A Fault Diagnosis Method for Unbalanced Data of Coal Mill Based on Bayesian Network

Abstract

A fault diagnosis method for unbalanced data of coal mills based on Bayesian networks, belonging to the field of generalized zero-sample fault diagnosis of feed water pump sets in thermal power generation. Aiming at the problems of uneven distribution of fault data, difficulty in diagnosing minority-class faults and poor interpretability of results in industrial scenarios, the invention synergistically optimizes the sample distribution through SMOTE data augmentation and Dirichlet prior smoothing, constructs a three-layer causal topology network of "fault layer - attribute layer - observation layer", combines expert knowledge constraints with data-driven learning to achieve high-precision fault classification, designs a dual-mode diagnosis mechanism, synchronously supports direct inference of fault nodes and indirect inference of attribute nodes, and provides interpretable physical-level diagnosis guidance while ensuring the accuracy of fault diagnosis. This method has been successfully applied to the coal mill system of coal-fired power units and can be extended to the transparent intelligent operation and maintenance of complex industrial equipment.

Description

Technical Field

[0001] The present invention belongs to the field of generalized zero-sample fault diagnosis of feedwater pump groups for thermal power generation, and provides a fault diagnosis method for unbalanced data of coal mills based on Bayesian networks. Background Art

[0002] The coal pulverizing system of a coal mill is an important auxiliary equipment in a coal-fired power plant. Its main function is to grind raw coal into pulverized coal that meets the combustion requirements and transport it to the boiler through primary air. The operating state of the coal pulverizing system directly affects the boiler combustion efficiency, the service life of the equipment, and the overall operating performance of the power plant. Once a coal mill fails, it may lead to unqualified pulverized coal fineness, reduced boiler combustion efficiency, and even cause serious equipment damage and economic losses. Therefore, developing an efficient and accurate fault diagnosis method for coal mills has important engineering significance for improving the safety, economy, and operating efficiency of power plants.

[0003] In actual industrial scenarios, fault diagnosis of coal mills faces a significant challenge, that is, the data distribution of different fault categories is unbalanced. Specifically, some fault categories occur relatively frequently, so the corresponding fault data is relatively sufficient; while some other fault categories have extremely scarce data due to low occurrence probability or limited acquisition conditions. This unbalanced data distribution makes the diagnostic model more inclined to learn the characteristics of fault categories with sufficient data during the training process, while ignoring the fault categories with scarce data, thus reducing the diagnostic accuracy of minority-class faults. In addition, multiple fault types in the coal pulverizing system may occur simultaneously or switch rapidly, further increasing the complexity of fault diagnosis.

[0004] Currently, methods such as deep learning, statistical analysis, and rule-based expert systems have been widely used in the field of fault diagnosis, but there are still the following limitations when dealing with the above problems: Deep learning relies on large-scale data for training and is good at extracting complex feature patterns from massive data. However, it has high requirements for the balance of data volume and class distribution. In the scenario of unbalanced fault category data, deep learning models often have the problem of overfitting the majority class, resulting in insufficient diagnostic ability for minority-class faults. In addition, the "black box" characteristic of deep learning models limits their interpretability in the industrial field and it is difficult to provide engineers with clear fault tracing information. Traditional statistical methods and rule-based diagnostic methods rely more on domain expert knowledge and can provide a certain degree of interpretability. However, these methods have high requirements for the balance of data distribution and sample size, and it is difficult to meet the diverse fault diagnosis needs in the industrial environment, especially when dealing with minority-class faults with scarce data.

