A Fault Detection Method and System for Coal Mills in Thermal Power Units Based on Multivariate Data Reconstruction

Abstract

This application relates to a fault detection method for coal mills in thermal power units based on multivariate data reconstruction. The method includes: acquiring time-series and historical data of multiple operating parameters of the coal mill during operation; constructing time-shifted feature vectors for each operating parameter based on the principle of maximizing mutual information; inputting the time-shifted feature vectors into a pre-trained TS-MSDAE model to obtain the reconstructed values ​​of each operating parameter; determining the reconstruction error by combining dynamic weights with the reconstructed values ​​and time-series data; determining a fault judgment threshold based on the sliding window statistics of the reconstruction error of normal samples; and determining whether the operating state of the coal mill is abnormal based on the fault judgment threshold and the reconstruction error. This application solves the problem of poor adaptability in traditional coal mill fault diagnosis methods. The TS-MSDAE model adopts a dual-scale historical state splicing architecture, which improves the model's adaptability to noise and operating condition changes and reduces the risk of overfitting.

Description

Technical Field

[0001] This application relates to the field of coal mill fault detection, and in particular to a method and system for coal mill fault detection in thermal power units based on multivariate data reconstruction. Background Technology

[0002] As one of the most important auxiliary equipment in thermal power generating units, the coal mill's operating status directly affects the stability and economy of the unit's operation. Common and serious operating faults in coal mills include coal blockage, coal shortage, and spontaneous combustion. These types of operating faults involve multiple coupled parameters, and the dynamic changes of each operating parameter will show corresponding trends.

[0003] Existing methods for pulverizer fault diagnosis mainly rely on physical mechanism modeling and fixed threshold monitoring (such as pulverizer current and outlet temperature thresholds), or on extracting time-frequency domain features (such as wavelet packet energy entropy and vibration spectrum) through feature engineering-driven models (such as support vector machines and random forests) to achieve state classification. However, fixed threshold methods (such as judging coal blockage by current > 120A) are difficult to adapt to dynamic changes such as coal type switching and equipment wear; model analysis methods based on SVM and random forests require manual construction of time-frequency domain features, but pulverizer parameters have strong nonlinear coupling, and manual features are difficult to fully characterize the fault modes; moreover, the proportion of fault data in actual operation is extremely low, while supervised learning models such as CNN and LSTM rely on a large number of labeled samples, which are prone to missed detections due to sample imbalance. Summary of the Invention

[0004] This application provides a method, system, electronic device, and storage medium for detecting coal mill faults in thermal power units based on multivariate data reconstruction, in order to at least solve the problem of poor adaptability of traditional coal mill fault diagnosis methods in related technologies.

[0005] In a first aspect, embodiments of this application provide a method for fault detection of coal mills in thermal power units based on multivariate data reconstruction, the method comprising: The system acquires time-series and historical data of multiple operating parameters of the coal mill during operation, including material flow parameters, equipment status parameters, and air volume and temperature parameters. Based on the principle of maximizing mutual information, a time-shift feature vector of each of the operating parameters is constructed according to the time-series data and the historical data. The time-shifted feature vector is input into the pre-trained TS-MSDAE model to obtain the reconstructed values ​​of each of the running parameters; the contribution of each running parameter to the reconstruction error is calculated based on the SHAP value to obtain the dynamic weight of each running parameter; the reconstruction error is determined by combining the dynamic weight with the reconstructed values ​​and the time series data. The fault judgment threshold is determined based on the sliding window statistics of the reconstruction error of normal samples. Based on the fault judgment threshold and the reconstruction error, it is determined whether the operating status of the coal mill is abnormal.

[0006] In some embodiments, constructing the time-shift feature vector of each of the operating parameters based on the principle of maximizing mutual information, according to the time-series data and the historical data, includes: Based on the time-series data and the historical data, construct the time-shift mutual information curve for each operating parameter; Based on the time-shift mutual information curve, determine the target time delay when the mutual information between different operating parameters is maximized; Based on the target time delay, the time-series data and the historical data are concatenated to obtain the time-shift feature vector.

