Fault diagnosis method for aluminum electrolysis cell digital twin model based on deep learning optimization
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGXI ACAD OF SCI
- Filing Date
- 2026-03-03
- Publication Date
- 2026-05-29
Smart Images
Figure CN121765655B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aluminum electrolysis technology, and in particular to a fault diagnosis method for a digital twin model of an aluminum electrolysis cell based on deep learning optimization. Background Technology
[0002] Aluminum electrolytic cells are core equipment in the modern aluminum smelting industry. Their operational status directly affects production efficiency, energy consumption, production safety, and equipment lifespan. During production, complex electrochemical and physical processes occur within the electrolytic cell under extreme high-temperature conditions. Their operating status is influenced by a combination of factors, including alumina concentration distribution, anodic effect, bottom sedimentation, and cathode damage. Timely and accurate diagnosis and early warning of these potential faults and abnormal states are crucial for achieving stable and efficient operation of the electrolytic cell, reducing energy consumption and production costs, and ensuring safe production.
[0003] Traditional electrolytic cell fault diagnosis relies heavily on operator experience, judging faults based on threshold alarms for limited parameters such as cell voltage and current series. This method is highly subjective, insensitive to early-stage and slowly changing faults, and unable to distinguish the specific type and root cause of the fault. With the development of automation and information technology, data-driven condition monitoring and fault diagnosis methods have begun to be applied to electrolytic cells. For example, some methods extract features through time-frequency analysis of current or voltage signals and use statistical process control or simple machine learning models for anomaly detection. However, these methods often suffer from insufficient feature extraction and weak model generalization ability when faced with the strong non-stationarity, nonlinearity, and strong noise interference of electrolytic cell signals, making it difficult to accurately separate and identify multiple concurrent or similar faults caused by different physical mechanisms. Summary of the Invention
[0004] This invention provides a fault diagnosis method for a digital twin model of an aluminum electrolytic cell based on deep learning optimization. It can monitor the operating status of the aluminum electrolytic cell, identify various process faults and equipment safety risks, distinguish between real faults and pseudo-effects caused by external interference, and improve the stability of the electrolytic cell operation.
[0005] The first aspect of this invention provides a fault diagnosis method for a digital twin model of an aluminum electrolytic cell based on deep learning optimization, comprising the following steps:
[0006] Acquire the current data of the aluminum electrolytic cell and preprocess the current data of the aluminum electrolytic cell;
[0007] The aluminum electrolytic cell is divided into three regions: a first region based on the feeder; a second region based on the physical layout of the anode; and a third region based on the cathode structure.
[0008] Based on the division results of the first region, the current data is decomposed to construct anode feature vectors and cathode feature vectors;
[0009] An anode fault diagnosis model is set up based on the division results of the second region; a cathode safety detection model is set up based on the division results of the third region; prior knowledge is constructed for the cathode safety detection model through Latin hypercube design and physical information neural network.
[0010] The anode feature vector is input into the anode fault diagnosis model to diagnose the fault status of the aluminum electrolytic cell; the cathode feature vector is input into the cathode safety detection model to detect the safety risk status of the aluminum electrolytic cell.
[0011] Based on the division results of the first region, set the region status information; based on the region status information, determine whether the regional changes in current data belong to false fault characteristics; when the regional changes in current data belong to false fault characteristics, determine that the current fault diagnosis warning is a false effect, and filter the fault diagnosis warning.
[0012] Furthermore, the step of dividing the aluminum electrolytic cell into a first region according to the feeder includes:
[0013] Based on the position and number of feeders above the electrolytic cell, the electrolytic cell is divided into control areas corresponding to the feeders;
[0014] The equivalent resistance of the control area corresponding to the feeder is calculated based on the acquired current data, and the area corresponding to the equivalent resistance is used as the area control unit.
[0015] Furthermore, the second region division based on the physical layout of the anodes of the aluminum electrolytic cell includes dividing all anode carbon blocks and corresponding guide rods into multiple anode regions according to their physical positions based on the arrangement of the anode groups of the aluminum electrolytic cell, with each anode region associated with a region control unit corresponding to the equivalent resistance.
[0016] The third region division based on the cathode structure of the aluminum electrolysis cell includes dividing the cathode into multiple cathode regions based on the physical separation of the cathode blocks, with each cathode region associated with a region control unit corresponding to the equivalent resistance.
