Machine learning based flue gas treatment facility predictive maintenance system and method

Abstract

The present application relates to the technical field of flue gas treatment, and discloses a flue gas treatment facility predictive maintenance system and method based on machine learning, which comprises the following steps: establishing a device physical topology database, then establishing a time-varying device coupling strength matrix and identifying maintenance vulnerable chains; obtaining historical production data, generating production rhythm feature vectors through multi-scale time-frequency analysis, and outputting maintenance timing Gantt charts in combination with the maintenance vulnerable chains; extracting influence feature vectors of each maintenance task, constructing a production loss prediction model, and predicting the total production loss of the current maintenance plan; the present application deeply integrates the device spatial coupling relationship and the production rhythm, accurately aligns the multi-device maintenance window with the production trough under the global coupling constraint, avoids the induction of chain production stoppage of the high-coupling device group due to out-of-order or parallel maintenance, and converts the "production constraint conflict" into a predictable zero-conflict window configuration in a data-driven manner, thereby reducing the unplanned downtime risk caused by the mismatch of the maintenance time.

Description

Technical Field

[0001] This invention relates to the field of flue gas treatment technology, and more specifically, to a predictive maintenance system and method for flue gas treatment facilities based on machine learning. Background Technology

[0002] In the field of flue gas treatment technology, coal gas purification systems are the core component of industrial flue gas source control, and their stable operation directly determines the flue gas treatment efficiency and emission compliance levels. A coal gas purification system consists of various interconnected devices such as valves, coolers, and pipelines, forming a tightly coupled industrial ecosystem through the transfer of energy and matter between the devices via coal gas and other media. Because coal gas purification requires continuous operation to avoid direct emissions of pollutants, system equipment maintenance must strictly avoid peak production periods and the risk of cascading failures. Therefore, precise adaptation of maintenance windows has become a key technical challenge in ensuring both effective flue gas treatment and production continuity.

[0003] In existing technologies, optimization of flue gas treatment systems largely focuses on parameter monitoring and control strategies. For example, Chinese patent application CN114819238A discloses a method and device for predicting the oxygen content in flue gas from a gas-fired boiler. By calling a preset prediction model, it generates the oxygen content result in flue gas based on operating data, achieving real-time parameter monitoring without sensors, reducing measurement errors and facilitating basic equipment maintenance. However, this solution is limited to the prediction and monitoring of a single parameter and does not involve the analysis of coupling relationships between multiple devices or maintenance sequence planning. Chinese patent application CN120595611A discloses an AI-based intelligent control method, system, and medium for flue gas treatment. It identifies operating conditions through unsupervised clustering and optimizes control parameters using differential evolution algorithms, achieving dynamic switching of strategies under different operating conditions. This optimizes operating costs while ensuring pollutant emissions meet standards. However, this solution focuses on the control and optimization of the flue gas treatment process and does not address the issue of coordinated adaptation between equipment maintenance and production rhythms.

[0004] However, none of the aforementioned existing technologies address the core contradiction in the maintenance of multiple devices in coal gas purification systems: the conflict between production constraints and the dynamic adaptation of maintenance needs. In coal gas purification systems, equipment forms a strong coupling relationship through the coal gas medium, and the production process exhibits significant multi-scale rhythmic characteristics, such as coal gas production fluctuating in 2-3 hour and 8-hour cycles depending on feeding and shift changes. Existing technologies neither quantify the coupling strength between devices nor identify suitable maintenance windows based on production rhythms, leading to a dilemma in maintenance decisions: fixed-cycle shutdowns for maintenance easily overlap with peak coal gas production periods, causing unplanned production losses; while temporary maintenance based on single load parameters, without identifying highly coupled equipment chains, can easily trigger cascading failures due to out-of-sequence maintenance, not only increasing the risk of production interruptions but also potentially causing substandard flue gas treatment due to equipment malfunctions, thus violating the core objectives of flue gas treatment. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, this invention provides a predictive maintenance system and method for flue gas treatment facilities based on machine learning. By constructing a physical topology database of equipment to identify maintenance-prone chains, and combining the rhythmic characteristics of historical production data to generate a maintenance time-series Gantt chart, the system can simultaneously predict production losses in the maintenance plan. This enables precise alignment of maintenance windows for multiple devices with production troughs, avoids chain failures caused by out-of-sequence maintenance of highly coupled equipment, effectively resolves production constraint conflicts, and significantly improves the scientific nature of gas purification system maintenance and production continuity.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Predictive maintenance methods for flue gas treatment facilities based on machine learning include:

[0008] Collect the equipment list of the gas purification system and the three-dimensional spatial coordinates of each device in the equipment list to construct a physical topology database of the equipment; construct a time-varying equipment coupling strength matrix based on the physical topology database of the equipment; identify and maintain easily disturbed chains according to the time-varying equipment coupling strength matrix;

[0009] Historical production data of the gas purification system is acquired, and multi-scale time-frequency analysis is performed on the historical production data to generate a production rhythm feature vector; based on the production rhythm feature vector and the maintenance perturbation chain, a maintenance time-series Gantt chart is generated.

[0010] The maintenance impact feature vector of each maintenance task is extracted from the maintenance timeline Gantt chart, and a production loss prediction model is constructed. Based on the maintenance impact feature vectors of all maintenance tasks and the production loss prediction model, the total production loss of the current maintenance plan is predicted.

[0011] The method for constructing the device physical topology database includes:

[0012] Establish a device connection matrix ECM. The device connection matrix ECM is a symmetric square matrix. The element ECMij in the i-th row and j-th column of ECM is set to 1 when there is a direct connection between device i and device j, and otherwise set to 0.

[0013] Each device is assigned a unique identifier, and the device's physical topology database is composed of the device's unique identifier, three-dimensional spatial coordinates, and device connection matrix.

[0014] The method for constructing the coupling strength matrix of the time-varying device includes:

[0015] Based on the device physical topology database, virtual tracer particle injection and trajectory tracking are performed to generate a medium flow trajectory dataset;

[0016] The media transfer time and temperature conduction influence coefficient between devices are calculated based on the media flow trajectory dataset; a time-varying device coupling strength matrix is ​​constructed based on the media transfer time and temperature conduction influence coefficient between devices.

[0017] The medium flow trajectory dataset includes at least the time when particles enter and leave the device, and the medium temperature when particles enter and leave the device.

[0018] The medium transfer time between the devices is calculated based on the time it takes for particles to enter and leave the devices, and the temperature conduction influence coefficient between the devices is calculated based on the medium temperature when particles enter and leave the devices.

[0019] The method for constructing a time-varying device coupling strength matrix based on the medium transfer time and temperature conduction influence coefficient between devices includes:

[0020] Based on the device connection matrix ECM, the shared pipe length between devices is calculated, and the comprehensive coupling strength between devices is calculated based on the medium transfer time, temperature conduction influence coefficient and shared pipe length.

[0021] The time-varying device coupling strength matrix uses the device's unique identifier as the row and column index and the overall coupling strength as the element value.

[0022] The method for identifying and maintaining easily disturbed chains based on time-varying device coupling strength matrices includes:

[0023] A depth-first search algorithm is used to identify candidate maintenance-prone chains based on the device connectivity matrix (ECM).

[0024] After the candidate maintenance-prone chains are determined, the chain length and chain strength of the candidate maintenance-prone chains are calculated.

[0025] If the length of a candidate maintenance-prone chain is greater than a length threshold and the chain strength is greater than a strength threshold, then the candidate maintenance-prone chain is confirmed as a maintenance-prone chain.

[0026] The method for identifying candidate maintenance-prone chains includes:

[0027] A depth-first search algorithm is employed, starting with each device as the initial node. Based on the device connectivity matrix (ECM), all direct adjacent nodes of the initial node are generated, forming the first layer of nodes in the search. Using these first-layer nodes as new starting points, the direct adjacent nodes are recursively generated, forming all reachable connection paths between devices. The algorithm then determines whether the overall coupling strength of all adjacent device pairs in each reachable connection path is greater than the strong coupling threshold α. strong If so, the reachable connection path will be used as a candidate maintenance perturbation chain.

[0028] The historical production data of the gas purification system includes at least the amount of gas produced.

[0029] The method for generating the production rhythm feature vector includes: constructing a time series of gas production, performing a fast Fourier transform on the time series of gas production to identify periodic components; recording the amplitude and phase of each identified periodic component, and assigning a unique periodic component identifier to each periodic component; the periodic component identifier, amplitude, and phase of each periodic component constitute the rhythm feature of the corresponding periodic component; and integrating the rhythm features of all periodic components to form the production rhythm feature vector.

[0030] The method for generating the maintenance time-series Gantt chart includes:

[0031] Calculate the production load index, and determine the effective idle window based on the production rhythm feature vector and the production load index;

[0032] Each maintenance disturbance chain is assigned a disturbance level, a maintenance task priority queue is constructed based on the maintenance disturbance chain level, and the best fit and matching of maintenance tasks with available free windows is performed.

[0033] By integrating the best fit and matching results and introducing multidimensional constraints, a visual maintenance time-series Gantt chart is generated.

[0034] The method for determining the effective free window includes:

[0035] Potential idle windows are determined based on the production load index. After determining the potential idle windows, a qualitative verification is performed using the production rhythm feature vector to confirm whether the potential idle windows are true low-temperature windows. The true low-temperature windows are then subjected to a persistence judgment. If the persistence judgment is passed, the true low-temperature window is determined to be a valid idle window.