[0005] To address the problem of unbalanced data distribution among fault categories, some studies have introduced data augmentation techniques and prior knowledge fusion techniques in recent years: Data augmentation techniques (such as Synthetic Minority Over-sampling Technique, SMOTE) expand the data distribution of minority-class faults by generating virtual samples, thereby alleviating the problem of data imbalance. Although the SMOTE technique has improved the diagnostic ability of minority-class faults to a certain extent, the virtual samples it generates may not fully reflect the true fault patterns, especially in high-dimensional and multi-fault scenarios, which may affect the diagnostic accuracy and robustness of the model. Bayesian networks combine domain knowledge with data-driven methods through causal relationship modeling, making the diagnostic model highly interpretable. However, in complex industrial scenarios, how to effectively construct the Bayesian network structure while taking into account the diagnostic accuracy of minority-class faults and the robustness of majority-class faults remains a technical challenge. In addition, the complexity of equipment operation in the industrial production environment further exacerbates the difficulty of fault diagnosis. For example, factors such as equipment aging, environmental fluctuations, and improper operation may lead to the diversity and uncertainty of fault patterns, and trigger the concurrency or rapid switching of multiple fault categories. Summary of the Invention

[0006] Aiming at the technical problems existing in the prior art, the present invention provides a method for diagnosing the unbalanced data fault of a coal mill based on a Bayesian network, which can balance the diagnostic performance of minority-class faults and majority-class faults in complex scenarios with multiple faults coexisting, and achieve high-efficiency interpretability of the diagnostic results.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] 1. Offline data processing

[0009] Step 1: Fault sample cleaning;

[0010] Specifically, clean the original data according to the integrity, that is, the missing rate < 5%, the legality, that is, the value is within the physical range, and the standard of deleting redundant records; use the interquartile range IQR method to detect extreme outliers, and define the outlier threshold as Q1 - 1.5IQR and Q3 + 1.5IQR, and eliminate those beyond the range.

[0011] Step 2: Unbalanced data augmentation;

[0012] Specifically, according to the fault labels, distinguish the majority class and the minority class, and use SMOTE to augment the minority-class samples: For each minority-class sample x i , search for k nearest-neighbor minority-class samples in the feature space; randomly select one of the neighbor samples x j , and perform linear interpolation to generate a new sample point X new=x i +λ(x j -x i ), where λ~U(0,1), and λ is a random interpolation coefficient; repeat until the number of samples in the minority class is the same as that in the majority class.

[0013] Step 3: Data normalization and discretization;

[0014] Specifically, the Min - Max method is used to map continuous sensor data to the interval [0,1]; based on the equal - frequency binning method, the normalized data is divided into 5 discrete intervals to ensure a uniform distribution of the sample size in each interval.

[0015] 2. Offline knowledge fusion

[0016] Step 4: Fault - attribute mapping modeling;

[0017] Specifically, first, define the key abnormal attributes of each fault category through domain knowledge; second, construct a fault - attribute matrix, which is specifically represented by the following formula:

[0018] ; where the rows correspond to the fault categories, the columns correspond to the attributes, and m ij =1 indicates that the fault will cause an obvious abnormality in the attribute , otherwise it is 0.

[0019] Step 5: Design of a three - layer Bayesian network architecture;

[0020] The specific structure is as follows:

[0021] Fault layer, i.e., the top layer: composed of all fault - category nodes, representing the possible fault modes of the system;

[0022] Fault - attribute layer, i.e., the middle layer: fault - attribute nodes extracted based on domain knowledge, describing the typical characteristics of faults;

[0023] Fault - variable layer, i.e., the bottom layer: contains monitoring - variable nodes, whose values are derived from the discretized sensor data and directly reflect the real - time state of the system.

[0024] 3. Offline model learning

[0025] Step 6: Bayesian network structure learning

[0026] (1) Construct the initial skeleton:

[0027] Forced connection: Based on expert knowledge, fix the directed edges of "fault node → fault - attribute node", corresponding to the logical dependencies of the fault - attribute matrix;

[0028] (2) Open search: Allow data-driven algorithms to explore the potential causal relationships between "fault attribute nodes → monitoring variable nodes" and between attribute / variable nodes.

[0029] Greedy structure search: Use the Hill-Climbing algorithm to iteratively perform edge addition, edge deletion, and edge reversal operations, and use the BIC scoring function to balance model accuracy and complexity:

[0030] ; where logP(D∣G) is the log-likelihood; ∣θ∣ is the total number of network parameters, and N is the number of samples.