[0017] Furthermore, the step of decomposing the current data based on the division results of the first region to construct anode feature vectors and cathode feature vectors includes the following steps:
[0018] Obtain the current data corresponding to each area control unit;
[0019] An improved empirical mode decomposition is performed on the current data of each regional control unit to obtain multiple intrinsic mode function components; and time-domain features, frequency-domain features, and time-frequency-domain features of the current data are extracted from the intrinsic mode function components.
[0020] The time-domain, frequency-domain, and time-frequency-domain features of the corresponding area control unit in the anode region are fused to generate the anode feature vector; the time-domain, frequency-domain, and time-frequency-domain features of the corresponding area control unit in the cathode region are fused to generate the cathode feature vector.
[0021] Furthermore, the construction of prior knowledge for the cathode safety detection model through Latin hypercube design and physical information neural network includes:
[0022] Set geometric and location parameters to characterize cathode faults, and use these parameters as design factors.
[0023] The fault parameter combination is generated by sampling within the space of the design factor using the Latin hypercube design method.
[0024] The multiphysics simulation model of the electrolytic cell was used to simulate each set of fault parameters, obtain the corresponding cathode current response data, and construct a simulation dataset.
[0025] A physical information neural network model is trained using a simulation dataset; the prior knowledge of the trained physical information neural network model is then transferred to the cathode safety detection model.
[0026] Furthermore, the step of inputting the anode feature vector into the anode fault diagnosis model to diagnose the fault state of the aluminum electrolytic cell includes:
[0027] The gradient boosting decision tree classifier is used to identify the fault type of the anode feature vector and calculate the predicted probability of the fault.
[0028] Based on the identified fault type and the calculated predicted probability, the location is correlated with the corresponding anode area to generate a fault warning.
[0029] Furthermore, the step of inputting the cathode feature vector into the cathode safety detection model to detect the safety risk status of the aluminum electrolysis cell includes:
[0030] The cathode feature vector is input into the cathode safety detection model that has transferred prior knowledge, and the output is a health score that represents the health status of the cathode.
[0031] Based on the current density uniformity, historical trend and health score of the cathode region, assess the risk level of the cathode anomaly corresponding to the cathode region.
[0032] When the risk level exceeds the preset safety threshold, a safety warning is triggered, and a fault warning is generated.
[0033] Furthermore, the region status information includes:
[0034] Real-time estimated value and rate of change of the equivalent resistance corresponding to the area control unit;
[0035] The operating status information of the corresponding feeder for the area control unit;
[0036] The average anode current density and cathode current density of the area corresponding to the area control unit;
[0037] The balance coefficient between the current in the area control unit and the current in the adjacent areas.
[0038] Furthermore, the pseudo-fault characteristics include:
[0039] The first false fault characteristic is that the number of control units in the area where the current abnormality occurs is less than the first false fault characteristic judgment threshold.
[0040] The second false fault characteristic is that the duration of the current abnormality is less than the second false fault characteristic judgment threshold; after the current abnormality ends, the current returns to the normal fluctuation baseline.
[0041] The third false fault characteristic is that the current abnormality occurs in the control unit of the area where the current exceeds the third false fault characteristic judgment threshold.
[0042] Furthermore, determining whether regional changes in current data constitute false fault characteristics based on regional status information includes the following steps:
[0043] When a fault warning is triggered, the status information of the control unit in the area affected by the warning and the event logs of the adjacent control units are obtained.
[0044] Determine whether the current anomaly meets the first, second, and third pseudo-fault characteristics;
[0045] If the conditions are met, it is determined to be a spurious effect and a warning filter is applied; if the conditions are not met, it is determined to be a real fault warning.
[0046] As can be seen from the above technical solutions, the present invention has the following advantages:
[0047] This invention achieves refined fault location and early identification by integrating physical spatial mapping and signal feature decomposition. Specifically, it divides the electrolytic cell into logical regions using a three-tiered physical structure consisting of the feeder, anode group, and cathode block, transforming the complex spatial distribution system into independently analyzable computational units. Based on this, an improved empirical mode decomposition method is used to process the regional current data, adaptively separating the intrinsic mode function components characterizing different physical processes. Multi-domain features are extracted from these components, constructing feature vectors that strictly correspond to specific anode and cathode regions. This allows the system to accurately correlate global monitoring signals to local physical components, overcoming the fuzzy fault location shortcomings of traditional methods and achieving sensitive capture and spatial location of subtle early-stage fault features.
