A method and system for predictive maintenance of ferroalloy off-gas purification

Abstract

This application discloses a method and system for predictive maintenance of ferroalloy exhaust gas purification, relating to the field of exhaust gas purification. The method includes acquiring and preprocessing real-time operating condition data and real-time performance data of a ferroalloy exhaust gas purification system to obtain preprocessed data; establishing an operating condition-performance correlation model based on historical operating condition data and historical performance data from a pre-set historical database, and extracting degradation characteristic parameters reflecting changes in the health status of the purification unit from the preprocessed data based on the operating condition-performance correlation model; using a time series prediction algorithm to predict the remaining effective operating time of the purification unit based on the degradation characteristic parameters; triggering graded early warnings based on the remaining effective operating time and outputting performance optimization instructions corresponding to the graded early warnings; adjusting real-time performance data based on the performance optimization instructions, and optimizing the operating condition-performance correlation model based on the adjusted real-time performance data. This application has the effect of predicting and identifying the performance degradation of the purification unit and enabling forward-looking control.

Description

Technical Field

[0001] This application relates to the field of exhaust gas purification, and in particular to a method and system for predictive maintenance of ferroalloy exhaust gas purification. Background Technology

[0002] The exhaust gas from ferroalloy submerged arc furnaces exhibits extreme operating conditions, including high temperatures, high dust levels, and significant fluctuations in pollutant concentrations, posing a significant challenge to exhaust gas purification systems. Existing maintenance methods have significant shortcomings, resulting in low reliability, uneconomical operation, and high environmental compliance risks for purification equipment within the purification unit. First, current technologies cannot isolate the impact of operating condition fluctuations on performance indicators, relying solely on simple threshold comparisons of raw measurements for alarms. This fails to distinguish whether the performance degradation of the purification equipment is due to inherent equipment deterioration or external changes in operating conditions, leading to distorted health status assessments and a lack of scientific basis for maintenance decisions. Second, the absence of a mechanism to predict remaining effective operating time means maintenance is based solely on scheduled inspections or post-failure repairs, failing to provide early warnings of equipment failure. This results in a high risk of sudden shutdowns, frequent production interruptions, and the potential for environmental accidents involving excessive emissions due to sudden equipment failure. Finally, operational parameter control relies on feedback adjustments based on instantaneous emission values, lacking proactive optimization for equipment degradation trends. This leads to persistently high system energy consumption and, in the later stages of purification equipment lifespan, accelerated wear and tear on core consumables due to overcompensation, shortening the equipment's lifespan. These problems collectively constrain the stable operation and sustainable development of ferroalloy exhaust gas purification systems. Summary of the Invention

[0003] To address the issues of being unable to distinguish between operating condition fluctuations and equipment degradation, lacking remaining life prediction, leading to delayed maintenance and high risk of equipment failure, this application provides a method and system for predictive maintenance of ferroalloy exhaust gas purification.

[0004] In the first aspect, the method for predictive maintenance of ferroalloy exhaust gas purification provided in this application adopts the following technical solution: S1, acquire and preprocess the real-time operating condition data and real-time performance data of the ferroalloy exhaust gas purification system to obtain preprocessed data; S2. Establish a condition-performance correlation model based on historical operating condition data and historical performance data in a preset historical database, and extract deterioration characteristic parameters reflecting changes in the health status of the purification unit from the preprocessed data based on the condition-performance correlation model; the purification unit is a device for removing particulate matter or gaseous pollutants from ferroalloy exhaust gas. S3. Based on the aforementioned degradation characteristic parameters, use a time series prediction algorithm to predict the remaining effective operating time of the purification unit in reverse. S4. Trigger a tiered warning based on the remaining effective running time, and output a performance optimization instruction corresponding to the tiered warning; S5. Adjust the real-time performance data based on the performance optimization instructions, and optimize the working condition-performance correlation model based on the adjusted real-time performance data.

[0005] By adopting the above technical solution, real-time operating condition data and real-time performance data of the ferroalloy exhaust gas purification system are acquired and preprocessed. An operating condition-performance correlation model is established based on a pre-set historical database to automatically extract degradation characteristic parameters of the purification unit, enabling automatic and reliable monitoring of the purification unit's health status changes without manual intervention. A time series prediction algorithm is used to predict the remaining effective operating time of the purification unit, achieving early warning of potential failures. Based on the prediction results, tiered warnings are triggered and performance optimization instructions are output to guide subsequent performance data adjustments to dynamically match the degradation trend of the purification unit, ensuring optimal operating status. Real-time performance data is adjusted according to the performance optimization instructions, and the operating condition-performance correlation model is optimized to form a closed-loop adaptive update mechanism for prediction. This solution simplifies the health assessment process of the purification unit, improves the accuracy and predictive nature of predictive maintenance, pre-determines the degradation and failure time of the purification unit, reduces the risk of unexpected downtime and post-failure maintenance costs, optimizes the purification efficiency of the exhaust gas purification system, and ensures the stable operation, timely maintenance, and environmental compliance of the ferroalloy exhaust gas purification system.

[0006] Optionally, S2 specifically includes constructing a baseline pressure difference model and a baseline efficiency model under healthy conditions based on historical operating condition data and historical performance data in a preset historical database; inputting the preprocessed real-time operating condition data into the baseline pressure difference model and the baseline efficiency model respectively to independently calculate the theoretical pressure difference and theoretical purification efficiency; the theoretical pressure difference and the theoretical purification efficiency respectively correspond to the actual pressure difference and actual purification efficiency in the preprocessed real-time performance data to calculate the degradation characteristic parameters.

[0007] By adopting the above technical solution, a baseline pressure difference model and a baseline efficiency model under healthy conditions are constructed based on a preset historical database, automatically calculating the theoretical pressure difference and theoretical purification efficiency. By precisely comparing the theoretical values ​​with the actual pressure difference and actual purification efficiency in real-time performance data, degradation characteristic parameters are independently calculated, achieving automated and high-precision quantitative assessment of the purification unit's health status. This method eliminates the influence of human intervention errors and reduces subjective errors. The independent calculation and comparison between the baseline model and real-time data ensures the reliability and representativeness of degradation characteristic parameters, providing a solid and accurate data foundation for subsequent prediction of remaining effective operating time. It also enhances the early warning capability of predictive maintenance, effectively preventing sudden failures of purification equipment, extending the lifespan of the purification unit, and optimizing the purification efficiency of ferroalloy exhaust gas.

[0008] Optionally, the theoretical pressure difference and the theoretical purification efficiency correspond to the actual pressure difference and the actual purification efficiency in the preprocessed real-time performance data, respectively. The degradation characteristic parameters are calculated by combining the actual pressure difference and the theoretical pressure difference to obtain the normalized pressure difference; the ratio of the actual purification efficiency to the theoretical purification efficiency to obtain the equivalent efficiency; and the derivatives of the normalized pressure difference and the equivalent efficiency to obtain the degradation rate. The normalized pressure difference, the equivalent efficiency, and the degradation rate are combined to obtain the degradation characteristic parameters.