[0031] (3) Industrial mechanism constraints:

[0032] Direction constraint: Prohibit attribute / variable nodes from pointing backward to the fault node;

[0033] Whitelist: Forcefully retain the directed edges determined by the process logic, such as attribute A → variable B;

[0034] Blacklist: Prohibit connections between physically irrelevant nodes, such as temperature attribute → vibration variable;

[0035] Step 7: Bayesian network parameter learning;

[0036] (1) Calculate the joint distribution frequency of a node and its parent nodes based on the training data to generate the initial CPT;

[0037] (2) To avoid the zero-probability problem, use Dirichlet prior smoothing and introduce a moderate prior smoothing coefficient in each value combination:

[0038] ; where N(X i , Pa(X i )) is the joint frequency of X i and its parent nodes, and α is the prior smoothing coefficient.

[0039] Step 8: Dual-mode diagnosis strategy;

[0040] (1) Direct inference of fault nodes: Input real-time monitoring variable evidence and calculate the posterior probability of the fault node; if the maximum posterior probability exceeds the set threshold τ1, then output the corresponding fault category; otherwise, mark it as "unknown fault".

[0041] (2) Indirect inference of attribute nodes: Calculate the posterior probability of the fault attribute node, binarize it into an abnormal pattern; perform similarity matching with the fault-attribute matrix, and if the cosine similarity > τ2, output the most matching fault or "unknown fault".

[0042] Step 9: Parameter joint optimization;

[0043] Joint optimization is carried out through grid search combined with K-fold cross-validation, and the optimal configuration of [α, τ 1, τ2] is selected. Taking the weighted score F1-score of diagnostic accuracy and false alarm rate as the optimization goal, the optimal parameter combination is determined.

[0044] Step 10: Model solidification and deployment;

[0045] Lock the optimal network structure, CPT parameters and diagnostic threshold, and export them as a lightweight inference engine; integrate it into the coal mill monitoring system to support real-time inference at a rate of hundreds of times per second.

[0046] 4. Online fault diagnosis

[0047] Step 11: Online diagnosis process

[0048] (1) Data preprocessing: Real-time sensor data is mapped to the values of the observation layer nodes after normalization and binning;

[0049] (2) Dual-mode synchronous inference:

[0050] Direct inference mode: Calculate the posterior probability of the fault node and output the fault category with the maximum probability;

[0051] Indirect inference mode: Calculate the posterior probability of the attribute node, generate an abnormal mode and match it with the fault-attribute matrix.

[0052] The beneficial effects of the present invention are as follows:

[0053] By performing balancing and discretization learning on the data in the offline stage, it can handle complex industrial scenarios where multiple faults exist in parallel or switch quickly; the Bayesian network combines domain expert knowledge with data-driven, and the network structure itself reflects the causal relationship between faults, attributes, and monitoring variables, facilitating subsequent analysis and traceability; using Dirichlet prior smoothing effectively reduces the negative impact of unbalanced data on parameter learning and improves the accuracy and stability of diagnosis; online inference is efficient: The inference mechanism based on the Bayesian network only needs to update and propagate in the network for the given evidence, and the computational complexity is relatively controllable, enabling real-time or quasi-real-time online diagnosis of faults.

[0054] Through the above steps and method configurations, the recognition and diagnosis accuracy of multi-fault scenarios can be significantly improved, meeting the requirements of industrial systems for reliability and interpretability. This method not only performs well in scenarios with many fault types and unbalanced data, but can also be extended and applied to other similar industrial process monitoring and fault diagnosis tasks. Description of the drawings

[0055] Figure 1It is a schematic flowchart of a fault diagnosis method for unbalanced data of a coal mill based on a Bayesian network in an embodiment of the present invention.

[0056] Figure 2 It is a schematic diagram of the Bayesian network structure in an embodiment of the present invention.

[0057] Figure 3 It is a schematic flowchart of the coal pulverizing process of the coal mill in an embodiment of the present invention.