[0048] Secondly, this invention achieves comprehensive and interpretable state assessment and risk warning by constructing a dual-model diagnostic system specifically for both anode and cathode and combining it with prior physical knowledge. Addressing the frequent but relatively explicit mechanisms of anode system failures, a gradient boosting decision tree model is employed for rapid fault classification and probabilistic output. For the challenge of direct observation and scarce fault samples in cathode systems, this invention innovatively generates simulation data through Latin hypercube design and multiphysics simulation, and utilizes a physical information neural network to learn the physical laws within the data, transferring prior knowledge to the detection model. This enables the system to rapidly diagnose common anode faults and achieve health scoring and risk classification for latent cathode faults.
[0049] Finally, by integrating multi-dimensional regional status information and establishing false fault feature filtering rules, highly reliable early warnings and effective suppression of false alarms are achieved. This is accomplished by synchronously calculating dynamic status information such as the equivalent resistance change rate, feeder status, current density, and balance of each region. When an initial warning is triggered, the system can determine, based on contextual information, whether the anomaly conforms to false fault characteristics such as local isolation, instantaneous outbreak, or global synchronization. Through the filtering mechanism, false signals caused by external interference such as the neighboring tank effect and power grid fluctuations can be effectively identified, thus filtering out a large number of false alarms before final reporting. This improves the accuracy and reliability of alarm information, allowing maintenance personnel to focus on handling real faults and avoiding ineffective inspections and production interventions caused by false alarms, achieving the maintenance goal of cost reduction and efficiency improvement.
[0050] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from an examination of the following, or may be learned from the practice of the invention. Attached Figure Description
[0051] Figure 1 The method flowchart provided by the present invention. Detailed Implementation
[0052] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0053] Example 1
[0054] The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal; no specific limitation is made. The method in this application will be described below from the perspective of system implementation. Figure 1 The method for fault diagnosis of aluminum electrolytic cells based on a deep learning-optimized digital twin model includes the following steps:
[0055] Acquire the current data of the aluminum electrolytic cell and preprocess the current data of the aluminum electrolytic cell;
[0056] Specifically, the current data of the aluminum electrolytic cell is acquired by arranging non-contact fiber optic sensors within the cell. The preprocessing of the current data includes using a low-pass filter to remove high-frequency noise and normalizing the data to eliminate dimensional interference.
[0057] The aluminum electrolytic cell is divided into three regions: a first region based on the feeder; a second region based on the physical layout of the anode; and a third region based on the cathode structure.
[0058] Since an electrolytic cell is a spatially distributed system, faults often exhibit locality. A three-level partitioning method maps the physical space into computable logical units.
[0059] The method of dividing the aluminum electrolytic cell into a first region according to the feeder includes:
[0060] Based on the position and number of feeders above the electrolytic cell, the electrolytic cell is divided into control areas corresponding to the feeders;
[0061] The equivalent resistance of the control area corresponding to the feeder is calculated based on the acquired current data, and the area corresponding to the equivalent resistance is used as the area control unit.
[0062] Based on the actual location and number of feeders installed above the electrolytic cell, the entire electrolytic cell is divided into several independent control areas, so that each control area basically corresponds to the material coverage range of a feeder.
[0063] After spatial division, an equivalent resistance is estimated for the region based on the acquired current data and calculated using Kirchhoff's current law. This equivalent resistance reflects production data such as the electrolyte conductivity and alumina concentration within the controlled region.
[0064] The second region division based on the physical layout of the anodes of the aluminum electrolytic cell includes dividing all anode carbon blocks and corresponding guide rods into multiple anode regions according to their physical positions based on the arrangement of the anode groups of the aluminum electrolytic cell. Each anode region is associated with a region control unit corresponding to the equivalent resistance.
[0065] Based on the physical arrangement of the anode carbon block groups and their guide rods within the electrolytic cell, the entire anode system is divided into several continuous anode regions. After division, each anode region is associated with a corresponding or spatially corresponding region control unit based on the positional relationship between the anode regions and the feeder.
[0066] The third region division based on the cathode structure of the aluminum electrolysis cell includes dividing the cathode into multiple cathode regions based on the physical separation of the cathode blocks, with each cathode region associated with a region control unit corresponding to the equivalent resistance.
[0067] Based on the physical separation of the cathode carbon blocks or steel rod assemblies, the cathode system is divided into multiple cathode regions. Each cathode region is associated with a corresponding set of region control units located above it. By establishing this association, abnormal changes in the monitored current and equivalent resistance can be inferred to reflect the health status of the corresponding cathode region below, such as whether local damage or changes in the current path have occurred.
[0068] Based on the division results of the first region, the current data is decomposed to construct anode feature vectors and cathode feature vectors;
[0069] Specifically, it includes the following steps:
[0070] Obtain the current data corresponding to each area control unit;
[0071] An improved empirical mode decomposition is performed on the current data of each regional control unit to obtain multiple intrinsic mode function components; and time-domain features, frequency-domain features, and time-frequency-domain features of the current data are extracted from the intrinsic mode function components.