[0036] A machine learning-based predictive maintenance system for flue gas treatment facilities, used to implement the aforementioned machine learning-based predictive maintenance method for flue gas treatment facilities, the system comprising:

[0037] Coupled Association Identification Module: Used to collect the equipment list of the gas purification system and the three-dimensional spatial coordinates of each equipment in the equipment list, and construct an equipment physical topology database; based on the equipment physical topology database, construct a time-varying equipment coupling strength matrix; and identify and maintain easily disturbed chains according to the time-varying equipment coupling strength matrix.

[0038] Maintenance time-series Gantt chart generation module: used to acquire historical production data of the gas purification system, perform multi-scale time-frequency analysis on the historical production data, generate production rhythm feature vectors; and generate maintenance time-series Gantt charts based on the production rhythm feature vectors and maintenance perturbation chains.

[0039] Loss prediction module: used to extract the maintenance impact feature vector of each maintenance task from the maintenance timeline Gantt chart, build a production loss prediction model, and predict the total production loss of the current maintenance plan based on the maintenance impact feature vectors of all maintenance tasks and the production loss prediction model.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] This invention transforms the traditionally invisible inter-equipment medium-thermal coupling into quantifiable dynamic strength by constructing an equipment physical topology database and a time-varying coupling strength matrix. It further identifies easily disruptive chains that could trigger cascading failures once maintenance is initiated. Based on this, it extracts production rhythm feature vectors using multi-scale time-frequency features of historical production data. This ensures that maintenance window selection no longer relies on static load thresholds but is strictly synchronized with the daily-shift-feeding multi-cycle low-valley phases of gas production. This embeds the "coupling-rhythm" dual constraint into the maintenance sequence Gantt chart during the generation stage. Subsequently, using the maintenance impact feature vector as a link, a production loss prediction model performs a unified loss assessment of planned multi-equipment maintenance tasks, ensuring that any window adjustment iterates within the foreseeable range of total loss. Thus, the entire solution transforms the "production constraint conflict of multi-equipment maintenance windows" into a zero-conflict state that can be explicitly calculated and eliminated during the planning stage. This achieves closed-loop control of coupling relationships, production rhythms, and maintenance losses, significantly improving the maintenance executability and operational stability of complex flue gas treatment systems in continuous production scenarios. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart illustrating the principle of a predictive maintenance method for flue gas treatment facilities based on machine learning, provided in an embodiment of the present invention.

[0044] Figure 2 A flowchart illustrating the principle of identifying and maintaining easily disturbed chains, provided for an embodiment of the present invention;

[0045] Figure 3 A flowchart illustrating the principle of determining a valid free window is provided in this embodiment of the invention.

[0046] Figure 4 This is a functional block diagram of a predictive maintenance system for flue gas treatment facilities based on machine learning, provided as an embodiment of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1

[0049] Please see Figure 1 As shown, this embodiment provides a predictive maintenance method for flue gas treatment facilities based on machine learning, including:

[0050] Step S10: Collect the equipment list of the gas purification system and the three-dimensional spatial coordinates of each device in the equipment list, and construct a physical topology database of the equipment; construct a time-varying equipment coupling strength matrix based on the physical topology database of the equipment; identify and maintain easily disturbed chains according to the time-varying equipment coupling strength matrix;

[0051] Step S10 focuses on the construction of equipment coupling relationships and identification of vulnerable chains in the gas purification system. Through physical topology modeling, media flow tracking, coupling strength quantification, vulnerable chain mining and visualization, it solves the problems of isolated equipment cognition, implicit coupling relationships and difficulty in predicting chain risks in traditional maintenance, laying the foundation for subsequent maintenance sequence arrangement and loss control.

[0052] Further, step S10 includes:

[0053] Step S11: Collect the equipment list of the gas purification system and the three-dimensional spatial coordinates of each equipment in the equipment list, establish the equipment connection matrix and construct the equipment physical topology database;

[0054] The equipment list in the gas purification system refers to a list of all components from the outlet pipe of the sealed furnace to the outlet pipe of the generator set, including valves, coolers, and pipes. The equipment list is collected by directly extracting basic information through the intelligent monitoring system for waste gas treatment to ensure data integrity. Each piece of equipment is assigned a unique identifier IDi, and its three-dimensional spatial coordinates (Xi, Yi, Zi) are recorded. These coordinates represent the position of each piece of equipment in the factory coordinate system, where Xi represents the distance coordinate of equipment i along the east-west direction, Yi represents the distance coordinate of equipment i along the north-south direction, and Zi represents the distance coordinate of equipment i along the vertical elevation direction. These coordinates are obtained through laser scanners or CAD model measurements to reflect the actual physical layout. The equipment connection matrix ECM is a symmetric square matrix. Its element ECMij is set to 1 when there is a direct connection between equipment i and equipment j, otherwise it is set to 0. Direct connections can be direct pipe connections or medium channels. Pipe connections rely on independent pipe components as intermediaries, while medium channels are flow paths formed by the equipment's own structure or direct docking between equipment, requiring no additional pipe intermediaries; these are "pipeless direct connections" between equipment. The Device Connection Matrix (ECM) is constructed by traversing the device list and verifying physical interface matches. This construction process starts from the device list, comparing coordinates and interface specifications pairwise to mark connections, and then reflecting the connection paths between devices through the relationships between matrix elements. This matrix form is used because it can efficiently represent graph structure relationships, facilitates subsequent path lookups, avoids the inefficiency of traversing list-based storage, and allows for rapid calculation of path existence through matrix multiplication when identifying indirect connections. The Device Physical Topology Database contains unique identifiers, three-dimensional spatial coordinates, and device connection matrices for each device. The Device Physical Topology Database transforms isolated device information into a networked topology, facilitating the capture of implicit interactions.

[0055] In existing technologies, equipment is treated as an independent entity, ignoring media conduction coupling, which makes it impossible to predict cross-equipment impacts during maintenance. Establishing an equipment connection matrix and integrating 3D spatial coordinates to construct an equipment physical topology database quantifies spatial relationships to reveal potential risk paths. For example, maintenance of an upstream valve may affect the thermal stress of a downstream cooler, solving the problem of isolated equipment perception. The equipment physical topology database facilitates the transition from overall solution to coupling analysis. Without this step, the subsequent construction of a time-varying equipment coupling strength matrix will lack basic connection data, causing coupling strength calculations to deviate from physical reality, failing to align with production rhythms, and particle trajectory tracking cannot define flow boundaries without a connection matrix. This step connects static equipment information with dynamic media simulation, laying the foundation for coupling quantification and making the identification of maintenance-prone chains possible. After the equipment physical topology database is constructed, it directly supports particle injection in step S12, ensuring that trajectory tracking is limited to actual connection paths. When coordinating with subsequent step S20, the 3D spatial coordinates in the equipment physical topology database assist in matching idle windows, avoiding maintenance conflicts and integrating physical layout into time-series planning.

[0056] Step S12: Based on the device physical topology database, perform virtual tracer particle injection and trajectory tracking to generate a medium flow trajectory dataset;

[0057] Virtual tracer particles refer to massless points that simulate gas molecules, used to track medium flow without interfering with the actual system. During the injection process, particle batches are generated at the outlet of the sealed furnace every preset time interval τ. τ is determined by system response time and load balancing calculations; for example, τ is set to 10 seconds to capture transient changes. Trajectory tracking simulates the movement of particles along paths defined by the device connection matrix. Initial attributes such as temperature, pressure, and velocity are initialized from real-time sensor data to match the current operating conditions. When a particle enters device i, the entry time Tin,i is recorded and its attributes are updated; when it leaves, the departure time Tout,i is recorded. The residence time Di of the particle in device i is the difference between Tout,i and Tin,i.

[0058] Particle tracking can simulate complex flows rather than simplified pipe models, generating detailed media flow trajectory datasets. These datasets include unique particle identifiers, unique device identifiers, the times when the particle enters and leaves the device, and the media temperatures at the time of entry and exit. This data is presented in tabular form. For example, for particle PID001, the media flow trajectory dataset is tabulated as PID001, IDi, Tin,i, Tout,i, θin,i, and θout,i. Here, θin,i represents the media temperature at the inlet of device i when the particle enters i, and θout,i represents the media temperature at the outlet of device i when the particle leaves i. Thermal changes are captured by synchronously collecting data at the inlet and outlet sensing points of device i. The temperatures here are not the temperatures of the virtual tracer particles themselves (virtual particles are massless points with no physical properties), but rather the real-time temperatures of the medium (gas) being tracked by the particles at the inlet and outlet of device i. Existing technologies neglect flow complexity, resulting in static coupling assessments. Virtual tracer particle injection and trajectory tracking can dynamically capture conduction paths, revealing short-circuit or blockage risks, improving prediction accuracy, solving the problem of implicit coupling quantification, and advancing the evolution of solutions towards time-varying analysis. Without this step, the medium transmission time in step S13 cannot be calculated, leading to a static coupling matrix that cannot be coordinated with loss prediction. This step bridges physical topology and dynamic parameters, providing a computational foundation and enabling the identification of susceptible chains based on realistic simulation. By limiting particle paths through the device's physical topology database, invalid tracking is avoided, data accuracy is enhanced, and precise mapping of coupling relationships is achieved.