[0058] Figure 4 It is a schematic diagram of the fault-attribute matrix of the coal mill in an embodiment of the present invention. Detailed implementation manners

[0059] The following combines the accompanying drawings and embodiments to further describe in detail the specific implementation manners of the present invention. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0060] Aiming at the problems existing in the prior art, the present invention proposes a fault diagnosis method for unbalanced data of a coal mill based on a Bayesian network. Aiming at the problem of uneven distribution of fault category data in the industrial scenario, combined with the data enhancement technology SMOTE and domain prior knowledge, it solves the technical pain point of difficult diagnosis of minority-class faults. Through offline data cleaning, sample balancing, normalization and discretization processing, and domain knowledge sorting to construct a fault-attribute matrix, combined with Bayesian network causal modeling, a Bayesian network model structure with a three-layer structure as shown in Figure 2 is established, and the network structure is optimized by the Hill-Climbing algorithm. In the parameter learning stage, the Dirichlet prior smoothing method is adopted to alleviate the influence of small samples and zero-probability problems on the model, and improve the stability and robustness of the diagnostic model. Finally, the optimized model is deployed in the online environment, and through the combination of direct inference of fault nodes and indirect inference of attribute nodes, real-time or quasi-real-time fault diagnosis is realized, significantly improving the accuracy and interpretability of the diagnosis.

[0061] The core goal of the present invention is to solve the challenges of uneven distribution of fault category data in industrial fault diagnosis, difficult diagnosis of minority-class faults, and coexistence of multiple faults in a complex industrial environment. By introducing SMOTE data enhancement and Dirichlet prior smoothing technologies, the diagnostic ability of minority-class faults is improved; through the causal modeling ability of the Bayesian network, the high interpretability of the diagnostic results is ensured; and through the online inference mechanism, the balance between real-time performance and diagnostic robustness is achieved. Finally, a precise and reliable fault diagnosis solution is provided for industrial systems to ensure the safe and efficient operation of equipment.

[0062] Specifically, the present invention provides a fault diagnosis method for unbalanced data of a coal mill based on a Bayesian network, combined with Figure 1, including the following steps:

[0063] 1. Offline data processing

[0064] Step 1: Fault sample cleaning;

[0065] Specifically, clean the original data according to the integrity (i.e., the missing rate < 5%) and legality (i.e., the value is within the physical range), and delete redundant records. Detect extreme outliers using the interquartile range (IQR) method, and define the outlier threshold as Q1 - 1.5IQR and Q3 + 1.5IQR. Remove the data points outside this range.

[0066] Step 2: Imbalanced data augmentation;

[0067] According to the fault labels, count and distinguish the majority class and the minority class, and use SMOTE to augment the minority class samples: For each minority class sample x i , search for k nearest neighbor minority class samples in the feature space; randomly select one of the neighbor samples x j , and perform linear interpolation to generate a new sample point X new =x i +λ(x j -x i ), where λ ~ U(0,1) and λ is the random interpolation coefficient; repeat until the number of minority class samples is the same as that of the majority class samples.

[0068] Step 3: Data normalization and discretization;

[0069] Specifically, use the Min - Max method to map continuous sensor data to the [0,1] interval; divide the normalized data into 5 discrete intervals based on the equal - frequency binning method to ensure uniform distribution of the sample size in each interval.

[0070] 2. Offline knowledge fusion

[0071] Step 4: Fault - attribute mapping modeling;

[0072] Specifically, first define the key abnormal attributes of each fault category through domain knowledge; secondly, construct a fault - attribute matrix, which is specifically represented by the following formula:

[0073] ; where the rows correspond to the fault categories and the columns correspond to the attributes. m ij =1 indicates that the fault will cause obvious abnormality of the attribute , otherwise it is 0.