[0072] The time-domain, frequency-domain, and time-frequency-domain features of the corresponding area control unit in the anode region are fused to generate the anode feature vector; the time-domain, frequency-domain, and time-frequency-domain features of the corresponding area control unit in the cathode region are fused to generate the cathode feature vector.
[0073] Because the current signal within the electrolytic cell is affected by physical processes such as feeding and aluminum molten fluid fluctuations, it exhibits non-stationary and nonlinear characteristics. Therefore, the system employs an improved empirical mode decomposition method to process the current data of each unit. EMD (Empirical Mode Decomposition) adaptively decomposes complex signals into a set of intrinsic mode function components (IMFs) arranged from high to low frequencies. Each IMF can be considered an intrinsic oscillation mode, which may physically correspond to different processes; for example, high-frequency components correspond to bubble noise, while low-frequency components correspond to aluminum molten fluid flow. During feature extraction, time-domain features, frequency-domain features, and time-frequency-domain features are extracted from each IMF component obtained through decomposition.
[0074] Next, all features extracted from control units belonging to the same anode physical region are concatenated and normalized to form a comprehensive anode feature vector. This anode feature vector is used to represent the operating state of the corresponding anode region from an electrical perspective. Similarly, features from all control units belonging to the same cathode physical region are fused to generate a cathode feature vector, which is used to characterize the health status of that cathode region.
[0075] An anode fault diagnosis model is set up based on the division results of the second region; a cathode safety detection model is set up based on the division results of the third region; prior knowledge is constructed for the cathode safety detection model through Latin hypercube design and physical information neural network.
[0076] Specifically, the anode fault diagnosis model is constructed based on the second region segmentation results, using a gradient boosting decision tree as the core classifier and taking the anode feature vectors generated for each anode region as input. The anode feature vectors integrate the time-domain, frequency-domain, and time-frequency-domain features of the control unit current signal for the corresponding region. After training with historical fault data and simulation data, the model can output the fault type of the current anode region, such as uneven anode current distribution, anode detachment risk, or localized overheating, while also providing the corresponding predicted probability.
[0077] The cathode safety detection model is set based on the third region division results and adopts a deep neural network architecture. The input is the cathode feature vector corresponding to the cathode region, and the output is a health score and a comprehensive risk level assessment. To address the scarcity of real cathode fault data, the model uses Latin hypercube design to spatially sample the geometric and location parameters of cathode faults, combines multiphysics simulation to generate diverse fault response data, and uses this data to train the physical information neural network.
[0078] Specifically, the construction of prior knowledge for the cathode safety detection model through Latin hypercube design and physical information neural network includes:
[0079] Set geometric and location parameters to characterize cathode faults, and use these parameters as design factors.
[0080] The fault parameter combination is generated by sampling within the space of the design factor using the Latin hypercube design method.
[0081] The multiphysics simulation model of the electrolytic cell was used to simulate each set of fault parameters, obtain the corresponding cathode current response data, and construct a simulation dataset.
[0082] A physical information neural network model is trained using a simulation dataset; the prior knowledge of the trained physical information neural network model is then transferred to the cathode safety detection model.
[0083] First, the system parametrically characterizes cathode faults, abstracting physical damage into controllable geometric and location parameters; for example, the length and depth of cracks and erosions, and the location of the cracks and erosions within which cathode carbon blocks they occur. These parameters are defined as design factors to construct a multidimensional fault parameter space.
[0084] Next, the Latin hypercube design method is used for sampling within this parameter space. As a stratified sampling technique, the Latin hypercube design method ensures that all design factors and their combinations are uniformly explored with a limited number of samplings, thereby generating multiple sets of fault parameter combinations with spatial coverage and no correlation. Then, each set of fault parameters is input into the multiphysics simulation model of the electrolyzer. By running the simulation, the cathode current response data at each cathode region or measurement point under the corresponding fault state are extracted, generating a simulation dataset where each data sample corresponds to a set of known fault parameters and the corresponding system response.
[0085] The simulation dataset is used to train a Physical Information Neural Network (PINN). This trained PINN model is then used as a source of prior knowledge, and through transfer learning, the prior knowledge is transferred to the actually deployed cathode safety detection model.
[0086] The anode feature vector is input into the anode fault diagnosis model to diagnose the fault status of the aluminum electrolytic cell; the cathode feature vector is input into the cathode safety detection model to detect the safety risk status of the aluminum electrolytic cell.
