatic scheduling method for periodic maintenance tasks is used to execute the aforementioned automatic scheduling system for periodic maintenance tasks. The method includes: S1. Collect vibration signals, current signals, and environmental signals to generate multi-source raw data; perform spatial mapping on the multi-source raw data based on the building information model and geographic information model to construct a three-dimensional digital twin map and obtain synchronized digital twin data; S2. Perform temporal alignment convolution on the digital twin synchronization data within a preset time window to obtain alignment feature vectors; input the alignment feature vectors into an ordinary differential equation recursive network to generate a health index, remaining usable lifetime, and continuous degradation field tensor, and package device risk assessment data based on the health index, remaining usable lifetime, and continuous degradation field tensor. S3. Input the equipment risk assessment data into the symplectic gradient Hamilton-Monte Carlo model for sampling to obtain the initial scheduling solution; sequentially perform Boson sampling, topological Boson evolution and dual game optimization on the initial scheduling solution to obtain maintenance scheduling data; S4. Publish the maintenance scheduling data as a maintenance work order, use a mixed reality terminal to guide maintenance personnel to complete the maintenance and record retest data and images to form execution data; S5. Based on the execution data, reinforcement learning is used to update the parameters of the ordinary differential equation recurrent network and the parameters of the symplectic gradient Hamilton-Monte Carlo model. The updated parameters are then applied to steps S2 and S3 to form an adaptive loop of data generation, risk assessment, scheduling optimization and execution feedback.

[0143] In this invention, S1, Data Twin: Vibration, current, and environmental sensor data are synchronized in time and written into a 3D digital twin map to form synchronized digital twin data with spatial semantics. S2, Risk Assessment: Within a sliding window, temporally aligned convolution is used to eliminate sampling biases between different channels. Then, a recursive network inference using ordinary differential equations is used to generate a health index, remaining usable lifetime, and degradation field tensor, which are then encapsulated as equipment risk assessment data. S3, Global Scheduling: The risk assessment data is input into a symplectic gradient Hamiltonian-Monte Carlo model to obtain a low-potential initial solution. Subsequently, through Bose sampling, topological Bose evolution, and dual game multi-level optimization, maintenance scheduling data that satisfies resource constraints is output. S4, On-site Execution: The scheduling data is published as a maintenance work order. A mixed reality terminal visualizes the risk hotspot on-site and guides operations. After maintenance is completed, retested sensor data and images are automatically uploaded to generate execution data. S5. Strengthening the closed loop: After the execution data is evaluated by the reward function, reinforcement learning is used to adjust the recursive network time constant and Hamilton-Monte Carlo potential weights, update the parameters, and inject the risk assessment and scheduling module to achieve a continuous adaptive loop of data collection, risk calculation, task scheduling, and on-site execution.

[0144] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An automatic scheduling system for periodic maintenance tasks, characterized in that, include: The acquisition twin unit is used to collect vibration signals, current signals and environmental signals to generate multi-source raw data, and construct a three-dimensional digital twin map based on the building information model and geographic information model, and output synchronized digital twin data; The health assessment unit is used to perform temporally aligned convolution and ordinary differential equation recursive network operations on the digital twin synchronization data within a preset time window to generate health index, remaining usable lifetime and continuous degradation field tensor, and encapsulate them into equipment risk assessment data. The scheduling optimization unit is used to input equipment risk assessment data into the symplectic gradient Hamilton-Monte Carlo model and, through Bose sampling, topological Bose evolution and dual game optimization, output maintenance scheduling data. The execution learning unit is used to publish maintenance scheduling data, guide maintenance with the help of mixed reality terminals, record retest data and images to form execution data, and use reinforcement learning to update the parameters of the health assessment unit and scheduling optimization unit to form an adaptive loop.

2. The system according to claim 1, characterized in that, The acquisition twin unit includes a time synchronization module and a data mapping module. The time synchronization module writes a unified time identifier to all sensor data, and the data mapping module partitions and stores the multi-source raw data according to the device identifier and a fixed time window, and automatically updates the 3D digital twin map when a new device identifier is detected.

3. The system according to claim 1, characterized in that, The temporal alignment convolution of the health assessment unit uses a variable convolution kernel length, and the adaptive algorithm adjusts the convolution kernel length in real time according to the sum of the time difference between adjacent vibration signals and the time difference between environmental signals.

4. The system according to claim 3, characterized in that, During the recursive network inference stage of the ordinary differential equation, the health assessment unit adjusts the network time constant in real time according to the change magnitude of the aligned feature vector, and writes the updated hidden state into the cache after inference.

5. The system according to claim 1, characterized in that, The health assessment unit uses an index mapping method to integrate the health index and remaining usable life into a risk coefficient, and uses the spatial distribution value of the risk coefficient to update the continuous degradation field tensor.

6. The system according to claim 5, characterized in that, When the health assessment unit encapsulates the risk assessment data of the packaged equipment, it writes the spatial gradient of the continuous degradation field tensor and the corresponding risk coefficient into the same data structure.

7. The system according to claim 1, characterized in that, The scheduling optimization unit uses momentum vectors and random noise to drive variable updates in the symplectic gradient Hamiltonian-Monte Carlo model to generate an initial scheduling solution.

8. The system according to claim 7, characterized in that, The scheduling optimization unit performs Boson sampling on the initial scheduling solution to generate a set of candidate scheduling solutions. The set of candidate scheduling solutions is input into topological Boson evolution to obtain conflict resolution scheduling solutions. After matching resource constraints through dual game, the maintenance scheduling data is output.

9. The system according to claim 1, characterized in that, The execution learning unit displays images of equipment risk hot zones and maintenance operation instructions at the maintenance site through a mixed reality terminal. After maintenance is completed, it records retest data and images to form execution data. Based on the execution data, it generates parameter update quantities through reinforcement learning. The parameter update quantities are simultaneously written into the health assessment unit and the scheduling optimization unit to form an adaptive loop of data generation, risk assessment, scheduling optimization and execution feedback.

10. An automatic scheduling method for periodic maintenance tasks, used to execute the automatic scheduling system for periodic maintenance tasks according to any one of claims 1-9, characterized in that, The method includes: S1. Collect vibration signals, current signals, and environmental signals to generate multi-source raw data; perform spatial mapping on the multi-source raw data based on the building information model and geographic information model to construct a three-dimensional digital twin map and obtain synchronized digital twin data; S2. Perform temporal alignment convolution on the digital twin synchronization data within a preset time window to obtain alignment feature vectors; input the alignment feature vectors into an ordinary differential equation recursive network to generate a health index, remaining usable lifetime, and continuous degradation field tensor, and package device risk assessment data based on the health index, remaining usable lifetime, and continuous degradation field tensor. S3. Input the equipment risk assessment data into the symplectic gradient Hamilton-Monte Carlo model for sampling to obtain the initial scheduling solution; sequentially perform Boson sampling, topological Boson evolution and dual game optimization on the initial scheduling solution to obtain maintenance scheduling data; S4. Publish the maintenance scheduling data as a maintenance work order, use a mixed reality terminal to guide maintenance personnel to complete the maintenance and record retest data and images to form execution data; S5. Based on the execution data, reinforcement learning is used to update the parameters of the ordinary differential equation recurrent network and the parameters of the symplectic gradient Hamilton-Monte Carlo model. The updated parameters are then applied to steps S2 and S3 to form an adaptive loop of data generation, risk assessment, scheduling optimization and execution feedback.

Images