[0074] Step 5: Design of the three - layer Bayesian network architecture;

[0075] The specific structure is as follows:

[0076] Fault layer, i.e., the top layer: composed of all fault category nodes, representing the possible fault modes of the system;

[0077] Fault attribute layer, i.e., the middle layer: fault attribute nodes extracted based on domain knowledge, describing the typical characteristics of faults;

[0078] Fault variable layer, i.e., the bottom layer: contains monitoring variable nodes, whose values are derived from discretized sensor data, directly reflecting the real-time state of the system.

[0079] 3. Offline model learning

[0080] Step 6: Bayesian network structure learning

[0081] (1) Construct the initial skeleton:

[0082] Forced connection: Based on expert knowledge, fix the directed edges of "fault node → fault attribute node", corresponding to the logical dependencies of the fault-attribute matrix;

[0083] (2) Open search: Allow data-driven algorithms to explore the potential causal relationships between "fault attribute nodes → monitoring variable nodes" and between attribute / variable nodes.

[0084] Greedy structure search: Use the Hill-Climbing algorithm to iteratively perform edge addition, edge deletion, and edge reversal operations, and use the BIC scoring function to balance model accuracy and complexity:

[0085] ; where logP(D∣G) is the log-likelihood; ∣θ∣ is the total number of network parameters, and N is the number of samples.

[0086] (3) Industrial mechanism constraints:

[0087] Direction constraint: Prohibit attribute / variable nodes from pointing back to fault nodes;

[0088] White list: Forcefully retain the directed edges determined by process logic, such as attribute A → variable B;

[0089] Black list: Prohibit connections between physically unrelated nodes, such as temperature attribute → vibration variable.

[0090] Step 7: Bayesian network parameter learning;

[0091] (1) Calculate the joint distribution frequency of a node and its parent nodes based on the training data to generate the initial CPT;

[0092] (2) To avoid the zero-probability problem, use Dirichlet prior smoothing and introduce a moderate prior smoothing coefficient in each value combination: ; where N(X i ,Pa(Xi )) is X i The joint frequency with its parent node, and α is the prior smoothing coefficient.

[0093] Step 8: Dual-mode diagnosis strategy;

[0094] (1) Direct inference of fault nodes: Input the evidence of real-time monitoring variables, and calculate the posterior probability of fault nodes; if the maximum posterior probability exceeds the set threshold τ1, output the corresponding fault category; otherwise, mark it as "unknown fault".

[0095] (2) Indirect inference of attribute nodes: Calculate the posterior probability of fault attribute nodes, and binarize them into abnormal patterns; match with the fault-attribute matrix, if the cosine similarity > τ2, output the most matching fault or "unknown fault".

[0096] Step 9: Joint optimization of parameters;

[0097] Perform joint optimization through grid search combined with K-fold cross-validation, select the optimal configuration of [α, τ 1, τ2], and take the weighted score F1-score of diagnostic accuracy and false alarm rate as the optimization goal to determine the optimal parameter combination.

[0098] Step 10: Model solidification and deployment;

[0099] Lock the optimal network structure, CPT parameters and diagnostic thresholds, and export them as a lightweight inference engine; integrate it into the coal mill monitoring system to support real-time inference at a rate of hundreds of times per second.

[0100] 4. Online fault diagnosis

[0101] Step 11: Online diagnosis process

[0102] (1) Data preprocessing: Real-time sensor data is mapped to the values of the observation layer nodes after normalization and binning;

[0103] (2) Dual-mode synchronous inference:

[0104] Direct inference mode: Calculate the posterior probability of fault nodes and output the fault category with the maximum probability;

[0105] Indirect inference mode: Calculate the posterior probability of attribute nodes, generate abnormal patterns and match them with the fault-attribute matrix.

[0106] The key innovation of the present invention lies in the dual-stage diagnosis mode combining offline and online, which organically integrates SMOTE minority class enhancement, prior smoothing, and a three-layer Bayesian network structure. It can not only obtain a more balanced fault recognition effect in the case of small samples and data imbalance, but also make full use of industrial prior knowledge to construct a fault-attribute matrix, thereby achieving high-precision and highly interpretable diagnosis in the scenario of multiple concurrent faults. In particular, it is worth noting that the method designs two paths of direct inference and indirect inference for the fault layer, attribute layer, and variable layer respectively. It can not only quickly output the posterior probability of the fault node, but also trace the specific abnormal mechanism based on the attribute node, making the diagnosis result more intuitive and transparent, and facilitating subsequent maintenance and troubleshooting.