[0059] Step S13: Calculate the medium transfer time and temperature conduction influence coefficient between devices based on the medium flow trajectory dataset; construct a time-varying device coupling strength matrix based on the medium transfer time and temperature conduction influence coefficient between devices.

[0060] The entry and exit times of particles into and out of the devices, as well as the medium temperature at those times, are extracted from the medium flow trajectory dataset. The medium transfer time (MTTij) between devices i and j is calculated based on these entry and exit times: MTTij is defined as the average value of the particle's exit time (Tout,i) from the exit time (Tout,i) of device i to the entry time (Tin,j) of device j. This is obtained by calculating the difference between Tin,j and Tout,i for all particles passing through the path from i to j and taking the average, where Tin,j is the moment the particle enters device j. The averaging method smooths out noise effects and quantifies delays to classify coupling strength. Based on MTTij and a preset time threshold T... thre Based on the relationship, the coupling type between device i and device j is divided into three categories: strong, medium, and weak. Strong coupling indicates fast medium transmission between devices and rapid diffusion of maintenance impacts; weak coupling indicates slow transmission and gradual diffusion of impacts. This classification provides a basis for subsequent identification of susceptible chains. If MTTij < T thre Then the coupling type between device i and device j is determined to be strong coupling, if T thre ≤MTTij≤3T thre If MTTij > 3T, then it is considered moderate coupling. thre If , then it is a weak coupling. thre The setting is based on the system response time distribution. By statistically analyzing the average response time of media transmission under historical operating conditions, the 80th percentile is taken as T. thre For example, T thre It lasts for 30 seconds.

[0061] The temperature conduction influence coefficient is calculated based on the medium temperature when particles enter and leave the equipment: θout,i and θin,i are extracted from the medium flow trajectory dataset for device i, and the medium temperature change ΔTi = θout,i - θin,i for device i is calculated. ΔTi reflects the thermal effect of device i on the flowing medium. Simultaneously, θout,j and θin,j are extracted from device j, and the medium temperature change ΔTj = θout,j - θin,j for device j is calculated. θin,j is the medium temperature at the inlet of device j when particles enter; θout,j is the medium temperature at the outlet of device j when particles leave. ΔTj reflects the thermal response of the medium to device j after being acted upon by device i. The formula for calculating the temperature conduction influence coefficient βij between device i and device j is βij = ΔTj / ΔTi. The logic of this formula is that when device i exerts a thermal influence on device j through medium conduction, the medium temperature change ΔTj of device j is positively correlated with the medium temperature change ΔTi of device i; the larger the ratio, the more significant the thermal influence. If ΔTi=0, meaning that device i has no thermal effect, then βij takes the value of 0, indicating that device i has no direct thermal conduction effect on j.

[0062] Based on the Equipment Connection Matrix (ECM), the shared pipe segment length Lij between equipment i and equipment j is calculated. Specifically, based on the direct connection relationships between equipment recorded in the ECM, the sets of pipe-type equipment directly connected to equipment i and equipment j are extracted. The intersection of these two sets yields the pipe segment directly connected to both equipment i and j. The length of each pipe segment is calculated using its three-dimensional spatial coordinates within the intersection. The sum of the lengths of all directly connected pipe segments yields the shared pipe segment length Lij. The introduction of Lij stems from the influence of the tightness of physical connections on coupling. A longer shared pipe segment results in greater consistency in the transmission of the medium under environmental interference, increasing the probability of state changes between equipment i and j being transmitted to each other through the shared pipe segment. For example, a longer shared pipe segment makes it easier for pipe vibrations in equipment i to be transmitted to equipment j, exacerbating mechanical wear in equipment j.

[0063] The comprehensive coupling strength between devices is calculated based on the medium transfer time, temperature conduction influence coefficient, and shared pipe length. The comprehensive coupling strength Cij between device i and device j needs to be normalized by the reciprocal of MTTij, the temperature conduction influence coefficient βij, and the shared pipe length Lij, and then weighted and summed, i.e., Cij=w1×N(1 / MTTij)+w2×N(βij)+w3×N(Lij). Where N(∙) represents the linear normalization function, mapping each parameter to the [0,1] interval to avoid evaluation bias caused by dimensional differences; w1 is the weighting coefficient of medium transfer time, w2 is the weighting coefficient of temperature conduction influence coefficient, and w3 is the weighting coefficient of shared pipe length, the sum of the three is 1. The weighting is determined by the degree of influence of each factor on coupling through the analytic hierarchy process. The influence of heat conduction is the most critical to the safety of equipment operation, so w2 takes the highest value. For example, w1=0.3, w2=0.4, w3=0.3. The introduction of 1 / MTTij makes the contribution of this index greater the shorter the transfer time, and it is positively correlated with the coupling strength; the normalization process ensures that the parameters of the three dimensions are comparable, and the weighted summation realizes the fusion of multi-dimensional information. The construction of the time-varying device coupling strength matrix DCSM uses the unique identifier of the device as the row and column index and the comprehensive coupling strength as the element value. The calculated Cij is filled into the corresponding position. Since MTTij and βij change dynamically with the working conditions, DCSM is updated once every preset time interval (consistent with the particle injection interval τ) to realize the real-time quantification of the coupling relationship.

[0064] Step S13 addresses the shortcomings of static coupling relationships, inaccurate single-dimensional evaluation, and the inability to update coupling strength in real time. By integrating dynamic transmission time, thermal conductivity, and physical connection length, it achieves multi-dimensional dynamic quantification of coupling relationships. Compared to existing technologies that rely solely on static connections for evaluation, it can more accurately reflect the actual degree of influence between devices. Step S13 can not only identify the coupling relationships of directly connected devices but also capture the implicit influence of indirectly connected devices through the calculation of Cij. For example, devices i and j may not be directly connected but form a strong coupling with intermediate devices through shared pipe sections. The calculation of Cij will reflect this indirect association, avoiding the omission of indirect couplings by traditional methods.

[0065] Step S14: Based on the time-varying equipment coupling strength matrix and equipment connection matrix, identify maintenance disturbance chains with cascading failure risk, and assign a disturbance level to each maintenance disturbance chain;

[0066] Please see Figure 2 As shown, a depth-first search algorithm is used, based on the device connection matrix (ECM), to identify candidate maintenance-prone chains. The depth-first search algorithm is chosen because of its ability to efficiently explore paths, making it suitable for finding continuous, strongly coupled device chains in a device topology network. Compared to breadth-first search, it can locate risky chains on long paths more quickly. The search process is as follows: Starting with each device as the initial node, all direct adjacent nodes of the initial node (devices corresponding to elements of 1 in the ECM) are generated based on the device connection matrix (ECM), serving as the first layer of nodes in the search. Using the first layer nodes as new starting points, direct adjacent nodes are recursively generated to form all reachable connection paths between devices. The termination condition is when the path length reaches the system's maximum device chain length L_max. L_max is determined by statistically analyzing the number of devices in the longest historical fault chain, typically 20 devices. Finally, it is determined whether the overall coupling strength of all adjacent device pairs in each reachable connection path is greater than the strong coupling threshold α. strong If so, then the reachable connection path is selected as a candidate maintenance perturbation chain; where α strong The base strength threshold α is equal to 1.2 times. α is set based on the minimum comprehensive coupling strength value that triggered cascading failures in historical maintenance cases. This is achieved by statistically analyzing the comprehensive coupling strength of equipment pairs that experienced secondary failures after maintenance over the past five years, and taking the minimum value as α. strong It is equal to 1.2 times the base strength threshold α to ensure that the selected candidate chains have extremely high coupling strength.

[0067] For example, setting α strong =0.8, devices A→B→C→D form a reachable connection path, where the combined coupling strength between adjacent devices A and B is 0.9 (>α). strong The combined coupling strength of adjacent devices to B and C is 0.7 (< α). strongThe combined coupling strength of adjacent devices to C and D is 0.9 (>α). strong Since the combined coupling strength between B and C is not greater than the strong coupling threshold α, strong The impact of maintaining A can only be transmitted to B, and cannot be transmitted to C and D through the weak points of BC. This path does not have the characteristic of "affecting the whole body by affecting one part", so it cannot be used as a candidate maintenance-prone chain. Only when the combined coupling strength of AB, BC, and CD is greater than 0.8 can the reachable connection path A→B→C→D satisfy the basic condition for the transmission of chain risks.

[0068] After the candidate maintenance-prone chain is determined, its chain length Lchain and chain strength Schain are calculated. Lchain represents the number of devices included in the candidate maintenance-prone chain, and Schain is the geometric mean of the overall coupling strength of all adjacent devices on the candidate maintenance-prone chain. The geometric mean is used instead of the arithmetic mean because it is more susceptible to extreme values, highlights the influence of strongly coupled nodes in the chain, and avoids misjudgments caused by the arithmetic mean being lowered by weakly coupled nodes. If the Lchain of the candidate maintenance-prone chain is greater than the length threshold Lmin and the Schain is greater than the strength threshold Smin, then the candidate maintenance-prone chain is confirmed as a maintenance-prone chain (MSC). Lmin is determined by statistically analyzing the shortest number of devices in historical fault chains, typically taking the minimum length of common fault chains in the system. Smin is the average Schain value of historical strongly coupled chains. If both conditions are met, it is confirmed as a maintenance-prone chain (MSC). The chain length Lchain being greater than Lmin ensures that the number of devices covered by the chain reaches a certain scale. If the chain length is too short, such as only 2-3 devices, even with high coupling strength, the cascading impact of its failure will be limited and will not constitute a systemic risk. The chain strength Schain is greater than Smin to ensure that the coupling relationship between devices on the chain reaches a "strong correlation" level. Schain, as the geometric mean of the overall coupling strength of all adjacent devices on the chain, directly reflects the overall tightness of the chain's coupling. Only when both conditions are met can it be said that the candidate chain is both "long enough" (wide impact range) and "strong enough" (fast risk propagation), possessing the potential to trigger systemic cascading failures, and is therefore identified as a maintenance-sensitive chain (MSC) that requires key attention.