[0009] By adopting the above technical solution, theoretical pressure difference and theoretical purification efficiency are automatically generated based on a benchmark model. The actual pressure difference and actual purification efficiency are innovatively normalized to their corresponding theoretical values, accurately calculating the normalized pressure difference and equivalent efficiency reflecting the instantaneous state. Furthermore, the degradation rate is dynamically captured through derivative calculations, forming a three-dimensional degradation characteristic parameter system that integrates static deviation and dynamic trends. This solution achieves real-time quantitative analysis of the health status of the purification unit, breaking through the limitations of traditional single-parameter monitoring. The normalized pressure difference directly characterizes the degree of physical blockage, the equivalent efficiency accurately maps the performance degradation magnitude, and the degradation rate predicts the direction of fault evolution in advance. The synergistic analysis of these three degradation characteristic parameters significantly improves the sensitivity and reliability of the purification equipment's condition assessment, providing multi-dimensional dynamic input for time series prediction. Meanwhile, by automatically extracting key degradation characteristic parameters through mathematical modeling, the risk of misjudgment by human experience is completely avoided, improving the objectivity, accuracy and stability of degradation characteristic parameters and significantly enhancing the early warning capability of the purification unit for early failure risks. This effectively prevents sudden and severe failures such as sudden pressure rise and efficiency drop, maximizes the lifespan of the purification equipment, reduces the cost of maintenance and repair of the purification equipment, and ensures the continuous and efficient operation of the exhaust gas purification system.

[0010] Optionally, S3 specifically includes extrapolating the trend curve of the degradation characteristic parameter within a future preset time period using a time series prediction algorithm; and calculating the time when the trend curve reaches the preset performance failure threshold based on a preset performance failure threshold to obtain the remaining effective running time.

[0011] By adopting the above technical solution, the future trend curve of degradation characteristic parameters is dynamically extrapolated using a time series prediction algorithm, overcoming the limitations of traditional static degradation thresholds. By using a preset performance failure threshold, the time point when the trend curve reaches the critical state within a preset future time period is calculated in reverse, achieving accurate reverse extrapolation of the remaining effective operating time. This solution innovatively combines forward prediction with reverse verification. The time series prediction algorithm extrapolates to capture the nonlinear evolution law of the degradation process, while the reverse calculation defines the absolute boundary conditions for the failure of the purification equipment, forming a dual verification mechanism. This solution significantly improves the anti-interference capability and applicability of the prediction results of the time series prediction algorithm to complex engineering, effectively solving the false alarm problem caused by data noise in traditional time prediction models. By quantifying the precise time window from the current operating state of the purification equipment to performance failure, a scientific decision-making basis is provided for subsequent graded early warning, ensuring that the triggering time of maintenance commands is strictly synchronized with the actual degradation process of the purification equipment. This achieves a fundamental shift from post-failure repair to pre-failure intervention, minimizing unplanned downtime due to purification equipment failure, extending the service life of the purification unit, reducing the frequency of spare parts replacement, and lowering post-failure maintenance costs.

[0012] Optionally, S5 specifically includes adjusting the real-time performance data according to the performance optimization instruction; calculating the pressure difference residual between the actual pressure difference and the theoretical pressure difference, and the purification efficiency residual between the actual purification efficiency and the theoretical purification efficiency; and optimizing the operating condition-performance correlation model using the pressure difference residual and the purification efficiency residual.

[0013] By adopting the above technical solution, real-time performance data is dynamically adjusted according to performance optimization instructions, achieving precise calibration of the purification unit's operating parameters. A dual-channel quantitative deviation feedback mechanism is constructed by calculating the pressure differential residual between the actual and theoretical pressure differentials, and the purification efficiency residual between the actual and theoretical purification efficiencies. Based on the residual data, the operating condition-performance correlation model is continuously optimized, forming an adaptive closed loop of instruction execution, deviation analysis, and model iteration. This solution overcomes the limitations of traditional static calculation models. The pressure differential residual directly reveals the hydrodynamic deviations caused by filter media blockage, and the purification efficiency residual accurately maps the degree of catalyst activity degradation. The dual residuals synergistically drive the real-time updating of the benchmark model parameters. This significantly improves the dynamic adaptability of the benchmark pressure differential model and benchmark efficiency model, effectively eliminating the distortion problems of the benchmark model caused by aging of purification equipment and operating condition drift. Automatic correction of the theoretical value calculation logic through residual feedback ensures the timeliness and accuracy of deterioration characteristic parameter extraction, ensuring strict synchronization between graded early warning instructions and the actual state of the purification equipment, and realizing the self-evolution capability of the predictive maintenance system.

[0014] Optionally, the construction of the benchmark pressure difference model and benchmark efficiency model under healthy conditions adopts a collaborative modeling framework based on federated graph neural networks. Specifically, it includes constructing a global graph structure through heterogeneous historical operating condition data and historical performance data of distributed device clusters, dynamically aggregating feature representations of similar operating condition samples using graph attention mechanism, and embedding the specific physical constraints of the purification unit as regularization terms to generate the benchmark pressure difference model and benchmark efficiency model.

[0015] By adopting the above technical solutions, a collaborative modeling framework based on federated graph neural networks is used to construct the benchmark pressure difference model and benchmark efficiency model, achieving collaborative modeling under multi-source data fusion. This solution overcomes the limitations of traditional single-device modeling. The federated framework efficiently integrates distributed heterogeneous data while protecting data privacy, eliminating data silos. The graph attention mechanism adaptively focuses on key operating condition sample features, improving the accuracy and robustness of feature extraction in the benchmark model. Physical constraint embedding ensures that the benchmark model strictly follows the actual physical laws of the purification unit, preventing overfitting and enhancing generalization ability. This solution significantly improves the accuracy and adaptability of the benchmark pressure difference model and benchmark efficiency model, reduces the development and maintenance costs of the model, and provides a more reliable theoretical basis for the extraction of degradation feature parameters.

[0016] Optionally, the collaborative modeling framework of the federated graph neural network also includes capturing the cross-time-period dependencies of the distributed device cluster by constructing a spatiotemporal graph convolutional network, designing a causal inference module to analyze the causal relationship between the evolution of operating conditions and the degradation feature parameters, and introducing a lightweight federated distillation strategy to compress the global graph structure.