[0087] During operation, the generated anode and cathode feature vectors are input into the corresponding anode fault diagnosis model or cathode safety detection model, respectively. The anode fault diagnosis model performs rapid analysis on the input and outputs the diagnostic results for the anode region. The cathode safety detection model outputs the health score or risk level for each cathode region.
[0088] Specifically, the step of inputting the anode feature vector into the anode fault diagnosis model to diagnose the fault state of the aluminum electrolysis cell includes:
[0089] The gradient boosting decision tree classifier is used to identify the fault type of the anode feature vector and calculate the predicted probability of the fault.
[0090] Based on the identified fault type and the calculated predicted probability, the location is correlated with the corresponding anode area to generate a fault warning.
[0091] The anode fault diagnosis model is a trained Gradient Boosting Decision Tree (GBDT) classifier, comprising two levels of analysis and output. The first level is fault type identification and probability quantification. The anode fault diagnosis model infers from the input anode feature vector and outputs a fault type classification result. Examples include "severely uneven anode current distribution," "suspected early stage of anode detachment," "risk of local overheating," and "abnormal contact resistance of the conductor rod." Simultaneously, it also outputs the corresponding predicted probability to reflect the model's confidence level in its current judgment.
[0092] The second level is the spatial localization of the fault and the encapsulation of early warning information. Based on the regional mapping relationship corresponding to the source of the feature vector, the system automatically associates the identified fault type and probability with the specific anode physical area represented by that feature vector. Then, a structured fault warning is generated based on the fault type, predicted probability, and the associated anode area location.
[0093] Specifically, the step of inputting the cathode feature vector into the cathode safety detection model to detect the safety risk status of the aluminum electrolysis cell includes:
[0094] The cathode feature vector is input into the cathode safety detection model that has transferred prior knowledge, and the output is a health score that represents the health status of the cathode.
[0095] Based on the current density uniformity, historical trend and health score of the cathode region, assess the risk level of the cathode anomaly corresponding to the cathode region.
[0096] When the risk level exceeds the preset safety threshold, a safety warning is triggered, and a fault warning is generated.
[0097] The cathode safety detection model maps input features to a quantified output health score through forward propagation. The health score is a continuous value between 0 and 1, reflecting the health status of the corresponding cathode region in multiple physical structures, such as material integrity and thermal stress state. A higher score indicates a more ideal cathode condition in the aluminum electrolytic cell; a lower score indicates a worse condition.
[0098] The current density uniformity is quantified by calculating the ratio or standard deviation of the average current density of the cathode region to the average current density of all adjacent cathode regions. A significant deviation of the uniformity from the normal range is a typical spatial characteristic of current concentration or shunting caused by cathode damage or abnormal precipitation.
[0099] The historical trend analysis analyzes the changing trends of the cathode region by recalling its historical health score sequence over a past time window. For example, it calculates the moving average of the score, the downward slope, or detects whether there is a sustained decline over multiple consecutive periods. A steady downward trend is a more significant indicator of risk than a single isolated low score.
[0100] When assessing the risk level of cathode anomalies in a cathode region, spatial distribution, temporal evolution, and instantaneous state—namely, current density uniformity, historical trend, and health score—are weighted and comprehensively evaluated to output a graded risk level; for example, risk levels can be set as normal, low risk, medium risk, and high risk. The calculated risk level is then compared with a preset safety threshold. When the risk level reaches or exceeds the medium or high risk threshold, a clear and intervention-required safety hazard is identified in the cathode region, triggering a safety warning. The fault warning information can be structured, including the specific cathode region number, the determined risk level, and the decision-making basis, enabling personnel to quickly locate the risk.
[0101] Based on the division results of the first region, set the region status information; based on the region status information, determine whether the regional changes in current data belong to false fault characteristics; when the regional changes in current data belong to false fault characteristics, determine that the current fault diagnosis warning is a false effect, and filter the fault diagnosis warning.
[0102] In an aluminum electrolysis workshop, when an anodic effect occurs in one electrolytic cell, the strong magnetic field change may induce brief, fault-like abnormal fluctuations in the current data of adjacent normal cells—a spurious effect. If not distinguished, this can lead to numerous false alarms. Intelligent judgment can be made using area status information. For example, the system detects a sudden drop in current in a certain area and triggers an alarm, but the anomaly is limited to that area; or the anomaly lasts for a very short time. By combining comprehensive information, the system can determine whether the alarm is a spurious effect, thereby filtering it and avoiding misoperation.