[0107] Meanwhile, the present invention optimizes the parameter configuration through adaptive threshold search, namely grid search and cross-validation, reduces the deviation of manual parameter tuning, and ensures the diagnosis efficiency and model stability in the online stage. Its generality or adaptability is also quite prominent: without changing the basic principle, it can be widely applied to different industrial scenarios such as chemical industry, metallurgy, electric power, and machinery manufacturing by simply replacing the domain knowledge and monitoring data in the offline stage, providing an efficient and flexible solution for the timely discovery and accurate positioning of diverse fault types. The following further describes the present invention in detail through specific embodiments.

[0108] (1) Experimental scenario and data configuration

[0109] Taking the coal pulverizing system of a certain supercritical coal-fired unit as the research object, the simplified schematic diagram of the coal pulverizing process is as Figure 3 shown. Raw coal enters the feeder from the raw coal bunker through the raw coal bunker coal dropping pipe, is conveyed to the feeder coal dropping pipe by the feeder, and enters the coal mill for grinding into pulverized coal. The cold and hot primary air are mixed through the regulating valve and then blown into the coal mill, and the mixed primary air carries the pulverized coal after grinding and blows it into the boiler.

[0110] Twelve key monitoring variables are selected, including the opening degree of the cold air valve, the opening degree of the hot air valve, the coal feeding flow rate, the moisture content of raw coal, the primary air volume, the primary air temperature, the amount of raw coal in the coal mill, the amount of pulverized coal in the coal mill, the current of the coal mill, the amount of pulverized coal blown out, the outlet temperature, and the moisture content in the pulverized coal. Six types of typical faults are defined, as shown in Table 1, and eight fault attributes are defined, as shown in Table 2. The training set simulates the industrial data imbalance scenario, with the majority class fault containing 480 samples, the minority class fault containing 20 samples, and the ratio of 1:120. The test set uses balanced data to verify the generalization ability. Through ablation experiments and multi-scenario tests, the independent and combined effects of SMOTE data augmentation and Dirichlet prior smoothing are compared, and the effectiveness of the proposed method is verified. The schematic diagram of the coal mill fault-attribute matrix is as Figure 4 shown.

[0111] Table 1 Coal mill fault types and fault description table

[0112]

[0113] Table 2 Description Table of Coal Mill Fault Attributes

[0114]

[0115] On this basis, the fault diagnosis of the coal mill process is realized by using Steps 1 to 11.

[0116] (2) Results of Ablation Experiments

[0117] In the ablation experiment, F6 is used as the minority-class fault, and the other 5 classes are used as the majority-class faults. The effects of SMOTE data augmentation and Dirichlet prior smoothing on the model performance are verified respectively. The experimental results are shown in Table 3.

[0118] Table 3 Results Table of Ablation Experiments

[0119]

[0120] According to the results, it can be seen that the recall rate Recall and F1 score of the traditional MLE method for the minority class are only 0.79 - 0.84, indicating that it is difficult to handle the problem of data imbalance. After introducing SMOTE alone, the F1 score of F6 is increased to 0.90, proving that data augmentation effectively expands the feature space of the minority class; while when only using Dirichlet prior smoothing, the F1 score does not increase significantly, and even shows a slight decrease, indicating that the smoothing technology needs to cooperate with data augmentation. When SMOTE and Dirichlet smoothing are applied jointly, the diagnostic performance of the model for F6 reaches the optimal level (F1 = 0.985), and the recall rate is increased to 98.5%, significantly reducing the risk of missed detection.