[0017] By adopting the above technical solution, a spatiotemporal graph convolutional network, a causal inference module, and a lightweight federated distillation strategy are integrated into the federated graph neural network collaborative modeling framework. The spatiotemporal graph convolutional network accurately models the cross-time-period dependencies of distributed device clusters, solving the shortcomings of traditional methods in capturing dynamic correlations. The causal inference module deeply analyzes the true causal relationship between operating parameters and degradation characteristic parameters, eliminating the risk of misjudgment caused by confounding variables. The lightweight federated distillation strategy efficiently compresses global graph structure knowledge, overcoming the engineering bottleneck of limited computing resources in industrial sites. This solution significantly improves the generalization ability and timeliness of the benchmark pressure difference / efficiency model. Spatiotemporal modeling makes the calculation of theoretical pressure difference / efficiency values ​​adapt to the collaborative evolution law of multiple devices. Causal inference ensures the interpretability of the extraction of degradation characteristic parameters. Model compression enables low-latency deployment of a thousand-node-level prediction system on edge devices. A technical closed loop of dynamic correlation modeling, root cause analysis, and quantitative implementation is formed, enabling the predictive maintenance process to have self-evolutionary capabilities, while ensuring that the ferroalloy exhaust gas purification system continuously meets ultra-low emission standards under all operating conditions.

[0018] Optionally, the time series prediction algorithm adopts a dual-channel LSTM network architecture based on wavelet packet decomposition. Constructing the dual-channel LSTM network architecture specifically includes decomposing the degradation feature parameter sequence into a first component and a second component using a wavelet packet decomposition algorithm and constructing the dual-channel LSTM network. The first channel performs trend fitting on the first component representing the degradation of the purification unit, and the second channel models abnormal fluctuations in the second component representing sudden disturbances. An adaptive adjustment module is added before the output layer of the dual-channel LSTM network to dynamically adjust the weight coefficients of the dual channels. A variable step-size rolling prediction mechanism is adopted, automatically shortening the prediction window and triggering online parameter updates of the dual-channel LSTM network when the prediction deviation exceeds a set threshold.

[0019] By adopting the above technical solutions, the time series prediction algorithm innovatively employs a dual-channel LSTM network architecture based on wavelet packet decomposition, ensuring the robustness of the time series prediction algorithm model under complex operating conditions. Combined with a variable step-size rolling prediction mechanism, when the prediction deviation exceeds the limit, the prediction window is automatically shortened and the parameters of the online dual-channel LSTM network model are updated, achieving self-correction in the prediction process. This solution overcomes the limitations of traditional single models. Wavelet packet decomposition eliminates noise interference, the dual-channel architecture independently analyzes steady-state decay and transient disturbances, adaptive weight adjustment enhances the dynamic adaptability of the dual-channel LSTM network model, and the variable step-size mechanism balances prediction accuracy and computational efficiency, significantly improving the accuracy and flexibility of remaining effective running time prediction and effectively reducing the false alarm rate caused by sudden failures of the purification unit.

[0020] Optionally, the adaptive adjustment module adopts a policy optimization mechanism based on a meta-learning framework. Specifically, it dynamically generates the allocation strategy of the weight coefficients of the dual-channel network by constructing a meta-knowledge base and designing a meta-optimizer, and introduces an online knowledge distillation mechanism to compress the complexity of the dual-channel LSTM network.

[0021] By adopting the above technical solutions, the adaptive adjustment module innovatively introduces a strategy optimization mechanism based on a meta-learning framework, forming a technical closed loop of dynamic optimization and lightweight synergy. This solution breaks through the limitations of traditional fixed-weight models; the meta-learning framework enables the weight allocation strategy to have self-evolution capabilities, accurately adapting to complex operating conditions such as sudden changes in flue gas load and raw material fluctuations. Online knowledge distillation, while retaining core time-series prediction capabilities, significantly reduces the computational load of the dual-channel LSTM network model, enabling efficient deployment of edge devices. The synergistic effect of the dual mechanisms greatly improves the prediction robustness and real-time performance of the dual-channel LSTM network model. The meta-knowledge base drives dynamic strategy optimization to eliminate prediction bias caused by sudden disturbances, while the lightweight model reduces response latency to milliseconds, meeting the stringent requirements of continuous production scenarios. This achieves a dual improvement in the prediction accuracy of remaining effective running time and actual engineering efficiency, effectively preventing unplanned downtime and extending the effective service life of the purification unit.

[0022] Secondly, this application provides a predictive maintenance system for ferroalloy exhaust gas purification, which includes a data acquisition module, a preprocessing module, a model building module, a feature extraction module, a time series prediction module, an early warning triggering and instruction generation module, a performance control module, and a model optimization module. The data acquisition module is deployed at the operation site of the purification unit and is used to collect real-time operating condition data and real-time performance data of the ferroalloy exhaust gas purification system. The preprocessing module is connected to the data acquisition module via a wired network and is used to preprocess the real-time operating condition data and the real-time performance data to obtain preprocessed data. The model building module is deployed on a central server and builds a working condition-performance correlation model based on historical working condition data and historical performance data stored in the historical database. The feature extraction module is communicatively connected to the model building module, and extracts degradation feature parameters from the preprocessed data by parsing the working condition-performance correlation model. The time series prediction module is integrated into the edge computing device and predicts the remaining effective running time based on the degradation feature parameters; The early warning triggering and instruction generation module is communicatively connected to the time series prediction module, and triggers different levels of early warning signals based on the remaining effective running time; and generates performance optimization instructions that match the different levels of early warning signals. The performance control module is embedded in the purification unit controller and is used to execute the performance optimization instructions to adjust the real-time performance data; The model optimization module is deployed on a central server, receives the adjusted real-time performance data fed back by the performance control module, and optimizes the operating condition-performance correlation model.

[0023] Understandably, the ferroalloy exhaust gas purification predictive maintenance system provided in the second aspect above is used to execute the method provided in this application. Therefore, the beneficial effects it can achieve can be referred to the beneficial effects in the corresponding method, and will not be repeated here.

[0024] In summary, this application includes at least one of the following beneficial technical effects: 1. Based on the preprocessing of real-time operating condition data and performance data, the establishment of the operating condition-performance correlation model, the extraction of degradation characteristic parameters, the prediction of remaining effective running time, and the closed-loop optimization mechanism, a seamless process from fully automatic data acquisition to early warning triggering is realized. This effectively solves the defects of traditional methods that rely on manual experience for adjustment and prediction response delay, thereby significantly reducing the failure maintenance cost of the purification unit, reducing the risk of unplanned downtime, and improving the operating efficiency of the ferroalloy exhaust gas purification system. 2. Based on a three-dimensional degradation feature parameter system of normalized differential pressure, equivalent efficiency and degradation rate, and a wavelet packet decomposition dual-channel LSTM prediction algorithm, the system achieves high-precision quantitative analysis of the health status of the purification unit and dynamic capture of future trends. Through multi-dimensional feature fusion and adaptive prediction mechanism, the reliability of the remaining effective operating time prediction calculation is ensured, providing a scientific basis for graded early warning. This effectively solves the industry problems of large limitations of single parameter monitoring and high deviation of time prediction results, thereby greatly improving the timeliness and accuracy of early warning for purification equipment prediction and maintenance. 3. Based on federated graph neural network collaborative modeling, meta-learning optimization strategy and modular system design, the system realizes closed-loop control of distributed data fusion, dynamic adaptive optimization of the model and efficient edge deployment, effectively solving the engineering bottlenecks of data silos, outdated models and limited resources, thereby enhancing the operational stability of the ferroalloy exhaust gas purification system, extending the life cycle of the core purification equipment and ensuring the environmental compliance of exhaust gas treatment. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of the predictive maintenance method for ferroalloy exhaust gas purification provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the predictive maintenance system for ferroalloy exhaust gas purification provided in the embodiments of this application; Figure 3This is a flowchart illustrating the construction of a baseline model using a collaborative modeling framework based on federated graph neural networks for predictive maintenance of ferroalloy exhaust gas purification, as provided in another embodiment of this application. Figure 4 This is a flowchart of a dual-channel LSTM network architecture for a method of predictive maintenance of ferroalloy exhaust gas purification provided in another embodiment of this application.