[0007] In some embodiments, determining the fault judgment threshold based on the sliding window statistic of normal sample reconstruction error includes: Within a sliding time window of the first preset time period, calculate the mean and standard deviation of the reconstruction error of normal samples; The fault judgment threshold is determined based on the mean, the standard deviation, and the adaptive sensitivity coefficient, wherein the adaptive sensitivity coefficient is determined based on the stability of the unit's operating conditions.

[0008] In some embodiments, the TS-MSDAE model structure is a stacked noise-reducing autoencoder that includes a multi-layer encoder and a multi-layer decoder; The encoder comprises a bottom layer, a middle layer, and a high layer. The bottom layer is used to capture instantaneous parameter fluctuations, the middle layer is used to fuse cross-parameter nonlinear correlations, and the high layer is used to model long-range causal chains and compress parameter features into a low-dimensional latent space.

[0009] In some embodiments, the method further includes: If the average reconstruction error of the collected normal samples exceeds twice the baseline error within the sliding time window of the second preset time period, the TS-MSDAE model is considered to need to perform incremental learning. The incremental learning process includes: locking the weight parameters of the encoder and adjusting the weight parameters of the decoder based on the incremental learning model.

[0010] In some embodiments, after determining whether the operating status of the coal mill is abnormal, the method further includes: When the reconstruction error exceeds the fault judgment threshold and triggers a state abnormality signal, the feature vector in the latent space of the TS-MSDAE model is extracted. Calculate the similarity between the feature vector and the feature vectors of all known fault prototypes stored in the fault memory bank; If the maximum similarity exceeds the preset matching threshold, the fault type of the current feature vector is considered to be the fault type corresponding to the prototype feature vector with the highest similarity. If the maximum similarity does not exceed the preset matching threshold, the current feature vector is temporarily stored and added to the fault memory bank as a new fault prototype feature vector after manual confirmation.

[0011] In some embodiments, after acquiring time-series data and historical data of multiple operating parameters of the coal mill during operation, the method further includes: The sliding window mean method is used to handle missing values ​​in the time series data and the historical data; The dimensional differences between the time-series data and the historical data are eliminated through standardization processing; By filtering out transient interference from sensors, outliers are removed from the time-series data and the historical data.

[0012] Secondly, embodiments of this application provide a fault detection system for coal mills in thermal power units based on multi-source data reconstruction, the system comprising: The data acquisition module is used to acquire time-series and historical data of multiple operating parameters of the coal mill during operation. The operating parameters include material flow parameters, equipment status parameters, and air volume and air temperature parameters. The feature extraction module is used to construct time-shift feature vectors for each of the operating parameters based on the principle of maximizing mutual information, according to the time-series data and the historical data. The data reconstruction module is used to input the time-shifted feature vector into a pre-trained TS-MSDAE model to obtain the reconstruction values ​​of each of the running parameters; calculate the contribution of each running parameter to the reconstruction error based on the SHAP value to obtain the dynamic weight of each running parameter; and determine the reconstruction error by combining the dynamic weight with the reconstruction value and the time series data. The status judgment module is used to determine the fault judgment threshold based on the sliding window statistics of the reconstruction error of normal samples, and to determine whether the operating status of the coal mill is abnormal according to the fault judgment threshold and the reconstruction error.

[0013] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the coal mill fault detection method for thermal power units based on multi-source data reconstruction as described in the first aspect above.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the coal mill fault detection method for thermal power units based on multi-source data reconstruction as described in the first aspect above.