[0069] The method for assigning a vulnerability level to each maintenance vulnerability chain includes: classifying equipment in the gas purification system into Level 1 critical equipment, Level 2 critical equipment, and Level 3 critical equipment. If a maintenance vulnerability chain includes Level 1 critical equipment, it is marked as Level 1 Maintenance Vulnerability Chain MSC1; if it does not include Level 1 critical equipment but includes Level 2 critical equipment, it is marked as Level 2 Maintenance Vulnerability Chain MSC2; if it does not include either Level 1 or Level 2 critical equipment, it is marked as Level 3 Maintenance Vulnerability Chain MSC3. Level 1 critical equipment refers to equipment that plays a decisive role in the core function of the gas purification system, whose failure directly leads to system shutdown, downstream generator interruption, or significant safety risks. This includes gas coolers, the main valve at the outlet of the sealed furnace, and the gas pretreatment equipment at the generator inlet. The gas cooler directly determines the media processing efficiency and the safety of downstream equipment; the main valve at the outlet of the sealed furnace controls the flow of the media source; and the gas pretreatment equipment at the generator inlet connects to the core production chain. Level 2 critical equipment refers to equipment that affects system operating efficiency but does not immediately cause shutdown; its failure only results in localized functional limitations. This includes sectional valve assemblies, core gas temperature / pressure sensors, and medium-sized filtration devices. Sectional valve assemblies regulate local media flow, core gas temperature / pressure sensors ensure monitoring accuracy, and medium-sized filtration devices can affect media purity. Level 3 critical equipment refers to equipment that performs auxiliary functions; its failure has a minimal impact on the overall system operation and is easily and quickly repairable. This includes end-of-line exhaust valves, ordinary connecting pipes, and small monitoring instruments in non-core areas. Failures of critical equipment in MSC1 are likely to cause system shutdown, therefore the joint maintenance assessment window is the shortest, for example, set at 24 hours. The impact of critical equipment in MSC2 spreads more slowly, so the joint maintenance assessment window is set at 72 hours. The impact of critical equipment in MSC3 is minimal, so the joint maintenance assessment window is set at one week. This time division matches the impact transmission speed reflected by the media transmission time in S13; a shorter assessment window for strong coupling chains ensures timely intervention.

[0070] Step S14 achieves accurate identification of risk chains through algorithmic search and quantitative indicator screening. It can not only identify risk chains that have already failed, but also predict potential risk chains based on changes in the Schain value. For example, if the Schain value of a chain continuously increases within a week, it indicates that its coupling strength is constantly increasing and it may develop into a high-risk chain. This chain can be included in the maintenance assessment in advance to avoid failure. Step S14 solves the problem of maintenance priority classification, providing a hierarchical basis for the construction of the maintenance task priority queue in S20. Without this step, S20 would be unable to distinguish the risk level of maintenance tasks, potentially prioritizing the maintenance of low-risk equipment, leading to a chain reaction of losses caused by the failure of high-risk chains.

[0071] Step S15: Map the device connection matrix, the time-varying device coupling strength matrix, and the maintenance disturbance chain onto a two-dimensional plane to generate a visualized device maintenance coupling map.

[0072] The core of step S15 is to solve the problems of difficult-to-understand complex coupling relationships and inefficient risk information transmission in traditional maintenance by using projection transformation, layout algorithm and multi-dimensional visual encoding, so as to provide concrete support for maintenance decision-making.

[0073] An orthogonal projection transformation method is used to map the three-dimensional spatial coordinates of equipment onto the XOY two-dimensional plane. This method can completely preserve the planar relative positional relationships between equipment, avoid visual confusion caused by the superposition of three-dimensional maps, and ensure that maintenance personnel can understand equipment relationships based on the familiar factory layout. A force-oriented layout algorithm is adopted to simulate the interaction of springs in a physical system: equipment is abstracted as nodes, and the connections between equipment are abstracted as springs. The higher the overall coupling strength, the greater the attractive force coefficient of the spring, the stronger the attraction between nodes, and the closer they are; the lower the overall coupling strength, the greater the repulsive force coefficient of the spring, the stronger the repulsion between nodes, and the farther they are. This layout method breaks through the limitations of fixed coordinate layout, and can automatically aggregate strongly coupled equipment clusters and disperse weakly coupled equipment. This allows maintenance personnel to quickly locate highly correlated equipment groups through visual clustering. Compared with traditional list-style data, the spatial distribution characteristics of equipment coupling are easier to identify.

[0074] The node size is determined by the maintenance complexity, which is calculated by integrating the historical average maintenance time T, the number of required spare parts types N*, and the operation difficulty coefficient D. The calculation formula is C=c1×N(T)+c2×N(N*)+c3×N(D). Here, T is the arithmetic mean of the maintenance time of the equipment over the past M times (M≥5), reflecting the basic maintenance time; N* is the total number of spare parts types required for routine maintenance, reflecting the complexity of material preparation; and D is the average engineer rating on a 1-10 scale, determined based on indicators such as operational qualification requirements, special tool needs, and space constraints, reflecting the difficulty of operation. c1, c2, and c3 are the weighting coefficients of the historical average maintenance time, the number of required spare parts types, and the operation difficulty coefficient, respectively. c1, c2, and c3 are determined using the analytic hierarchy process (AHP). For example, c1=0.4, c2=0.3, and c3=0.3, highlighting the core impact of historical maintenance time on complexity. The node size is set linearly according to the value of maintenance complexity C. For example, C=1 corresponds to a diameter of 20pt, and C=0 corresponds to 8pt, so that difficult-to-maintain equipment can be quickly located due to the larger node size.

[0075] Node colors are determined based on the equipment health score H, which is calculated using real-time sensor data, including real-time temperature, real-time pressure, and real-time vibration. The formula for calculating the equipment health score H is H = 1 - (wt × Rt + wp × Rp + wv × Rv). Where Rt is the temperature deviation rate, equal to |T... 实 -T 标均 | / (T 标高 -T 标低 ), where T 实 For real-time temperature, T 标低 T 标高 T represents the lower and upper limits of the equipment's standard operating temperature range. 标均 Rt represents the midpoint of the equipment's standard operating temperature range; Rp is the pressure deviation rate, calculated using the same logic as Rt; Rv is the vibration deviation rate, calculated using the same logic as Rt. wt, wp, and wv are the weighting coefficients for Rt, Rp, and Rv, respectively (e.g., for a cooler, wt=0.5, wp=0.3, wv=0.2). These values ​​are adjusted based on the equipment type and the susceptibility of the parameters to disturbance. For example, if the equipment is a cooler, set wt=0.5, wp=0.3, and wv=0.2. When H≥H 高 When H indicates that the device is healthy, the node color is green; when H 低 <H<H 高 When H ≤ H, it indicates that the device is in a sub-healthy state, and the node color is yellow; 低 When this time, it indicates that the device needs maintenance, and the node color is red. H 高 H is the minimum historical H value for fault-free equipment (e.g., 0.8). 低 This represents the maximum H value of the equipment in the 72 hours prior to the historical fault (e.g., 0.5). The line thickness is linearly correlated with the overall coupling strength; the greater the overall coupling strength, the thicker the line, visually representing the difference in coupling strength. Maintenance-prone chains are highlighted with different colors: MSC1 is highlighted in red, MSC2 in orange, and MSC3 in yellow.

[0076] Step S15 addresses the shortcomings of traditional maintenance methods, such as fragmented coupled information and reliance on experience for risk identification. Through multi-dimensional visual coding, it makes implicit coupling relationships and risk levels explicit. Maintenance personnel can quickly locate high-risk equipment using a combination of "yellow nodes (sub-healthy) + red highlighted lines (MSC1)". For example, if a sub-healthy valve is located in MSC1, it can be immediately included in the emergency maintenance assessment, preventing a chain reaction caused by the fault. Step S15 also solves the problem of visualizing maintenance decision-making information, providing a visual verification basis for matching maintenance tasks in S20. Without this step, the time-varying equipment coupling strength matrix in S13 and the maintenance disturbance chain information in S14 are difficult to translate into practical decisions, requiring maintenance personnel to spend a significant amount of time interpreting the data, resulting in low decision-making efficiency. The classification of vulnerable chains directly determines the highlighted colors and warning priorities in the maintenance graph. The red highlight of MSC1 matches the 24-hour joint maintenance assessment window, ensuring consistency between visual warnings and timely decision-making requirements. The synergy with S20 is reflected in the fact that the device clusters and node statuses in the visualized equipment maintenance coupling graph provide an intuitive reference for maintenance task combinations and window matching. Maintenance personnel can use the graph to identify strongly coupled devices not recognized by the algorithm, manually adjust maintenance plans, and avoid conflicts caused by algorithm limitations. This synergy shortens maintenance decision-making time and improves decision accuracy.