[0027] Attached reference numerals: 1. Data acquisition module; 2. Preprocessing module; 3. Model building module; 4. Feature extraction module; 5. Time series prediction module; 6. Early warning triggering and instruction generation module; 7. Performance control module; 8. Model optimization module. Detailed Implementation

[0028] This application discloses a method and system for predictive maintenance of ferroalloy exhaust gas purification, as detailed below. Figure 1 -Appendix Figure 4 This application will be described in further detail.

[0029] See attached document Figure 1 A method for predictive maintenance of ferroalloy exhaust gas purification includes S1, acquiring and preprocessing real-time operating condition data and real-time performance data of the ferroalloy exhaust gas purification system to obtain preprocessed data.

[0030] Real-time operating data refers to the sequence of input variables collected in real time from sensor networks deployed at key nodes of the exhaust gas purification system, such as dust collector inlet and outlet, desulfurization tower inlet and outlet, and chimney emission outlet, including exhaust gas temperature. Exhaust gas flow Oxygen content Humidity d(t) and inlet pollutant concentration, such as SO2 concentration NOx concentration Real-time performance data refers to the output status sequence collected from the same node, including the pressure difference between the inlet and outlet of the dust removal unit. Concentration of pollutants at the outlet after purification Desulfurization or denitrification efficiency The data includes the induced draft fan current or speed, and vibration signals. Preprocessed data refers to standardized time-series vectors formed after noise filtering, outlier removal, and time synchronization, ensuring the reliability of the collected real-time operating condition and performance data and making them usable for subsequent predictive analysis.

[0031] In practice, real-time operating condition data and performance data are first collected synchronously using sensors such as dust-resistant pressure / differential pressure transmitters, thermocouples, thermal gas mass flow meters, and online flue gas analyzers to form a raw data stream. Then, preprocessing is performed, first using a moving average filtering algorithm to smooth high-frequency noise. The filtering formula is as follows: ,in, This is the filtered sequence. The original sequence, The sliding window width is typically half-width, and 5-10 sampling points are usually selected to eliminate random measurement noise; then the following is applied. Criteria for removing outliers and calculating the mean of data segments and standard deviation If the data points satisfy These are identified as outliers and removed. After removal, forward padding or linear interpolation is used to fill in missing values ​​to ensure the integrity of the collected data. Finally, time alignment and vectorization are performed to unify all sensor data to a baseline sampling period, such as 1 minute, and a strictly synchronized operating condition vector is generated through timestamp alignment. With performance vector ,in , This forms a preprocessed dataset, providing the input data foundation for subsequent modeling.

[0032] S2. Establish a condition-performance correlation model based on historical operating condition data and historical performance data in the preset historical database, and extract deterioration characteristic parameters from the preprocessed data based on the condition-performance correlation model to reflect the changes in the health status of the purification unit; the purification unit is a device for removing particulate matter or gaseous pollutants from ferroalloy exhaust gas.

[0033] In this embodiment, S2 specifically includes constructing a baseline pressure difference model and a baseline efficiency model under healthy conditions based on historical operating condition data and historical performance data in a preset historical database; inputting the preprocessed real-time operating condition data into the baseline pressure difference model and the baseline efficiency model respectively to independently calculate the theoretical pressure difference and theoretical purification efficiency; the theoretical pressure difference and theoretical purification efficiency are respectively combined with the actual pressure difference and actual purification efficiency in the preprocessed real-time performance data to calculate the degradation characteristic parameters.

[0034] Specifically, the theoretical pressure difference and theoretical purification efficiency correspond to the actual pressure difference and actual purification efficiency in the pre-processed real-time performance data, respectively. The degradation characteristic parameters are calculated by calculating the ratio of the actual pressure difference to the theoretical pressure difference to obtain the normalized pressure difference; calculating the ratio of the actual purification efficiency to the theoretical purification efficiency to obtain the equivalent efficiency; and calculating the derivatives of the normalized pressure difference and the equivalent efficiency to obtain the degradation rate. The combination of the normalized pressure difference, equivalent efficiency, and degradation rate yields the degradation characteristic parameters.

[0035] The preset historical database refers to a set of historical health status data, including historical operating condition data and historical performance data, stored for the ferroalloy exhaust gas purification system under healthy conditions. Historical operating condition data includes exhaust gas temperature, flow rate, inlet pollutant concentration, oxygen content, and humidity. Historical performance data includes inlet and outlet pressure difference of the dust removal unit, outlet pollutant concentration after purification, and desulfurization or denitrification efficiency. The purification unit refers to equipment that removes particulate matter or gaseous pollutants from the ferroalloy exhaust gas, such as a bag filter or desulfurization tower. The operating condition-performance correlation model refers to a mathematical relationship model that maps operating condition inputs to expected healthy performance. Deterioration characteristic parameters refer to quantitative health status indicators extracted from real-time operating condition data and real-time performance data that are insensitive to fluctuations in operating conditions, including normalized pressure difference, equivalent efficiency, and degradation rate.