[0015] Compared to related technologies, the coal mill fault detection method for thermal power units based on multivariate data reconstruction provided in this application employs a dual-scale historical state splicing architecture in its TS-MSDAE model. This architecture extracts and fuses features from time-series data at different granularities, enabling the model to simultaneously capture local details and global trends, effectively addressing the complex fault propagation mechanisms in equipment such as coal mills. Multi-scale feature fusion improves the model's adaptability to noise and operating condition changes, reduces the risk of overfitting, and solves the problem of poor adaptability in traditional coal mill fault diagnosis methods. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a coal mill fault detection method for thermal power units based on multi-source data reconstruction according to an embodiment of this application; Figure 2 This is a network structure diagram of an autoencoder according to an embodiment of this application; Figure 3 This is a flowchart of a state monitoring process based on a multi-data reconstruction according to an embodiment of this application; Figure 4 This is a flowchart of a method for detecting operational faults in a coal mill of a thermal power unit according to an embodiment of this application; Figure 5 This is a structural block diagram of a coal mill fault detection system for thermal power units based on multi-data reconstruction, according to an embodiment of this application. Figure 6 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0018] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0019] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0020] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0021] This embodiment provides a method for fault detection of coal mills in thermal power units based on multivariate data reconstruction. Figure 1 This is a flowchart of a coal mill fault detection method for thermal power units based on multi-source data reconstruction according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps: Step S101: Obtain time-series data and historical data of multiple operating parameters of the coal mill during operation. The operating parameters include material flow parameters, equipment status parameters, and air volume and temperature parameters.

[0022] In this embodiment, the material flow parameters include coal feed rate, coal powder inventory, and coal powder outlet flow rate; the equipment status parameters include mill current; and the air volume and temperature parameters include hot air valve opening, cold air valve opening, primary air flow rate, primary air temperature, and mill outlet temperature.

[0023] In some embodiments, step S101 includes: Step S1011: The sliding window mean method is used to handle missing values ​​in the time series data and historical data.

[0024] Step S1012: Eliminate the dimensional differences between time-series data and historical data based on standardization processing.

[0025] Step S1013: Remove outliers from time-series and historical data by filtering sensor transient interference.

[0026] The data vector is defined as follows:

[0027] Perform three levels of preprocessing on the raw data.

[0028] Missing value handling: A sliding window mean method is used to ensure data continuity and avoid introducing high-frequency noise. The formula for calculating the sliding window mean is as follows:

[0029] in, t For time points with missing values, x i For the first i A sequence of parameters, x i fill The result is calculated for missing values.

[0030] Standardization: Eliminates dimensional differences and improves model convergence speed. The standardization formula is as follows:

[0031] in, μ i For the mean, σ i Standard deviation, t This is the value at the current time. N This is the window width.

[0032] Outlier filtering formula:

[0033] Values ​​that satisfy the above filtering formula are removed to filter out transient interference from the sensor and reduce the false trigger rate.

[0034] The preprocessed data has high integrity and dimensional consistency, which improves the accuracy of fault diagnosis.

[0035] Step S102: Based on the principle of maximizing mutual information, construct the time-shift feature vector of each operating parameter according to time series data and historical data.

[0036] As a complex dynamic system, the coal mill exhibits nonlinear, time-varying coupling relationships among its various monitoring parameters, which dynamically adjust as the equipment's operating status changes. Traditional fixed-delay analysis methods often struggle to accurately capture this dynamic characteristic, leading to incomplete fault feature extraction and limited diagnostic accuracy. This embodiment addresses this by concatenating historical and current data to construct multi-timescale feature vectors for delay optimization, enabling the diagnostic model to better capture the time-delay effects of fault propagation.

[0037]

[0038] x TS ∈R 27 For the concatenated multi-timescale feature vectors, R 27 Represents the 27-dimensional set of real numbers, where x i ( t (i=1,...,9) represents the feature vector of the current state. x i ( t -Δ τ 1) is a short-time feature vector. x i ( t -Δ τ 1) is a long-term feature vector.

[0039] In some embodiments, step S102 specifically includes: Step S1021: Based on time series data and historical data, construct the time-shift mutual information curve for each operating parameter.

[0040] Step S1022: Based on the time-shift mutual information curve, determine the target time delay when the mutual information between different operating parameters is maximized.

[0041] Step S1023: Based on the target time delay, the time series data and historical data are concatenated to obtain the time shift feature vector.

[0042] This embodiment continuously monitors the time-series data of various transmission parameters and uses mutual information from information theory as a metric to quantitatively analyze the time-varying correlation between different parameter pairs. The calculation of mutual information not only considers the linear correlation of parameter values ​​but also captures non-linear dependencies, enabling the system to discover deep associations that traditional correlation coefficients cannot detect. In the actual calculation process, the system constructs a time-shift mutual information curve for each pair of parameters to be analyzed. This curve shows the strength of their statistical dependence when there are different time shifts between the two parameter sequences. By finding the extreme points on the curve, the time delay that best reflects the true causal relationship between the two parameters can be determined.