[0077] Step S20: Obtain historical production data of the gas purification system, perform multi-scale time-frequency analysis on the historical production data, and generate a production rhythm feature vector; based on the production rhythm feature vector and the maintenance perturbation chain, generate a maintenance time series Gantt chart.

[0078] Step S20 solves the problems of rhythm fragmentation, blind window adaptation, and frequent resource conflicts in traditional maintenance scheduling by extracting multi-scale rhythms, identifying dynamic windows, matching priorities, and verifying constraints, providing core technical support for the precise coordination of maintenance plans and production plans.

[0079] Further, step S20 includes:

[0080] Step S21: Obtain historical production data of the gas purification system, perform multi-scale time-frequency analysis on the historical production data, and generate production rhythm feature vectors.

[0081] Historical production data refers to the set of hourly operating parameters extracted from the intelligent monitoring system for waste gas treatment within the most recent M1 month, mainly including coal gas production G. h Where h represents the h-th hour, G hThis reflects the processing load of the gas purification system. Multi-scale time-frequency analysis is a method for decomposing and extracting patterns at different time scales in production data. Its core advantage lies in breaking through the limitations of single-time-dimensional analysis, and can simultaneously capture production fluctuation characteristics at different cycles such as hourly, daily, and weekly. This is because high-carbon ferrochrome production is affected by multiple factors such as material change, shift change, and equipment operation cycle, and the rhythm exhibits obvious multi-scale superposition characteristics. Traditional single-scale analysis will miss key cycle information, leading to conflicts between maintenance windows and hidden production peaks.

[0082] The process of generating rhythmic feature vectors by performing multi-scale time-frequency analysis on historical production data is as follows: The amount of coal gas produced, G... h The time series of coal gas production {G1,G2,...,G} constitutes the total amount of coal gas produced. n}, where G n This represents the amount of gas produced in the nth hour, and the time series of gas production {G1, G2, ..., G...} represents the amount of gas produced in the nth hour. n Performing a Fast Fourier Transform (FFT) converts the time-domain signal into a frequency-domain signal, quantifying the energy distribution of different frequency components and thus identifying periodic components. Higher energy frequencies correspond to more significant repetitive fluctuations in the time domain. Identifying the period essentially involves calculating the corresponding time intervals (period = 1 / frequency) from these significant frequency components, thereby determining the recurring temporal patterns in the production process. For example, after performing a FFT on the gas production time series of a 25200 KVA high-carbon ferrochrome furnace, the energy peaks in the frequency domain correspond to frequencies of 1 / 24 hours. -1 1 / 8 hour -1 1 / (2-3) hour -1 After conversion, the 24-hour cycle, 8-hour cycle, and 2-3 hour cycle can be identified. These cycles correspond to the repetitive rhythms of different dimensions in production.

[0083] The 24-hour cycle corresponds to the daily rhythm, matching the factory's work schedule and the day-night operating characteristics of equipment, such as the nighttime load adjustment of some equipment. The 8-hour cycle corresponds to the shift change rhythm, as factories generally adopt a three-shift production system, and personnel operation switching causes regular fluctuations in production parameters at 8-hour intervals. The 2-3 hour cycle corresponds to the charging rhythm, which is a rhythm formed by the fixed interval of furnace charge addition in high-carbon ferrochrome production, usually every 2-3 hours, and the change in the amount of furnace charge causing fluctuations in gas production. For each identified periodic component, its amplitude and phase are recorded, and a unique periodic component identifier is assigned to it. The amplitude represents the fluctuation intensity of the periodic rhythm; the larger the value, the more significant the impact of the cycle on gas production. For example, the amplitude of the daily rhythm is usually the largest, indicating that the difference between day and night operation is the dominant production rhythm. The phase represents the starting time of the periodic rhythm, reflecting the time starting point of the rhythm fluctuation. For example, the phase of the charging rhythm corresponds to the start time of each charging, that is, the initial time when the gas production begins to fluctuate with the addition of charge. The periodic component identifier, amplitude, and phase of each periodic component are combined to form the rhythmic feature of the corresponding periodic component. The rhythmic features of all periodic components are integrated to form the production rhythm feature vector.

[0084] For example, the production rhythm feature vector (PRV) can be represented as [(T24, A24, φ24), (T8, A8, φ8), (T3, A3, φ3)], where (T24, A24, φ24) represents the rhythm feature of a 24-hour cycle, where T24 is the periodic component identifier of the 24-hour cycle, A24 is the amplitude of the daily rhythm, and φ24 is the phase of the daily rhythm; (T8, A8, φ8) represents the rhythm feature of an 8-hour cycle, where T8 is the periodic component identifier of the 8-hour cycle, A8 is the amplitude of the shift change rhythm, and φ8 is the phase of the shift change rhythm, i.e., the start time point, such as the handover time of each shift; (T3, A3, φ3) represents the rhythm feature of a 3-hour cycle, where T3 is the periodic component identifier of the 3-hour cycle, A3 is the amplitude of the charging rhythm, and φ3 is the start time point of the charging rhythm, such as the start time of each furnace charge addition. Production rhythm feature vectors condense multi-scale rhythm information into structured data, enabling quantitative representation of production patterns.

[0085] The Fast Fourier Transform (FFT) is employed because it can efficiently decompose periodic signals. Compared to other time-frequency analysis methods such as wavelet transform, it has lower computational complexity, is more suitable for the real-time requirements of industrial scenarios, and has higher accuracy in identifying the periodicity of strongly periodic signals. Multi-scale time-frequency analysis addresses the fragmented nature of production rhythm recognition in existing technologies. By comprehensively capturing the periodic characteristics at each level, subsequent window identification no longer relies on a single indicator, such as only looking at gas volume. Instead, it judges the trough period based on the overall production rhythm. For example, by combining the phase of the feeding rhythm, the load trough window after each feeding can be predicted. This load trough window has relatively stable in-furnace reactions, resulting in less disturbance to production during maintenance. This is an implicit window mining effect that traditional methods cannot achieve. Without this step, the subsequent calculation of the production load index will lack a rhythmic benchmark, making it impossible to distinguish between "normal fluctuations" and "true troughs," leading to the identified window overlapping with production peaks and causing unnecessary production stoppages and losses.

[0086] Step S22: Calculate the production load index, determine the effective idle window based on the production rhythm feature vector and the production load index, evaluate the purity of the effective idle window and determine the type of the effective idle window;

[0087] Production load index (PLI) equals the current gas production (Gc) and the average gas production (G) over the past 24 hours. avg The ratio. Among them, Gc is extracted in real time from the intelligent monitoring system for waste gas treatment, G... avg The gas production is calculated by taking the arithmetic mean of the hourly gas production over the past 24 hours. This calculation formula uses 24 hours as the average window, taking into account the diurnal rhythm of the high-carbon ferrochrome furnace, which dominates production fluctuations, to ensure G... avg It can smooth out short-term noise and capture daily trends, avoiding drastic index fluctuations caused by single-hour data.

[0088] Please see Figure 3 As shown, the method for determining the effective idle window based on the production rhythm feature vector and the production load index includes: when the production load index PLI is less than the load threshold PL low If the current time period is considered a potential idle window, it is excluded; otherwise, it is excluded. The current time period is defined as a low-load interval formed by consecutive PLI calculation points starting from the current moment. low The settings are based on the PLI distribution during production troughs in historical production data, determined by calculating the 10th percentile of PLI over the past M2 months to cover most low-load periods and avoid over-screening. For example, PLI... lowSet to 0.6. This determination method solves the problem of low accuracy caused by the reliance on static thresholds for identifying idle production windows in existing technologies. This is because static thresholds ignore daily trends, while the relative ratio of PLI allows the determination to adapt to dynamic operating conditions. It can capture troughs under seasonal changes, such as when the overall load is high in summer, it can still relatively identify windows and expand the available time period.

[0089] After identifying potential idle windows, a qualitative verification is performed using the Production Rhythm Feature Vector (PRV). If the qualitative verification passes, the potential idle window is confirmed as a true trough window; otherwise, it is excluded. The qualitative verification process uses a vector matching algorithm, comparing the start time of the potential idle window with the phase of each periodic component in the PRV. The matching degree is calculated as a weighted inverse sum of phase differences, with weights set according to amplitude. Higher amplitude periods have greater weights to highlight the dominant rhythm. If the matching degree is higher than a preset matching degree threshold Mthre, it is confirmed as a true trough window; otherwise, it is excluded. Mthre is obtained through statistical analysis of historical misjudgment cases, taking the minimum matching degree that leads to conflict as the basis. For example, Mthre is set to 0.7. For example, if the start time of a potential idle window matches phase φ3 of the 2-3 hour feeding rhythm in the PRV, i.e., it is in the stable reaction period after feeding, it is confirmed as a true low-temperature window because the gas production naturally decreases during this period, minimizing maintenance disturbances. If it matches phase φ8 of the 8-hour shift change rhythm, i.e., it is in a temporary low point before the shift change, it is excluded because subsequent load recovery is likely to cause conflicts. This verification method solves the defect of false low-temperature misjudgment caused by a single load index in the prior art, because the PLI only quantitatively screens low loads, while the PRV adds rhythmic context to distinguish between random fluctuations and periodic lows, which can reveal hidden opportunities. For example, feeding lows may be short-lived but have high stability, which can be converted into small windows, accumulating fragmented time for optimized scheduling. It also solves the problem of blind window adaptation. By qualitatively defining and connecting screening and continuous inspection, it promotes the solution towards precise collaboration. If PRV qualitative verification is lacking, potential idle windows are prone to include false troughs, which will cause the Gantt chart arrangement in step S24 to conflict with production peaks. Because PLI without PRV ignores the rhythmic nature, the random low points are hidden, making the window unstable. PRV qualitative verification can filter out false windows, so that effective identification is based on the real rhythm.