[0036] In practice, firstly, historical health status data—that is, the operating records of the purification equipment when it was not deteriorated—is retrieved from a pre-set historical database. Based on the process mechanism and historical health status data or machine learning methods, a fusion method is used to construct a baseline pressure difference model and a baseline efficiency model under the healthy state. The baseline pressure difference model is modeled using the Kalman-Koznix equation or Gaussian process regression algorithm, representing the theoretical pressure difference corresponding to a given operating condition under the healthy state, and describing the relationship between the theoretical pressure difference and operating condition variables. The formula is: The baseline efficiency model, derived through support vector regression or chemical reaction kinetics, characterizes the theoretical purification efficiency that should be achieved under healthy conditions and given operating conditions, describing the relationship between the theoretical purification efficiency and operating condition variables. The formula is as follows: The parameters of the baseline model are calibrated through regression analysis using historical health data to ensure its accuracy. Then, the standardized time-series vector from step S1, i.e., the preprocessed real-time operating data, is input into the baseline pressure difference model and the baseline efficiency model respectively to independently calculate the theoretical pressure difference and theoretical purification efficiency. Deterioration characteristic parameters are also calculated. First, the normalized pressure difference is obtained by the ratio of the actual pressure difference to the theoretical pressure difference from the real-time performance data, using the formula: This normalized differential pressure value directly reflects irreversible deterioration states such as filter bag clogging and damage, eliminating the impact of fluctuations in operating conditions such as flue gas flow rate and dust concentration. The equivalent efficiency is obtained by comparing the actual purification efficiency with the theoretical purification efficiency from real-time performance data, using the following formula: This equivalent efficiency value characterizes the activity decay of the catalyst or absorbent, eliminating the influence of fluctuations in operating conditions such as inlet concentration and temperature. Then, by differentiating the time series of normalized pressure difference and equivalent efficiency, the degradation rate is calculated using the following formula: , This degradation rate represents the rate of change of degradation characteristic parameters, helping to identify degradation trends. A sudden increase in the degradation rate, for example... An abnormally high level can serve as an early warning signal of accelerated degradation; finally, the normalized pressure difference, equivalent efficiency, and degradation rate are combined into a degradation feature parameter vector, which is then output to the subsequent prediction steps.

[0037] S3. Based on the degradation characteristic parameters, the remaining effective operating time of the purification unit is predicted in reverse using a time series prediction algorithm.

[0038] In this embodiment, S3 specifically includes extrapolating the trend curve of the deterioration feature parameters within a future preset time period using a time series prediction algorithm; and calculating the time when the trend curve reaches the preset performance failure threshold based on the preset performance failure threshold to obtain the remaining effective running time.

[0039] Among them, the degradation characteristic parameters refer to the quantitative indicators of health status extracted from step S2 that are insensitive to fluctuations in operating conditions, including normalized differential pressure, equivalent efficiency, and degradation rate. Time series prediction algorithms refer to computational models used to analyze time-series data and predict future trends, such as the ARIMA model or Long Short-Term Memory (LSTM) neural network. Remaining effective operating time (RUL) refers to the remaining safe operating time of a purification unit, such as a bag filter or desulfurization tower, before its performance reaches a failure threshold, expressed in time units, such as hours or days. Preset performance failure thresholds refer to performance limits set based on equipment manufacturer specifications, emission standards, or historical failure data statistics, such as normalized differential pressure thresholds or equivalent efficiency thresholds, used to determine the failure point of the purification equipment.

[0040] In practice, the historical sequence of degraded feature parameters from the output of step S2 is first retrieved, and a time series prediction algorithm is used for extrapolation prediction. For stationary data, an ARIMA model is selected; for nonlinear or long-sequence dependencies, an LSTM neural network is chosen. The model parameters of the time series prediction algorithm are fitted by training historical health data to predict the feature parameter trend curve for a preset future period, such as 24 hours. The formula is as follows: ,in To predict the trajectory curve, Predictor is the selected time series prediction algorithm. The predicted time period length is then determined. Next, based on a preset performance failure threshold... For example, normalized differential pressure threshold Alternatively, the equivalent efficiency threshold can be used for reverse RUL calculation; for progressive degradation, such as an increase in normalized differential pressure, the RUL can be calculated. The formula is to find the minimum time t that makes the predicted value first reach or exceed the threshold. ;against Degradation-type degradation, such as a decrease in equivalent efficiency, requires finding the minimum time t that causes the predicted value to first reach or fall below a threshold. The formula is as follows: Furthermore, degradation characteristic parameters can be normalized and mapped into intuitive health indicators. , , This allows for the visualization of health status.

[0041] S4. Trigger tiered alerts based on the remaining effective runtime and output performance optimization instructions corresponding to the tiered alerts.

[0042] Among them, the graded early warning refers to a three-level progressive alarm mechanism based on RUL and deterioration rate. The performance optimization command refers to automatically generated digital commands for proactively adjusting real-time performance data, including induced draft fan speed setpoints, desulfurizing agent or denitrification agent dosage setpoints, etc., aiming to delay the deterioration of purification equipment and extend RUL.

[0043] In practice, the RUL and degradation rate are first evaluated based on a preset performance failure threshold. If the RUL is within the first threshold range, for example... Furthermore, if the degradation rate is normal, a Level 1 warning is triggered, generating a maintenance work order and pushing it to the management platform. It is recommended to prepare spare parts in advance and develop a planned maintenance schedule. If the RUL is within the second threshold range, for example... Or the rate of degradation exceeds the acceleration threshold, i.e. For example, if the normalized differential pressure derivative is ≥0.5 / h, a level two early warning is triggered. Simultaneously, a performance optimization command is automatically output to the decision-making and execution layer integrated into the DCS or PLC system for adaptive feedforward performance optimization based on real-time performance data. The command may include slightly increasing the induced draft fan speed to reduce differential pressure load, or optimizing the desulfurizer dosage. The formula is: New setpoint = Original setpoint × (1 + k × Deterioration rate), where k is the adjustment coefficient, ensuring the adjustment is within a smooth constraint to controllably slow down the deterioration trend. If RUL is less than or equal to the third threshold range, for example... Or a transient fault is detected. For example, a sudden increase in differential pressure or a sharp drop in purification efficiency will trigger a level 3 warning, output an emergency alarm to the upper-level system and recommend immediate shutdown for maintenance, while simultaneously activating the safety protection mode to prevent excessive emissions.

[0044] S5. Adjust real-time performance data based on performance optimization instructions, and optimize the working condition-performance correlation model based on the adjusted real-time performance data.

[0045] In this embodiment, S5 specifically includes adjusting real-time performance data according to performance optimization instructions; calculating the pressure difference residual between the actual pressure difference and the theoretical pressure difference, and the purification efficiency residual between the actual purification efficiency and the theoretical purification efficiency; and optimizing the operating condition-performance correlation model using the pressure difference residual and the purification efficiency residual.

[0046] Among these, adjusting real-time performance data refers to the system re-collecting the performance dataset after the real-time performance data is processed according to performance optimization instructions. Pressure differential residual refers to the difference between the actual and theoretical pressure differential. Purification efficiency residual refers to the difference between the actual and theoretical purification efficiency. Optimizing the operating condition-performance correlation model refers to using residual feedback to update the mathematical parameters of the baseline pressure differential model and the baseline efficiency model, ensuring that the model accuracy improves as the real-time performance data evolves.