[0103] Specifically, the area status information includes:
[0104] Real-time estimated value and rate of change of the equivalent resistance corresponding to the area control unit;
[0105] The operating status information of the corresponding feeder for the area control unit;
[0106] The average anode current density and cathode current density of the area corresponding to the area control unit;
[0107] The balance coefficient between the current in the area control unit and the current in the adjacent areas.
[0108] The real-time estimate of the equivalent resistance reflects the instantaneous state of the electrolyte conductivity, electrode spacing, and alumina concentration within the region. The rate of change reflects the stability of the region's state. The operating status information includes the feeder's running, stopped, and faulty states, as well as the feeder's feeding cycle. This operating status information is used to correlate abnormal electrical signals with operational events.
[0109] The average anodic and cathodic current densities of the region are quantitative indicators obtained by dividing the measured total current of the region by the effective conductive area of the corresponding electrode. Current density can eliminate the influence of differences in electrode size, making the states of different regions and the same region at different times comparable. For example, an abnormally high anodic current density may indicate a risk of local overheating, while non-uniform cathodic current density is directly related to cathodic damage or deposition problems.
[0110] The uniformity coefficient of the current in the region and the current in adjacent regions is calculated by comparing the current in this region with the average value of the current in several surrounding regions, and is used to quantify the uniformity of the current distribution in the spatial plane.
[0111] Specifically, the pseudo-fault characteristics include:
[0112] The first false fault characteristic is that the number of control units in the area where the current abnormality occurs is less than the first false fault characteristic judgment threshold.
[0113] The second false fault characteristic is that the duration of the current abnormality is less than the second false fault characteristic judgment threshold; after the current abnormality ends, the current returns to the normal fluctuation baseline.
[0114] The third false fault characteristic is that the current abnormality occurs in the control unit of the area where the current exceeds the third false fault characteristic judgment threshold.
[0115] The first pseudo-fault feature is used to identify the local isolation of anomalies. Physical faults within the aluminum electrolysis cell, such as uneven anode current distribution, localized overheating, or cathode damage, can propagate spatially through the electromagnetic field and thermal convection of the molten pool, causing related anomalies in multiple adjacent control units. Conversely, strong electromagnetic induction interference generated by the anode effect in neighboring cells typically has a very limited impact range, possibly inducing significant transient current fluctuations only in one or two control units closest to the interference source. Therefore, a low threshold for judging the first pseudo-fault feature can be set. When the number of detected abnormal regions falls below this threshold, the feature flag is triggered, indicating that the anomaly is more likely to originate from external localized interference. For example, the number of abnormal regions is less than 10% of the total number of regions.
[0116] The second pseudo-fault characteristic is used to capture the transient nature and recoverability of anomalies. When equipment or process failures occur, the resulting changes in current or resistance are usually continuous and gradual, and will not fully recover to the normal baseline on their own before manual intervention. In contrast, pseudo-effects, whether caused by external electromagnetic induction or instantaneous imbalance of materials within the tank, often result in abnormal signals that are sudden spikes or brief drops, lasting extremely short periods. Furthermore, after the interference factor disappears, the signal quickly and automatically returns to its normal fluctuation range. The short duration can be quantified by setting a threshold for judging the second pseudo-fault characteristic, and this can be combined with whether the signal smoothly recovers to the historical normal fluctuation baseline for judgment.
[0117] The third pseudo-fault feature is used to identify the global synchronicity of anomalies. When anomalies occur almost synchronously in all or most area control units, the source of the impact may be a common part of the entire electrolyzer or the external environment. For example, drastic fluctuations in series current, instantaneous disturbances in the rectifier system, or voltage flicker in the power grid can all cause this global anomaly. Since these situations are not caused by the deterioration of the health of a local component within the electrolyzer itself, they are also classified as pseudo-faults that need to be filtered. Such events can be captured by setting a very high threshold for judging the third pseudo-fault feature. For example, the number of anomalous areas exceeds 90% of the total number of areas.
[0118] The method for determining whether regional changes in current data constitute false fault characteristics based on regional status information includes the following steps:
[0119] When a fault warning is triggered, the status information of the control unit in the area affected by the warning and the event logs of the adjacent control units are obtained.
[0120] Determine whether the current anomaly meets the first, second, and third pseudo-fault characteristics;
[0121] If the conditions are met, it is determined to be a spurious effect and a warning filter is applied; if the conditions are not met, it is determined to be a real fault warning.