[0090] The true low-price window is continuously evaluated. If the evaluation passes the continuous evaluation, the true low-price window is determined as a valid idle window (IW). The method for evaluating the continuous evaluation is as follows: if PLI is continuously less than PL for th hour... lowIf th is greater than the minimum maintenance time tmin, it is confirmed as a valid idle window IW; otherwise, it is excluded. tmin is calculated based on the shortest maintenance time statistics for each equipment type, taking the average of historical data plus a safety margin. For example, tmin is set to 1 hour. Continuity judgment solves the problem of frequent maintenance resource conflicts. Continuity confirmation connects verification and purity assessment, promoting the scheme towards constraint optimization. Without continuity judgment, the true low-temperature window is prone to brief interruptions, causing the matching in step S23 to fail.

[0091] The purity of an effective idle window (IW) is assessed by statistically analyzing the occurrence rate of critical production events within the window. If the number of critical production events within the effective idle window is less than the event threshold Elow, the effective idle window is marked as high purity; otherwise, it is marked as restricted and only suitable for low-impact maintenance. Elow is calculated based on historical window conflict cases, taking the average event value of conflict-free windows. For example, Elow is set to 1. Event location utilizes PRV phase, such as the feeding rhythm phase φ3 indicating the start time, to avoid the window covering order delivery, etc. This purity method solves the defect in the prior art where window availability ignores event interference, because events such as quality inspection can amplify the impact of maintenance. Purity marking classifies windows, redirecting restricted windows to low-impact tasks, indirectly extending the total available time and improving scheduling flexibility.

[0092] The type of valid idle window (IW) is determined by its duration; if the duration is greater than 8 hours, it is classified as a large idle window (IW). L Suitable for large equipment maintenance; 2-8 hours is a medium-sized idle window (IW). M Suitable for valve maintenance; less than 2 hours is a small idle window (IW). S This approach is suitable for rapid maintenance. The effective idle window (IW) type classification solves the problem of blind window adaptation because duration matching aligns task types; for example, small windows accumulate to fill fragmented time, reducing overall waiting time. Step S22 improves window recognition accuracy through layer-by-layer integration of PLI filtering, PRV verification, persistence judgment, purity assessment, and type classification. Hidden low-level detection and event avoidance work together to reduce disturbances and lower the conflict rate.

[0093] Step S23: Construct a maintenance task priority queue based on the maintenance disturbance chain level, and perform the best fit and matching of maintenance tasks with available free windows;

[0094] The construction of the maintenance task priority queue is based on the maintenance vulnerability chain level MSC', integrating the equipment health score H and maintenance complexity C to form a priority calculation model P. The calculation formula of the priority model is: P=wm×N(MSC')+wh×N(1-H)+wc×N(C). Here, wm, wh, and wc are the weight coefficients of the maintenance vulnerability chain level, equipment health, and maintenance complexity, respectively, and their sum is 1. Through the analytic hierarchy process (AHP), it is determined that the maintenance vulnerability chain level has the most significant impact on system risk; therefore, wm takes the highest value, for example, wm=0.5, wh=0.3, and wc=0.2. The quantification rule for the maintenance vulnerability chain level MSC' is: Level 1 maintenance vulnerability chain MSC1 corresponds to a value of 1, Level 2 maintenance vulnerability chain MSC2 corresponds to 0.6, and Level 3 maintenance vulnerability chain MSC3 corresponds to 0.2. This quantification directly follows the vulnerability level classification result from step S14. The equipment health score H is obtained through the calculation model in step S15. 1-H represents the risk of equipment failure, and the larger the value, the higher the priority. The maintenance complexity C is obtained through the complexity calculation formula in step S15. The larger the value, the higher the maintenance difficulty, and resources should be allocated in priority.

[0095] The priority queue is sorted in descending order, meaning maintenance tasks with higher P values ​​are ranked higher. For example, a gas cooler in MSC1 has H=0.4 and C=0.9. The priority calculation model is P=0.5×1+0.3×(1-0.4)+0.2×0.9=0.5+0.18+0.18=0.86;

[0096] Another terminal exhaust valve in MSC3 has H=0.7 and C=0.2. The priority calculation model is P=0.5×0.2+0.3×(1-0.7)+0.2×0.2=0.1+0.09+0.04=0.23. Therefore, the maintenance task of the gas cooler takes priority over the terminal exhaust valve. This multi-dimensional weighted sorting method solves the problem of delayed maintenance of high-risk equipment caused by traditional scheduling that only prioritizes fault status, achieving a comprehensive balance between risk and difficulty.

[0097] The matching of maintenance tasks with available free windows employs a two-way filtering mechanism. First, it matches tasks based on their estimated duration and window type: tasks with an estimated duration greater than 8 hours are only matched with IW. L Tasks requiring 2-8 hours are compatible with IW M Tasks shorter than 2 hours are compatible with IW. SDetermining the estimated duration requires retrieving historical maintenance duration data for the equipment, taking the arithmetic average, and adding a 20% safety margin. This safety margin is used to handle unexpected technical issues. Secondly, time verification is performed using the joint maintenance assessment window of the maintenance disruption chain: Task MSC1 must be adapted within 24 hours, MSC2 within 72 hours, and MSC3 within one week. If the verification fails, other windows are re-matched. Finally, the coupling relationship between tasks is verified using the visualized equipment maintenance coupling map in step S15. If two tasks belong to the same maintenance disruption chain, they must be scheduled for joint maintenance within the same window to avoid cascading risks caused by individual maintenance.

[0098] This adaptation process addresses the shortcomings of traditional maintenance scheduling, such as blind task and window adaptation and scattered arrangement of coupled tasks, achieving optimal resource allocation. It works in conjunction with the maintenance disturbance chain in step S14, the coupling graph in step S15, and the window attributes in step S22. The disturbance chain level determines the adaptation timeliness, the coupling graph reveals task relationships, and the window attributes limit the adaptation scope. The combination of these three ensures that the matching result meets both risk control requirements and production rhythm constraints. Without this step, maintenance tasks cannot be arranged in an orderly manner according to risk level. High-risk tasks may be delayed because low-risk tasks occupy windows, and the scattered maintenance of coupled equipment can easily trigger cascading failures, leading to a double increase in maintenance costs and production losses.

[0099] Step S24: Integrate the best fit matching results and introduce multi-dimensional constraints to generate a visualized maintenance time series Gantt chart.

[0100] The integration of matching results requires the establishment of a task-window mapping table to clarify the window time period, window type, window purity and planned maintenance duration for each maintenance task. The field design of the mapping table must include core information such as task identifier, device ID, vulnerable chain level, estimated duration, window time period and window attributes to ensure data integrity and traceability.

[0101] The multi-dimensional constraint checks must be performed sequentially in the order of personnel, spare parts, and tools. Personnel constraint checks are implemented by constructing a skill matching matrix. The matrix row index represents the maintenance personnel identifier, the column index represents the equipment type, and the element value is the personnel's proficiency in maintaining the equipment (on a scale of 1 to 5). Proficiency is assessed based on the personnel's historical maintenance records and performance evaluations. During the check, it must be ensured that the proficiency of personnel matched for each task is not lower than 4 points, and that there is no overlap in personnel assignments within the same time period. Spare parts constraint checks rely on the spare parts management system to retrieve the inventory quantity of spare parts required for the task and compare it with the spare parts demand. The spare parts demand is determined according to the equipment maintenance manual, and the inventory quantity must be greater than or equal to the sum of the demand and the safety stock. The safety stock is calculated based on the spare parts procurement cycle and consumption rate. If the inventory is insufficient, the procurement process is triggered, and the task sequence is adjusted. Tool constraint checks obtain the tool occupancy time periods through the tool management system to ensure that the tools required for the task are idle within the window time period. If there are tool conflicts, alternative tools are coordinated or the task sequence is adjusted.

[0102] When constraint checks fail, optimization adjustments must be performed, using a greedy algorithm to reallocate resources: prioritizing the resource needs of high-priority tasks, while low-priority tasks can be postponed to the next window; tasks within the same easily disrupted chain should be prioritized for joint maintenance to reduce resource consumption; in case of tool conflicts, alternative tools should be used first, and if no alternative tools are available, the window time period should be adjusted. During the adjustment process, the task-window mapping table must be updated in real time and synchronized to the visualization device to maintain the coupling graph, ensuring consistency between the graph and the Gantt chart data.