[0121] Verification of Robustness for Multiple Minority-Class Fault Scenarios

[0122] As shown in Table 4, in a single minority-class fault scenario (such as F1 - F6 being used as the minority class respectively), the F1 scores of the model are all stable above 0.92, and the fault diagnosis accuracies of both the minority class and the majority class can be guaranteed. As shown in Table 5, further setting multiple complex minority-class fault scenarios (such as 2 - 3 minority-class faults), the model still maintains high robustness: F1 > 0.98 for 2 minority-class faults and F1 = 1 for 3 minority-class faults, indicating that SMOTE can independently generate synthetic samples for each class to avoid feature confusion. It should be noted that when the attribute features of multiple minority-class faults are significantly different, such as F2 "bearing overheat" and F5 "powder outlet pipe blockage", the model can accurately distinguish them by strengthening their respective causal paths.

[0123] Table 4 Results Table of Single Minority-Class Fault Experiment

[0124]

[0125] Table 5 Experimental results table of multiple minority-class faults

[0126]

[0127] (3)Interpretability verification and industrial application value

[0128] The diagnostic mode at the attribute layer reveals the fault mechanism through abnormal attribute combinations. For example, when "abnormal fineness of pulverized coal A7 + increased outlet pressure A5" is detected, it is determined that the powder outlet pipe F5 is blocked, and it is prompted to check the grinding components or clean the pipeline; "abnormal bearing temperature A3 + enhanced vibration A2" indicates F2 bearing overheating or damage, and the lubrication system is preferentially repaired. Compared with the direct probability inference F1 = 0.985, although the indirect inference mode F1 = 0.9287 has slightly lower accuracy, it provides a physical-level attribution explanation, supports engineers to quickly locate the root cause of the fault, and misdetection cases can be traced back to specific attribute abnormalities, such as "when misjudging F4, it is actually caused by the A8 pressure difference fluctuation".

[0129] The comprehensive experimental results show that the method proposed in this study, which combines SMOTE enhancement, Dirichlet smoothing, and causal network, achieves high accuracy in the fault diagnosis of coal mills in the context of data imbalance, with a significant improvement compared to traditional methods. Its hierarchical diagnostic mechanism, rapid early warning at the fault layer → physical attribution at the attribute layer, taking into account both real-time performance and interpretability, provides an innovative path from "black-box prediction" to "transparent decision-making" for industrial scenarios. In practical applications, this method has successfully reduced unplanned outages by 70%, verifying its engineering practical value.

[0130] The above description is only the preferred embodiments of the present disclosure and the description of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the technical features disclosed in the embodiments of the present disclosure but not limited to having similar functions.