[0122] When the anode fault diagnosis model or the cathode safety detection model issues an initial warning, the real-time status information of the control unit and its adjacent units in the area indicated by the warning is obtained; then, the historical event logs of these related areas during the same period are queried to check whether there are any planned process operations or chain reactions of related equipment.
[0123] Next, the system determines whether the current anomaly meets the first, second, and third pseudo-fault characteristics. If the current warning event meets any one or more of these pseudo-fault characteristics, the system determines that the alarm is a false alarm and performs a filtering operation on the warning, storing it only in the log for analysis. If the current event does not meet any of the pseudo-fault characteristics, the system determines it is a real fault and issues a warning.
[0124] In summary, this invention achieves refined fault location and early identification by integrating physical space mapping and signal feature decomposition. Specifically, it uses a three-tiered physical structure—feeder, anode group, and cathode block—to logically divide the electrolytic cell into regions, transforming the complex spatial distribution system into independently analyzable computational units. Based on this, an improved empirical mode decomposition method is employed to process regional current data, adaptively separating intrinsic mode function components characterizing different physical processes. Multi-domain features are extracted from these components, constructing feature vectors that strictly correspond to specific anode and cathode regions. This allows the system to accurately correlate global monitoring signals to local physical components, overcoming the fuzzy fault location shortcomings of traditional methods and achieving sensitive capture and spatial localization of subtle early-stage fault features.
[0125] Secondly, this invention achieves comprehensive and interpretable state assessment and risk warning by constructing a dual-model diagnostic system specifically for both anode and cathode and combining it with prior physical knowledge. Addressing the frequent but relatively explicit mechanisms of anode system failures, a gradient boosting decision tree model is employed for rapid fault classification and probabilistic output. For the challenge of direct observation and scarce fault samples in cathode systems, this invention innovatively generates simulation data through Latin hypercube design and multiphysics simulation, and utilizes a physical information neural network to learn the physical laws within the data, transferring prior knowledge to the detection model. This enables the system to rapidly diagnose common anode faults and achieve health scoring and risk classification for latent cathode faults.
[0126] Finally, by integrating multi-dimensional regional status information and establishing false fault feature filtering rules, highly reliable early warnings and effective suppression of false alarms are achieved. This is accomplished by synchronously calculating dynamic status information such as the equivalent resistance change rate, feeder status, current density, and balance of each region. When an initial warning is triggered, the system can determine, based on contextual information, whether the anomaly conforms to false fault characteristics such as local isolation, instantaneous outbreak, or global synchronization. Through the filtering mechanism, false signals caused by external interference such as the neighboring tank effect and power grid fluctuations can be effectively identified, thus filtering out a large number of false alarms before final reporting. This improves the accuracy and reliability of alarm information, allowing maintenance personnel to focus on handling real faults and avoiding ineffective inspections and production interventions caused by false alarms, achieving the maintenance goal of cost reduction and efficiency improvement.
[0127] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.
[0128] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.
[0129] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0130] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.
[0131] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A fault diagnosis method for a digital twin model of an aluminum electrolytic cell based on deep learning optimization, characterized in that, Includes the following steps: Acquire the current data of the aluminum electrolytic cell and preprocess the current data of the aluminum electrolytic cell; The aluminum electrolytic cell is divided into three regions: a first region based on the feeder; a second region based on the physical layout of the anode; and a third region based on the cathode structure. Based on the division results of the first region, the current data is decomposed to construct anode feature vectors and cathode feature vectors; An anode fault diagnosis model is set up based on the division results of the second region; a cathode safety detection model is set up based on the division results of the third region; prior knowledge is constructed for the cathode safety detection model through Latin hypercube design and physical information neural network. The anode feature vector is input into the anode fault diagnosis model to diagnose the fault status of the aluminum electrolytic cell; the cathode feature vector is input into the cathode safety detection model to detect the safety risk status of the aluminum electrolytic cell. Based on the division results of the first region, set the region status information; based on the region status information, determine whether the regional changes in the current data belong to pseudo-fault characteristics; When regional variations in current data indicate a false fault characteristic, the current fault diagnosis warning is determined to be a false effect, and the fault diagnosis warning is filtered out; where: The step of decomposing the current data based on the division results of the first region to construct the anode feature vector and the cathode feature vector includes the following steps: Obtain the current data corresponding to each area control unit; An improved empirical mode decomposition is performed on the current data of each regional control unit to obtain multiple intrinsic mode function components; and time-domain features, frequency-domain features, and time-frequency-domain features of the current data are extracted from the intrinsic mode function components. The time-domain, frequency-domain, and time-frequency-domain features of the corresponding area control unit in the anode region are fused to generate the anode feature vector; the time-domain, frequency-domain, and time-frequency-domain features of the corresponding area control unit in the cathode region are fused to generate the cathode feature vector. The construction of prior knowledge for the cathode safety detection model through Latin hypercube design and physical information neural network includes: Set geometric and location parameters to characterize cathode faults, and use these parameters as design factors. The fault parameter combination is generated by sampling within the space of the design factor using the Latin hypercube design method. The multiphysics simulation model of the electrolytic cell was used to simulate each set of fault parameters, obtain the corresponding cathode current response data, and construct a simulation dataset. A physical information neural network model is trained using a simulation dataset; the prior knowledge of the trained physical information neural network model is then transferred to the cathode safety detection model.