[0047] In practice, the first step is to execute performance optimization commands. Operating parameters are adjusted through the control logic module integrated into the DCS or PLC system. For example, the induced draft fan speed setpoint may be slightly increased to ensure the adjustment rate remains within smoothness constraints to control energy consumption, or the desulfurizer dosage may be optimized. Under the premise of meeting emission constraints, performance data is adjusted reasonably to guide future performance trajectories. The degradation rate should be minimized or effectively maximized by delaying the time it takes to reach the preset performance failure threshold as gradually as possible. After adjusting the initial real-time performance data, updated performance data, including actual differential pressure and actual purification efficiency, should be collected again in real time. The residuals should be calculated separately, with the differential pressure residual formula as follows: The formula for the purification efficiency residual is: Finally, using the residual optimization model, and through Gaussian process regression or support vector regression algorithms, the parameters of the benchmark pressure difference model and benchmark efficiency model obtained in step S2 are recalibrated, such as the regression coefficients, using the residual sequence as input. The optimization is achieved by minimizing the objective function of the sum of squared residuals, as shown in the formula: Alternatively, the preset performance failure threshold can be dynamically updated based on the residual distribution, for example, setting the updated preset performance failure threshold to the 95th percentile of the historical residuals.

[0048] See attached document Figure 2 This is a schematic diagram of a predictive maintenance system for ferroalloy exhaust gas purification in an embodiment of this application.

[0049] A predictive maintenance system for ferroalloy exhaust gas purification includes a data acquisition module 1, a preprocessing module 2, a model building module 3, a feature extraction module 4, a time series prediction module 5, an early warning triggering and instruction generation module 6, a performance control module 7, and a model optimization module 8.

[0050] The data acquisition module 1 consists of a dust sensor array deployed at the purification unit site, such as the inlet and outlet of the bag filter and the reaction zone of the desulfurization tower. It includes a differential pressure transmitter to collect differential pressure data, thermocouples to measure the exhaust gas temperature, a thermal gas mass flow meter to monitor the flow rate, an online flue gas analyzer to obtain pollutant concentration and oxygen content, and vibration sensors to record the status of the purification equipment. These sensors are connected to an industrial switch via a 4 to 20 mA analog signal or Modbus protocol to collect real-time operating data, such as exhaust gas temperature, inlet flow concentration, and real-time performance data, such as differential pressure, purification efficiency, and fan current.

[0051] The preprocessing module 2 is integrated into the edge computing gateway device. It receives real-time operating condition data and real-time performance data transmitted by the data acquisition module 1 via industrial Ethernet, executes a moving average filtering algorithm to eliminate high-frequency noise and outliers with three times the standard deviation to remove abnormal points, and then performs timestamp alignment to generate standardized operating condition vectors and performance vectors. Finally, it outputs the preprocessed data to the model building module 3.

[0052] Model building module 3 runs on the central server hardware platform, calls the health status dataset in the historical database, and constructs the working condition-performance correlation model through the Gaussian process regression algorithm, such as the benchmark pressure difference function and the benchmark efficiency function. The model parameters are stored after being calibrated by regression of historical data.

[0053] The feature extraction module 4 is connected to the model building module 3 to analyze the output of the working condition-performance correlation model in real time, calculate the degradation feature parameters from the preprocessed data, including normalized pressure difference, equivalent efficiency and its derivative degradation rate, and form a feature vector to push to the time series prediction module 5.

[0054] The time series prediction module 5 is embedded in an edge computing device, such as an industrial-grade minicomputer, loaded with a long short-term memory neural network engine, inputs feature vectors to extrapolate future trend curves and reverse-calculates the remaining effective running time, and outputs the prediction results to the early warning triggering and instruction generation module 6.

[0055] As a functional extension of the distributed control system, the early warning triggering and instruction generation module 6 triggers a three-level early warning signal based on the remaining effective running time threshold, generates corresponding performance optimization instructions, such as the fan speed adjustment coefficient or the desulfurizer dosage correction value, and sends them through the industrial protocol interface.

[0056] The performance control module 7 integrates programmable logic controller hardware, drives actuators such as induced draft fan frequency converters and metering pumps through analog output cards, constrains and adjusts the rate to prevent system jitter when executing performance optimization instructions, and adjusts performance data in real time.

[0057] The model optimization module 8 receives the adjusted real-time performance data from the central server, calculates the differential pressure residual and efficiency residual, and uses optimization algorithms to dynamically update the parameters of the operating condition-performance correlation model and the preset performance failure threshold.

[0058] See attached document Figure 3-4 The following is a description of another embodiment of the method and system provided in this implementation.

[0059] Based on the method and system for predictive maintenance of ferroalloy exhaust gas purification described in Example 1, this Example 2 adds some specific implementation methods.

[0060] See attached document Figure 3 In this embodiment, the benchmark pressure difference model and benchmark efficiency model under healthy conditions are constructed using a collaborative modeling framework based on federated graph neural networks. Specifically, this includes constructing a global graph structure using heterogeneous historical operating condition data and historical performance data from a distributed device cluster, dynamically aggregating feature representations of similar operating condition samples using a graph attention mechanism, and embedding specific physical constraints of the purification unit as regularization terms to generate the benchmark pressure difference model and benchmark efficiency model.

[0061] Specifically, the collaborative modeling framework of federated graph neural networks also includes capturing the cross-time dependencies of distributed device clusters by constructing a spatiotemporal graph convolutional network, designing a causal inference module to analyze the causal relationship between operating condition evolution and deterioration feature parameters, and introducing a lightweight federated distillation strategy to compress the global graph structure.