Claims

1. A method for diagnosing coal mill imbalance data faults based on Bayesian networks, characterized in that: Step 1: Offline data processing: preprocess the collected data, clean the fault samples, enhance the unbalanced data, and normalize and discretize the data; Step 2: Offline knowledge fusion: Model the fault-attribute mapping and construct a three-layer Bayesian network for fault classification; Step 2.1: Fault-attribute mapping modeling: Define the key abnormal attributes of each fault category through domain knowledge; then construct the fault-attribute matrix, which is expressed by the following formula: ; where rows correspond to fault categories, columns correspond to attributes, m ij =1 indicates fault Will cause attributes The obvious abnormality is 0 otherwise; Step 2.1: Three-layer Bayesian network architecture design: Fault layer, i.e. top layer: consists of all fault category nodes, representing the possible fault modes of the system; Fault attribute layer, i.e., the middle layer: fault attribute nodes extracted based on domain knowledge, describing typical characteristics of faults; Fault variable layer, i.e. bottom layer: contains monitoring variable nodes, whose values ​​are derived from sensor data after discretization, reflecting the real-time status of the system; Step 3: Offline model learning: Learn the Bayesian network structure and Bayesian network parameters, mark fault information through the dual-mode diagnosis strategy, optimize the parameter combination, and obtain the final model; 3.1: Bayesian network structure learning: 3.1.1: Build the initial skeleton: Forced connection: Fixed the directed edge of "fault node → fault attribute node" based on expert knowledge, corresponding to the logical dependency of the fault-attribute matrix; Open search: Allows data-driven algorithms to explore the potential causal relationship between "fault attribute node → monitoring variable node" and attribute / variable nodes; 3.1.2: Greedy structure search: The Hill-Climbing algorithm is used to iteratively perform edge addition, edge deletion, and edge reversal operations, and the BIC scoring function is used to balance the model accuracy and complexity: ; Where logP(D|G) is the log likelihood; |θ| is the total number of network parameters, and N is the number of samples; 3.1.3: Industrial mechanism constraints: Direction constraint: prohibits attribute / variable nodes from pointing in the opposite direction to the fault node; Whitelist: Forces the retention of directed edges determined by process logic; Blacklist: prohibits connections between physically unrelated nodes; 3.2: Bayesian network parameter learning: 3.2.1: Calculate the joint distribution frequency of the node and its parent node based on the training data to generate the initial CPT; 3.2.2: To avoid the zero probability problem, use Dirichlet prior smoothing and introduce a moderate prior smoothing coefficient in each value combination: ; where N(X i ,Pa(X i )) is X i The joint frequency with its parent node, α is the prior smoothing coefficient; 3.3: Dual-mode diagnostic strategy: 3.3.1: Direct inference of faulty nodes: Input real-time monitoring variable evidence and calculate the posterior probability of the faulty node; if the maximum posterior probability exceeds the set threshold τ1, the corresponding fault category is output; otherwise, it is marked as "unknown fault"; 3.3.2: Indirect inference of attribute nodes: Calculate the posterior probability of the fault attribute node and binarize it into an abnormal pattern; perform similarity matching with the fault-attribute matrix, i.e., cosine similarity>τ2, and output the most matching fault or "unknown fault"; 3.4: Joint Parameter Optimization: Through grid search combined with K-fold cross validation, joint optimization is performed to select [α,τ 1, τ2] optimal configuration, taking the weighted score F1-score of diagnostic accuracy and false alarm rate as the optimization target, and determining the optimal parameter combination; 3.5: Model solidification and deployment: The optimal network structure, CPT parameters and diagnostic thresholds are locked and exported as a lightweight inference engine; integrated into the coal mill monitoring system to support real-time inference at the rate of hundreds of times per second; Step 4: Perform online diagnosis.

2. A method for diagnosing coal mill imbalance data faults based on Bayesian network according to claim 1, characterized in that: In the step 1, the specific steps are: Step 1.1: Fault sample cleaning: Clean the original data according to the standards of completeness, i.e., missing rate <5%, legality, i.e., values ​​within the physical range, and deletion of redundant records; use the interquartile range IQR method to detect extreme outliers, define the anomaly thresholds as Q1-1.5IQR and Q3+1.5IQR, and exclude those beyond the range; Step 1.2: Imbalanced data enhancement: According to the fault label, statistically distinguish the majority class and the minority class, and use SMOTE to augment the minority class samples: For each minority class sample x i , search for k nearest neighbor minority class samples from the feature space; randomly select one of the nearest neighbor samples x j , and perform linear interpolation to generate new sample points X new =x i +λ(x j -x i ),λ~U(0,1), where λ is the random interpolation coefficient; repeat until the number of minority class samples is the same as the number of majority class samples; Step 1.3: Data normalization and discretization: Use the Min-Max method to map the continuous sensor data to the [0,1] interval; The normalized data were divided into five discrete intervals based on the equal frequency binning method to ensure that the sample size in each interval was evenly distributed.

3. A method for diagnosing coal mill imbalance data faults based on Bayesian network according to claim 1, characterized in that: In the step 4, the specific steps are: 4.1: Data preprocessing: Real-time sensor data is normalized, binned and mapped to observation layer node values; 4.2: Dual-mode simultaneous inference: Direct inference mode: calculate the posterior probability of the faulty node and output the maximum probability fault category; Indirectly infer patterns: calculate the posterior probability of attribute nodes, generate abnormal patterns and match them with the fault-attribute matrix.

Images