2. The fault diagnosis method for a digital twin model of an aluminum electrolytic cell based on deep learning optimization according to claim 1, characterized in that, The step of dividing the aluminum electrolysis cell into a first region according to the feeder includes: Based on the position and number of feeders above the electrolytic cell, the electrolytic cell is divided into control areas corresponding to the feeders; The equivalent resistance of the control area corresponding to the feeder is calculated based on the acquired current data, and the area corresponding to the equivalent resistance is used as the area control unit.
3. The fault diagnosis method for a digital twin model of an aluminum electrolytic cell based on deep learning optimization according to claim 1, characterized in that, The second region division based on the physical layout of the anodes of the aluminum electrolytic cell includes dividing all anode carbon blocks and corresponding guide rods into multiple anode regions according to their physical positions based on the arrangement of the anode groups of the aluminum electrolytic cell. Each anode region is associated with a region control unit corresponding to the equivalent resistance. The third region division based on the cathode structure of the aluminum electrolysis cell includes dividing the cathode into multiple cathode regions based on the physical separation of the cathode blocks, with each cathode region associated with a region control unit corresponding to the equivalent resistance.
4. The fault diagnosis method for a digital twin model of an aluminum electrolytic cell based on deep learning optimization according to claim 1, characterized in that, The step of inputting the anode feature vector into the anode fault diagnosis model to diagnose the fault state of the aluminum electrolysis cell includes: The gradient boosting decision tree classifier is used to identify the fault type of the anode feature vector and calculate the predicted probability of the fault. Based on the identified fault type and the calculated predicted probability, the location is correlated with the corresponding anode area to generate a fault warning.
5. The fault diagnosis method for a digital twin model of an aluminum electrolytic cell based on deep learning optimization according to claim 1, characterized in that, The step of inputting the cathode feature vector into the cathode safety detection model to detect the safety risk status of the aluminum electrolysis cell includes: The cathode feature vector is input into the cathode safety detection model that has transferred prior knowledge, and the output is a health score that represents the health status of the cathode. Based on the current density uniformity, historical trend and health score of the cathode region, assess the risk level of the cathode anomaly corresponding to the cathode region. When the risk level exceeds the preset safety threshold, a safety warning is triggered, and a fault warning is generated.
6. The fault diagnosis method for a digital twin model of an aluminum electrolytic cell based on deep learning optimization according to claim 1, characterized in that, The regional status information includes: Real-time estimated value and rate of change of the equivalent resistance corresponding to the area control unit; The operating status information of the corresponding feeder for the area control unit; The average anode current density and cathode current density of the area corresponding to the area control unit; The balance coefficient between the current in the area control unit and the current in the adjacent areas.
7. The fault diagnosis method for a digital twin model of an aluminum electrolytic cell based on deep learning optimization according to claim 1, characterized in that, The pseudo-fault characteristics include: The first false fault characteristic is that the number of control units in the area where the current abnormality occurs is less than the first false fault characteristic judgment threshold. The second false fault characteristic is that the duration of the current abnormality is less than the second false fault characteristic judgment threshold; after the current abnormality ends, the current returns to the normal fluctuation baseline. The third false fault characteristic is that the current abnormality occurs in the control unit of the area where the current exceeds the third false fault characteristic judgment threshold.
8. The fault diagnosis method for a digital twin model of an aluminum electrolytic cell based on deep learning optimization according to claim 7, characterized in that, The step of determining whether regional changes in current data constitute false fault characteristics based on regional status information includes the following steps: When a fault warning is triggered, the status information of the control unit in the area affected by the warning and the event logs of the adjacent control units are obtained. Determine whether the current anomaly meets the first, second, and third pseudo-fault characteristics; If the conditions are met, it is determined to be a spurious effect and an early warning filter is applied; if the conditions are not met, it is determined to be a real fault warning.