[0062] The distributed equipment cluster refers to heterogeneous historical operating data uploaded by edge computing nodes in multiple ferroalloy plants, such as exhaust gas temperature T, flow rate Q, and inlet pollutant concentration. Historical performance data is used. A global graph structure is constructed on the central server, where each purification unit serves as a graph node. The node feature vector includes the equipment model (string identifier), cumulative runtime, and statistical parameters of operating conditions, namely the mean μ and variance σ². The edge weights between nodes are determined by calculating the similarity of operating conditions, such as the Euclidean distance between normalized temperature differences and flow rate differences. A graph attention mechanism is used to dynamically aggregate the feature representations of nodes with similar operating conditions: a multi-head attention layer is designed to calculate the association weights between the target node and its neighboring nodes, with the weight coefficient formula as follows: ,in denoted as the attention weight of node i to j, which is a dimensionless scalar; softmax is the normalized exponential function; LeakyReLU is the linear unit activation function with leakage correction, and its leakage slope is 0.01; a is the trainable attention vector, whose dimension matches its corresponding feature; T denotes the vector transpose; This represents the concatenation of features after a linear transformation, where W is a trainable weight matrix. Let i be the feature vector of node i / j. This is the concatenation operator; the aggregated feature vector is used to characterize the common patterns of the devices. ,in The aggregated output feature vector; ∑ is the activation function, such as Sigmoid; ∑ is the summation operation; j iterates through neighboring nodes. Specific physical constraints embedded in the cleanup unit are used as regularization terms, and a physical consistency penalty term based on the filtering dynamics equation, such as the Kalman-Korzny equation, is added to the loss function. ,in This is a physical loss item; For model prediction of pressure difference Pascal, is the filter media characteristic coefficient, a dimensionless constant determined by the material porosity; The gas dynamic viscosity is Pascal-second; The dust concentration is expressed in grams per cubic meter. The chemical reaction equilibrium constraint term in its filtration kinetic equation is also relevant. ,in For chemical loss items, For model prediction efficiency percentage, Let T be the efficiency function based on reaction kinetics, and T be the temperature in °C. This represents the volume percentage of oxygen content. A spatiotemporal graph convolutional network is used to capture cross-time-period dependencies. One-dimensional convolutional layers are stacked in the time dimension to extract historical sequences, such as the periodic patterns of a 72-hour operating window. The 3×1 kernel size indicates a time step of 3. In the spatial dimension, graph convolution operations are used to fuse the synchronous state evolution features of adjacent nodes, i.e., equipment with similar operating conditions. A causal inference module is designed based on a dual machine learning framework. A residual neural network is used to fit the conditional expectation of operating parameters with respect to degradation features, and the counterfactual causal effect size is calculated. ,in Y represents the causal effect quantity; Y represents the degradation characteristic parameter, such as the normalized pressure difference. For intervention operations, X is a condition variable, such as inlet concentration, and x is a set value; As a baseline intervention, a lightweight federated distillation strategy is introduced to compress the model. Local baseline models (student models) are trained on the devices, while a central server aggregates knowledge from across nodes to generate a global teacher model, using KL divergence loss. Key characteristics of distillation, among which This is due to distillation losses; Output the distribution for the teacher model; Output distribution for student models; ultimately generate a benchmark differential pressure model that can be deployed to edge devices. and benchmark efficiency model ,in , This is a graph neural network function.

[0063] See attached document Figure 4 In this embodiment, the time series prediction algorithm adopts a dual-channel LSTM network architecture based on wavelet packet decomposition. The construction of the dual-channel LSTM network architecture specifically includes decomposing the degradation feature parameter sequence into a first component and a second component using the wavelet packet decomposition algorithm and constructing a dual-channel LSTM network. The first channel performs trend fitting on the first component representing the degradation of the purification unit, and the second channel models abnormal fluctuations on the second component representing sudden disturbances. An adaptive adjustment module is added before the output layer of the dual-channel LSTM network to dynamically adjust the weight coefficients of the two channels. A variable step size rolling prediction mechanism is adopted, which automatically shortens the prediction window and triggers online parameter updates of the dual-channel LSTM network when the prediction deviation exceeds a set threshold.

[0064] Specifically, the adaptive adjustment module adopts a policy optimization mechanism based on the meta-learning framework. Specifically, it constructs a meta-knowledge base and designs a meta-optimizer to dynamically generate the weight coefficient allocation strategy for the dual channels, and introduces an online knowledge distillation mechanism to compress the complexity of the dual-channel LSTM network.

[0065] Among them, the wavelet packet decomposition algorithm refers to decomposing a sequence of deterioration characteristic parameters, such as normalized pressure difference or equivalent efficiency, into a low-frequency first component representing the long-term deterioration trend of the purification unit and a high-frequency second component representing sudden disturbances, such as operating condition fluctuations, through multi-scale analysis. The dual-channel LSTM network refers to an architecture composed of two independent long short-term memory neural networks. The first channel inputs the first component for trend fitting, such as using multi-layer LSTM units to capture long-term dependencies, while the second channel inputs the second component for abnormal fluctuation modeling, such as learning short-term mutations through a gating mechanism. The adaptive adjustment module refers to a dynamic weight optimization component located before the dual-channel output layer, used to adjust the weight coefficients of the dual-channel output fusion in real time. The variable step-size rolling prediction mechanism refers to a strategy that dynamically adjusts the prediction window length based on prediction deviation and triggers updates to the dual-channel LSTM network.

[0066] In practice, the wavelet packet decomposition algorithm is first initialized: the db4 wavelet basis function is selected, and the decomposition formula is as follows: ,in To degrade the feature parameter sequence, These are wavelet coefficients. Let be the wavelet packet basis function, and n be the scale index, representing different frequency components. After decomposition, the low-frequency component is taken as the first component, and the high-frequency component as the second component. A dual-channel LSTM network is constructed, with each channel containing three LSTM layers and a hidden layer dimension of 128. The training objective of the first channel is to minimize the trend fitting error. ,in The loss function used to train the first channel LSTM network is the trend fitting error loss; N is the number of samples, i.e., the number of points in the time series. These are the actual values, the actual measured values ​​of the degradation characteristic parameters; The predicted value is the output of the first channel LSTM network; the second channel uses the anomaly detection loss function. ,in This is the anomaly detection loss function, used to train the second-channel LSTM network; The probability density function, i.e., given the model parameters The probability of the actual value being below; These are the parameters for the second-channel LSTM network. An adaptive adjustment module is added before the output layer: this module, based on a meta-learning framework, constructs a meta-knowledge base to store historical prediction biases and optimal weight allocations, such as weight coefficients. and satisfy Design a meta-optimizer, such as a policy gradient-based reinforcement learning network, to dynamically generate a weight allocation policy, as shown in the formula: ,in The weighting coefficients for the first channel output, ranging from 0 to 1. The sigmoid function is an activation function that maps the input to a value between 0 and 1. The weight matrix is ​​the same as the training parameter matrix. These are the hidden states or feature vectors from a two-channel LSTM network. The bias vectors are trainable. The network complexity is compressed using an online knowledge distillation mechanism to train a lightweight student network, such as a single-layer LSTM, mimicking the teacher network output. The loss function is... ,in Here is the distillation loss function used to train the student network; For Kullback-Leibler divergence, The output probability of the global teacher network. (where λ is the output probability of the lightweight student network and λ is a hyperparameter balancing KL divergence and L2 loss). The predicted values ​​are for a dual-channel LSTM network. The predicted values ​​are for a lightweight student network. Finally, a variable-step rolling forecast is implemented, with an initial forecast window set to 24 hours, and the forecast bias is calculated. ,in The Euclidean distance between the actual and predicted values. For the actual measurement sequence of the deterioration characteristic parameters, The output sequence of a dual-channel LSTM network, if The threshold, where the threshold is a preset deviation threshold, such as 1.5 times the historical average deviation, will automatically shorten the window to 12 hours and trigger the online parameter update of the dual-channel LSTM network. For example, the Adam optimizer can be used to fine-tune the weights to form a closed-loop adaptive mechanism, which supports the remaining life prediction accuracy of the ferroalloy exhaust gas purification system under high temperature and high dust fluctuation conditions.