[0103] The maintenance sequence Gantt chart is generated using a horizontal time axis design. The horizontal axis represents the time scale, and the vertical axis represents the maintenance task identifier. Each task is presented as a bar, and the length of the bar represents the planned downtime. The planned downtime is equal to the sum of the planned maintenance time and the equipment start-up and shutdown time. The equipment start-up and shutdown time is the total time from the issuance of the shutdown command to complete shutdown, and from the issuance of the start command to the restoration of standard operating conditions. It needs to be differentiated according to the equipment type (e.g., cooler start-up and shutdown need to include thermal equilibrium time, valve start-up and shutdown are mechanical action time). The specific value is determined by taking the arithmetic mean of the actual start-up and shutdown times of the equipment in the past M* times (M*≥5). The visual encoding of the Gantt chart must be consistent with the coupling map in step S15: tasks with different levels of susceptibility to disturbance are colored differently, MSC1 is red, MSC2 is orange, and MSC3 is yellow; the thickness of the bars represents the maintenance complexity, with thicker lines for higher complexity; task bars corresponding to high-purity windows are given a green border, and those corresponding to restricted windows are given a yellow border. The Production Rhythm Feature Vector (PRV) is reflected in the timeline annotations. Dashed lines are added to the corresponding phase positions of the daily rhythm, shift change rhythm, and material feeding rhythm to intuitively show the fit between the maintenance window and the production rhythm, ensuring that maintenance arrangements avoid production peaks.

[0104] This visualization method addresses the shortcomings of traditional maintenance plans, such as fragmented information and difficulty in detecting constraint conflicts. It enables all relevant personnel to intuitively grasp the relationship between task timing, resource allocation, and production rhythm, thereby improving decision-making efficiency. In conjunction with the production rhythm in step S21 and the matching results in step S23, the production rhythm guides the timing arrangement, the matching results determine the task content, and constraint checks ensure execution feasibility. The combination of these three elements makes the Gantt chart both realistic and practical.

[0105] Step S30: Extract the maintenance impact feature vector of each maintenance task from the maintenance timeline Gantt chart, construct a production loss prediction model, and predict the total production loss of the current maintenance plan based on the maintenance impact feature vectors of all maintenance tasks and the production loss prediction model.

[0106] Further, step S30 includes:

[0107] Step S31: Extract maintenance features from the maintenance time-series Gantt chart, and combine them with the time-varying equipment coupling strength matrix to construct the maintenance impact feature vector for each maintenance task;

[0108] The maintenance features extractable from the maintenance sequence Gantt chart include: the ID of the device to be maintained, the planned maintenance duration, the level of the maintenance disturbance chain to which it belongs, the maintenance window type, and the window purity. The ID of the device to be maintained is used to associate basic equipment information with historical maintenance data, ensuring the accuracy of feature tracing. The planned maintenance duration reflects the period during which the task occupies production; the longer the duration, the more significant the disturbance to the production rhythm may be. The level of the maintenance disturbance chain directly follows the risk classification result of step S14; a higher level means a greater risk of cascading failures during maintenance, indirectly amplifying production losses. The maintenance window type is associated with the window classification in step S22; different types of windows have different production load bases and stability, and the maintenance of high-purity large windows usually has less interference to production than that of restricted small windows. The comprehensive coupling strength in the time-varying equipment coupling strength matrix DCSM quantifies the degree of correlation between devices. This is achieved by comparing the comprehensive coupling strength between the device to be maintained and all devices within the system with a comprehensive coupling strength greater than α. strongThe coupling influence degree of the equipment to be maintained is obtained by summing the comprehensive coupling strength of the equipment. The calculation logic of the coupling influence degree is based on the principle that "the tighter the coupling of the equipment, the more significant the interference of maintaining the equipment on the operation of related equipment, and the more cumulative the production losses caused by it." For example, a gas cooler is in the first-level maintenance disturbance chain, and its coupling influence degree covers the sum of the comprehensive coupling strength with key equipment such as the main valve and pretreatment equipment, which can comprehensively reflect the chain influence range of maintenance behavior. The construction of the maintenance influence feature vector (MIV) requires the structured integration of the above features. Its expression is MIV = [planned maintenance duration, maintenance disturbance chain level, maintenance window type, window purity, coupling influence degree]. The maintenance disturbance chain level needs to be quantified according to preset rules. The first-level maintenance disturbance chain corresponds to a value of 1, the second level corresponds to 0.6, and the third level corresponds to 0.2. The maintenance window type can be converted into a computable parameter through numerical encoding, such as 0.8 for large windows, 0.5 for medium windows, and 0.2 for small windows, to ensure that all elements in the vector are numerical to adapt to the model input. Window purity can be converted into a computable parameter through numerical encoding: 1.0 for high-purity windows and 0.4 for restricted windows. Step S31 addresses the deficiency of single loss prediction features in existing technologies. Traditional methods often only use maintenance duration as the basis for loss assessment, ignoring the influence of coupling relationships and window characteristics, resulting in a large deviation between prediction results and actual values. MIV, on the other hand, integrates features from four dimensions: time (planned maintenance duration), risk (vulnerability chain level), environment (window type + purity), and association (coupling influence), achieving a multi-dimensional characterization of maintenance impact. This allows subsequent predictions to capture hidden loss factors. For example, for maintenance tasks of the same duration, devices in strong coupling chains cause higher losses than devices in weak coupling chains. This difference can only be identified by incorporating the coupling influence.

[0109] Step S32: Based on historical maintenance records and corresponding production loss data, train a production loss prediction model;

[0110] The collection of historical maintenance records must cover complete data from the past M3 months. M3 is set based on the maximum value of the production rhythm cycle and the equipment maintenance cycle, ensuring the data includes at least one complete production-maintenance cycle. For example, M3 can be set to 6. Historical maintenance records must include characteristic data corresponding to MIVs, such as the actual maintenance duration of the maintenance task, the actual maintained equipment ID, the level of the susceptible chain, the execution window type, and the execution window purity, as well as corresponding production loss data. Production loss data is extracted synchronously from the waste gas treatment intelligent monitoring system and the production management system. Specific indicators include the percentage decrease in gas production during maintenance, the loss of kilowatt-hours of power generation, and the reduction in ferrochrome production. The percentage decrease in gas production is calculated as (average gas production in the 24 hours before maintenance - average gas production during maintenance) / average gas production in the 24 hours before maintenance. The loss in power generation and the reduction in ferrochrome production are obtained by comparing the production difference between the maintenance period and the historical period without maintenance. A historical maintenance impact feature vector is constructed by taking the actual maintenance duration, actual maintenance device ID, level of the susceptible chain, execution window type, and execution window purity as input to the training process. The corresponding production loss data is used as the true label for the training process to construct a sample dataset.

[0111] The production loss prediction model selected is the Random Forest algorithm. This algorithm achieves prediction by constructing multiple decision trees and taking the average of the prediction results. Its advantage lies in its strong ability to fit nonlinear relationships. The relationship between maintenance behavior and production loss is affected by multiple factors such as coupling strength and window characteristics, resulting in complex nonlinear correlations that traditional linear regression models cannot accurately capture. At the same time, Random Forest effectively reduces the risk of overfitting through Bootstrap sampling and random feature selection mechanisms, adapting to the characteristics of relatively limited sample size and noise in industrial scenarios. Compared with algorithms such as Support Vector Machine, its output results are more interpretable, and the influence weight of each dimension on the loss can be clearly defined through feature importance assessment, providing a basis for subsequent optimization schemes. The training process for the production loss prediction model requires dividing the sample dataset into a training set and a test set in a 7:3 ratio. The training set is used to construct a decision tree forest, while the test set is used to verify the model's performance. Mean absolute error (MAE) and coefficient of determination (COD) are used as evaluation metrics. MAE reflects the average deviation between predicted and actual values, while COD characterizes the model's ability to explain data patterns. Hyperparameters such as the number of decision trees, maximum depth, and number of node splitting features are adjusted using a grid search method until the model's evaluation metrics on the test set reach preset standards, such as a MAE of less than 5% and a COD greater than 0.8. Step S32 addresses the shortcomings of existing technologies, such as low accuracy and poor generalization ability in production loss prediction. Traditional empirical prediction methods rely on the subjective judgment of maintenance personnel, resulting in high errors. In contrast, the random forest-based model learns historical patterns through data-driven learning, controlling prediction errors within a reasonable range.

[0112] Step S33: Apply the trained production loss prediction model to predict the total production loss of the current maintenance plan and generate an optimization scheme.

[0113] The prediction process requires inputting the maintenance impact feature vectors of all maintenance tasks in the current maintenance timeline Gantt chart into the trained model one by one. The model outputs the predicted production loss value for a single maintenance task, and the total predicted production loss value of the current maintenance plan is obtained by superimposing the predicted production loss values ​​of all maintenance tasks. The predicted production loss value for a single maintenance task includes the percentage decrease in gas production during the predicted maintenance period relative to the baseline value before maintenance, the predicted power generation loss in kilowatt-hours during the predicted maintenance period, and the reduction in ferrochrome production.

[0114] After prediction, an optimization plan is generated by combining the feature importance assessment results output by the model. Feature importance assessment can clarify the impact weight of each dimension on the loss. If the coupling influence weight is the highest, multiple closely coupled tasks in the same maintenance disturbance chain are merged into the same window for joint maintenance to reduce the cumulative loss caused by multiple downtimes. If the maintenance window type weight is high, tasks that are adapted to small, restricted windows and have large estimated losses are adjusted to be executed in large, high-purity windows to reduce additional losses caused by interference. If the planned maintenance duration weight is high, tasks whose duration exceeds the window adaptation range are split into maintenance steps or resource allocation is optimized to shorten the duration.