[0067] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," "third," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. The terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "comprising" or "including" and similar terms mean that the elements or objects preceding "comprising" or "including" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. "Above," "below," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0068] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method of ferrous alloy tail gas clean-up predictive maintenance, characterized by: This includes S1, acquiring and preprocessing real-time operating condition data and real-time performance data of the ferroalloy exhaust gas purification system to obtain preprocessed data; S2. Establish a condition-performance correlation model based on historical operating condition data and historical performance data in a preset historical database, and extract deterioration characteristic parameters reflecting changes in the health status of the purification unit from the preprocessed data based on the condition-performance correlation model; the purification unit is a device for removing particulate matter or gaseous pollutants from ferroalloy exhaust gas. S3. Based on the aforementioned degradation characteristic parameters, use a time series prediction algorithm to predict the remaining effective operating time of the purification unit in reverse. S4. Trigger a tiered warning based on the remaining effective running time, and output a performance optimization instruction corresponding to the tiered warning; S5. Adjust the real-time performance data based on the performance optimization instructions, and optimize the working condition-performance correlation model based on the adjusted real-time performance data.

2. A method of predictive maintenance of ferroalloy off-gas purification according to claim 1, characterized in that: S2 specifically includes constructing a baseline pressure difference model and a baseline efficiency model under healthy conditions based on historical operating condition data and historical performance data in a preset historical database; inputting the preprocessed real-time operating condition data into the baseline pressure difference model and the baseline efficiency model respectively to independently calculate the theoretical pressure difference and the theoretical purification efficiency; the theoretical pressure difference and the theoretical purification efficiency are respectively combined with the actual pressure difference and the actual purification efficiency in the preprocessed real-time performance data to calculate the degradation characteristic parameters.

3. A method of predictive maintenance of ferroalloy off-gas cleaning according to claim 2, characterized in that: The theoretical pressure difference and the theoretical purification efficiency correspond to the actual pressure difference and the actual purification efficiency in the pre-processed real-time performance data, respectively. The deterioration characteristic parameters are calculated by calculating the ratio of the actual pressure difference to the theoretical pressure difference to obtain the normalized pressure difference. The equivalent efficiency is obtained by calculating the ratio of the actual purification efficiency to the theoretical purification efficiency; the degradation rate is obtained by calculating the derivative of the normalized pressure difference and the equivalent efficiency; the degradation characteristic parameters are obtained by combining the normalized pressure difference, the equivalent efficiency, and the degradation rate.

4. The method for predictive maintenance of ferroalloy exhaust gas purification according to claim 1, characterized in that: S3 specifically includes extrapolating the trend curve of the degradation characteristic parameter within a future preset time period using a time series prediction algorithm; and calculating the time when the trend curve reaches the preset performance failure threshold based on a preset performance failure threshold to obtain the remaining effective running time.

5. The method for predictive maintenance of ferroalloy exhaust gas purification according to claim 2, characterized in that: S5 specifically includes adjusting the real-time performance data according to the performance optimization instruction; calculating the pressure difference residual between the actual pressure difference and the theoretical pressure difference, and the purification efficiency residual between the actual purification efficiency and the theoretical purification efficiency; and optimizing the operating condition-performance correlation model using the pressure difference residual and the purification efficiency residual.

6. The method for predictive maintenance of ferroalloy exhaust gas purification according to claim 2, characterized in that: The construction of the benchmark pressure difference model and benchmark efficiency model under healthy conditions adopts a collaborative modeling framework based on federated graph neural networks. Specifically, it includes constructing a global graph structure through heterogeneous historical operating condition data and historical performance data of distributed device clusters, dynamically aggregating feature representations of similar operating condition samples using graph attention mechanism, and embedding the specific physical constraints of the purification unit as regularization terms to generate the benchmark pressure difference model and benchmark efficiency model.

7. The method for predictive maintenance of ferroalloy exhaust gas purification according to claim 6, characterized in that: The collaborative modeling framework of the federated graph neural network also includes capturing the cross-time-period dependencies of the distributed device cluster by constructing a spatiotemporal graph convolutional network, designing a causal inference module to analyze the causal relationship between operating condition evolution and degradation feature parameters, and introducing a lightweight federated distillation strategy to compress the global graph structure.

8. The method for predictive maintenance of ferroalloy exhaust gas purification according to claim 4, characterized in that: The time series prediction algorithm employs a dual-channel LSTM network architecture based on wavelet packet decomposition. Constructing this dual-channel LSTM network architecture specifically involves decomposing the degradation feature parameter sequence into a first component and a second component using the wavelet packet decomposition algorithm and then building the dual-channel LSTM network. The first channel performs trend fitting on the first component representing the degradation of the purification unit, while the second channel models abnormal fluctuations in the second component representing sudden disturbances. An adaptive adjustment module is added before the output layer of the dual-channel LSTM network to dynamically adjust the weight coefficients of the two channels. A variable step-size rolling prediction mechanism is also employed, automatically shortening the prediction window and triggering online parameter updates for the dual-channel LSTM network when the prediction deviation exceeds a set threshold.

9. The method for predictive maintenance of ferroalloy exhaust gas purification according to claim 8, characterized in that: The adaptive adjustment module adopts a policy optimization mechanism based on a meta-learning framework. Specifically, it dynamically generates the weight coefficient allocation strategy of the dual-channel network by constructing a meta-knowledge base and designing a meta-optimizer, and introduces an online knowledge distillation mechanism to compress the complexity of the dual-channel LSTM network.

10. A predictive maintenance system for ferroalloy exhaust gas purification, characterized in that, It includes a data acquisition module (1), a preprocessing module (2), a model building module (3), a feature extraction module (4), a time series prediction module (5), an early warning triggering and instruction generation module (6), a performance control module (7), and a model optimization module (8); The data acquisition module (1) is deployed at the operation site of the purification unit and is used to collect real-time operating data and real-time performance data of the ferroalloy exhaust gas purification system. The preprocessing module (2) is connected to the data acquisition module (1) via a wired network and is used to preprocess the real-time operating condition data and the real-time performance data to obtain preprocessed data; The model building module (3) is deployed on the central server and builds a working condition-performance correlation model based on the historical working condition data and historical performance data stored in the historical database. The feature extraction module (4) is communicatively connected to the model building module (3), and extracts degradation feature parameters from the preprocessed data by parsing the working condition-performance correlation model. The time series prediction module (5) is integrated in the edge computing device and predicts the remaining effective running time based on the degradation feature parameters; The early warning triggering and instruction generation module (6) is communicatively connected to the time series prediction module (5), triggers different levels of early warning signals according to the remaining effective running time, and generates performance optimization instructions that match the different levels of early warning signals; The performance control module (7) is embedded in the purification unit controller and is used to execute the performance optimization instructions to adjust the real-time performance data. The model optimization module (8) is deployed on the central server, receives the adjusted real-time performance data fed back by the performance control module (7), and optimizes the working condition-performance correlation model.

Images