[0115] Step S33 addresses the technical problems of subjective and one-sided production loss prediction and the lack of clear optimization direction in traditional maintenance, while also compensating for the disconnect between prediction and decision-making. Step S33 forms a closed loop through data-driven prediction and feature-oriented optimization, enabling maintenance plans to further reduce production losses while keeping risks under control. In conjunction with the maintenance sequence Gantt chart in Step S24, the optimized plan can directly feed back into the Gantt chart adjustment, making the sequence arrangement more aligned with low-loss requirements. In conjunction with the time-varying equipment coupling strength matrix in Step S13, it ensures that the optimized plan accurately targets the hidden loss points caused by coupling correlations, achieving refined optimization beyond empirical judgment. Without this step, the loss prediction in Step S30 would remain only at the numerical level, unable to be transformed into practical maintenance plan improvements, making it difficult for the maintenance plan to continuously adapt to production rhythms and equipment coupling characteristics, and production losses could not be effectively controlled.

[0116] Example 2

[0117] This embodiment, based on Embodiment 1, provides a predictive maintenance system for flue gas treatment facilities based on machine learning, such as... Figure 4 As shown, it includes:

[0118] Coupled Association Identification Module: Used to collect the equipment list of the gas purification system and the three-dimensional spatial coordinates of each equipment in the equipment list, and construct an equipment physical topology database; based on the equipment physical topology database, construct a time-varying equipment coupling strength matrix; and identify and maintain easily disturbed chains according to the time-varying equipment coupling strength matrix.

[0119] Maintenance time-series Gantt chart generation module: used to acquire historical production data of the gas purification system, perform multi-scale time-frequency analysis on the historical production data, generate production rhythm feature vectors; and generate maintenance time-series Gantt charts based on the production rhythm feature vectors and maintenance perturbation chains.

[0120] Loss prediction module: used to extract the maintenance impact feature vector of each maintenance task from the maintenance timeline Gantt chart, build a production loss prediction model, and predict the total production loss of the current maintenance plan based on the maintenance impact feature vectors of all maintenance tasks and the production loss prediction model.

[0121] Furthermore, in the coupling association identification module, the method for constructing the device physical topology database includes:

[0122] Establish a device connection matrix ECM. The device connection matrix ECM is a symmetric square matrix. The element ECMij in the i-th row and j-th column of ECM is set to 1 when there is a direct connection between device i and device j, and otherwise set to 0.

[0123] Each device is assigned a unique identifier, and the device's unique identifier, three-dimensional spatial coordinates, and device connection matrix constitute the device's physical topology database.

[0124] In the coupling correlation identification module, the method for constructing the time-varying device coupling strength matrix includes:

[0125] Based on the device physical topology database, virtual tracer particle injection and trajectory tracking are performed to generate a medium flow trajectory dataset;

[0126] The media transfer time and temperature conduction influence coefficient between devices are calculated based on the media flow trajectory dataset; a time-varying device coupling strength matrix is ​​constructed based on the media transfer time and temperature conduction influence coefficient between devices.

[0127] In the coupling association identification module, the medium flow trajectory dataset includes at least the time when the particles enter and leave the device, and the medium temperature when the particles enter and leave the device;

[0128] The medium transfer time between the devices is calculated based on the time it takes for particles to enter and leave the devices, and the temperature conduction influence coefficient between the devices is calculated based on the medium temperature when particles enter and leave the devices.

[0129] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0130] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0131] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A predictive maintenance method for flue gas treatment facilities based on machine learning, characterized in that, The method includes: Collect the equipment list of the gas purification system and the three-dimensional spatial coordinates of each device in the equipment list to construct a physical topology database of the equipment; construct a time-varying equipment coupling strength matrix based on the physical topology database of the equipment; identify and maintain easily disturbed chains according to the time-varying equipment coupling strength matrix; The method for constructing the time-varying device coupling strength matrix includes: based on the device physical topology database, performing virtual tracer particle injection and trajectory tracking to generate a medium flow trajectory dataset; calculating the medium transfer time and temperature conduction influence coefficient between devices based on the medium flow trajectory dataset; and constructing the time-varying device coupling strength matrix based on the medium transfer time and temperature conduction influence coefficient between devices. The method for identifying maintenance-prone chains based on the time-varying device coupling strength matrix includes: establishing a device connection matrix (ECM); using a depth-first search algorithm; identifying candidate maintenance-prone chains based on the ECM; after determining the candidate maintenance-prone chains, calculating the chain length and chain strength of the candidate maintenance-prone chains; if the chain length of the candidate maintenance-prone chain is greater than a length threshold and the chain strength is greater than a strength threshold, then the candidate maintenance-prone chain is confirmed as a maintenance-prone chain. Historical production data of the gas purification system is acquired, and multi-scale time-frequency analysis is performed on the historical production data to generate a production rhythm feature vector; based on the production rhythm feature vector and the maintenance perturbation chain, a maintenance time-series Gantt chart is generated. The method for generating the maintenance time-series Gantt chart includes: calculating the production load index, determining the effective idle window based on the production rhythm feature vector and the production load index; assigning a disturbance level to each maintenance disturbance chain, constructing a maintenance task priority queue based on the maintenance disturbance chain level, and performing the best fit matching between maintenance tasks and the effective idle window; integrating the best fit matching results and introducing multi-dimensional constraints to generate a visualized maintenance time-series Gantt chart. The maintenance impact feature vector of each maintenance task is extracted from the maintenance timeline Gantt chart, and a production loss prediction model is constructed. Based on the maintenance impact feature vectors of all maintenance tasks and the production loss prediction model, the total production loss of the current maintenance plan is predicted.

2. The predictive maintenance method for flue gas treatment facilities based on machine learning according to claim 1, characterized in that, The method for constructing the device physical topology database includes: The device connection matrix ECM is a symmetric square matrix. The element ECMij in the i-th row and j-th column of ECM is set to 1 when there is a direct connection between device i and device j, and otherwise set to 0. Each device is assigned a unique identifier, and the device's physical topology database is composed of the device's unique identifier, three-dimensional spatial coordinates, and device connection matrix.

3. The predictive maintenance method for flue gas treatment facilities based on machine learning according to claim 2, characterized in that, The medium flow trajectory dataset includes at least the time when particles enter and leave the device, and the medium temperature when particles enter and leave the device. The medium transfer time between the devices is calculated based on the time it takes for particles to enter and leave the devices, and the temperature conduction influence coefficient between the devices is calculated based on the medium temperature when particles enter and leave the devices.

4. The predictive maintenance method for flue gas treatment facilities based on machine learning according to claim 3, characterized in that, The method for constructing a time-varying device coupling strength matrix based on the medium transfer time and temperature conduction influence coefficient between devices includes: Based on the device connection matrix ECM, the shared pipe length between devices is calculated, and the comprehensive coupling strength between devices is calculated based on the medium transfer time, temperature conduction influence coefficient and shared pipe length. The time-varying device coupling strength matrix uses the device's unique identifier as the row and column index and the overall coupling strength as the element value.

5. The predictive maintenance method for flue gas treatment facilities based on machine learning according to claim 4, characterized in that, The method for identifying candidate maintenance-prone chains includes: A depth-first search algorithm is employed, starting with each device as the initial node. Based on the device connectivity matrix (ECM), all direct adjacent nodes of the initial node are generated, forming the first layer of nodes in the search. Using these first-layer nodes as new starting points, the direct adjacent nodes are recursively generated, forming all reachable connection paths between devices. The algorithm then determines whether the overall coupling strength of all adjacent device pairs in each reachable connection path is greater than the strong coupling threshold α. strong If so, the reachable connection path will be used as a candidate maintenance perturbation chain.

6. The predictive maintenance method for flue gas treatment facilities based on machine learning according to claim 5, characterized in that, The historical production data of the gas purification system includes at least the amount of gas produced. The method for generating the production rhythm feature vector includes: constructing a time series of gas production, performing a fast Fourier transform on the time series of gas production to identify periodic components; recording the amplitude and phase of each identified periodic component, and assigning a unique periodic component identifier to each periodic component; the periodic component identifier, amplitude, and phase of each periodic component constitute the rhythm feature of the corresponding periodic component; and integrating the rhythm features of all periodic components to form the production rhythm feature vector.

7. The predictive maintenance method for flue gas treatment facilities based on machine learning according to claim 6, characterized in that, The method for determining the effective free window includes: Potential idle windows are determined based on the production load index. After determining the potential idle windows, a qualitative verification is performed using the production rhythm feature vector to confirm whether the potential idle windows are true low-temperature windows. The true low-temperature windows are then subjected to a persistence judgment. If the persistence judgment is passed, the true low-temperature window is determined to be a valid idle window.

8. A machine learning-based predictive maintenance system for flue gas treatment facilities, used to implement the machine learning-based predictive maintenance method for flue gas treatment facilities as described in any one of claims 1-7, characterized in that, The system includes: Coupled Association Identification Module: Used to collect the equipment list of the gas purification system and the three-dimensional spatial coordinates of each equipment in the equipment list, and construct an equipment physical topology database; based on the equipment physical topology database, construct a time-varying equipment coupling strength matrix; and identify and maintain easily disturbed chains according to the time-varying equipment coupling strength matrix. Maintenance time-series Gantt chart generation module: used to acquire historical production data of the gas purification system, perform multi-scale time-frequency analysis on the historical production data, generate production rhythm feature vectors; and generate maintenance time-series Gantt charts based on the production rhythm feature vectors and maintenance perturbation chains. Loss prediction module: used to extract the maintenance impact feature vector of each maintenance task from the maintenance timeline Gantt chart, build a production loss prediction model, and predict the total production loss of the current maintenance plan based on the maintenance impact feature vectors of all maintenance tasks and the production loss prediction model.

Images