[0043] Determining the delay based on maximizing parameter mutual information:

[0044] In the formula, ( x i , x j Let Δ be a pair of parameters to be analyzed. τ The time shift variable Δ is calculated using the maximum value set of independent variable points argmax, which is the time shift variable at which the mutual information between the two parameters is maximized. τ * .

[0045] The mutual information I(X,Y) is calculated as follows:

[0046] in p (X,Y) is the joint probability distribution function of X and Y, and p (x) and p (y) are the marginal probability distribution functions of X and Y, respectively.

[0047] A time-delay optimization method based on maximizing mutual information is adopted. By splicing historical states at two scales, the dynamic characteristics of fault propagation are solidified, which to some extent improves the deficiency of "time lag effect causing early warning delay" in coal mill fault diagnosis. Its structural optimization design (fixed dimension / integer quantization) meets the requirements of edge deployment, and the dual time window parameters are dynamically optimized by maximizing mutual information, which can adapt to changes in coal type and equipment aging. At the same time, as the equipment operates for a long time, the system will continuously update the time-delay relationship model of each parameter pair, automatically tracking the changes in time-delay characteristics caused by factors such as equipment aging and changes in operating conditions, so that the diagnostic system can maintain high accuracy over a long period of time.

[0048] Continue to refer to Figure 1 After obtaining the time-shifted feature vector, step S103 is executed.

[0049] Step S103: Input the time-shifted feature vector into the pre-trained TS-MSDAE model to obtain the reconstruction value of each running parameter; calculate the contribution of each running parameter to the reconstruction error based on the SHAP value to obtain the dynamic weight of each running parameter; and determine the reconstruction error by combining the dynamic weight with the reconstruction value and time series data.

[0050] To address the multi-dimensional and noisy nature of coal mill operation data, the TS-MSDAE model is adopted to mine deep-level feature information in the multi-dimensional data of the coal mill. A multi-data reconstruction model based on TS-MSDAE is constructed to extract deep-level robust features from normal operation data and learn the nonlinear mapping relationship between data and operation status, thereby realizing the monitoring of coal mill operation status.

[0051] Reconstruction error E of normal samples N This can be represented as the actual input value X. N With reconstructed value Z N The difference is shown in the following formula:

[0052] The input to the TS-MSDAE model is multivariate data X of the coal mill. In order to achieve efficient monitoring of the coal mill's operating status, a one-dimensional status monitoring index needs to be extracted from the multivariate reconstruction residual E.

[0053] The distance between two vectors is usually measured by similarity (currently, Euclidean distance is mostly used as the similarity index between vectors to measure the degree of deviation between input value and reconstructed value). However, considering that the similarity measure based on Euclidean distance ignores the different sensitivities of different parameters to faults, the traditional method of calculating reconstruction error treats all parameters equally, which leads to key fault features being submerged by noise from insensitive parameters. Differentiated processing of each feature is required. In this embodiment, a dynamic weighting method based on SHAP (Shapley Additive Explanations) values ​​is used to calculate the weighted reconstruction error.

[0054] Unlike Euclidean distance, SHAP (Shape-Up Point) is a game theory-based method for interpreting model output. It interprets the model's predictions by calculating the marginal contribution of each feature to the model's output. The core idea of ​​SHAP is to treat all features as "contributors" and calculate the average marginal contribution of each feature across different feature sequences.

[0055] Dynamic weight calculation based on SHAP values ​​includes calculating the contribution of each input feature to the reconstruction error and the normalized weight.

[0056] Calculate the contribution of each input feature to the reconstruction error ф i The calculation formula is:

[0057] Where E is the reconstruction error. x i For the first i The sequence of parameters is calculated by taking the partial derivative of the reconstruction error E with respect to each input parameter and then calculating the expected gradient IE of the output with respect to the input through backpropagation.

[0058] Normalized weight w i The calculation formula is:

[0059] Where j∈{1,2,3,……,27}, represents the dimension of the concatenated multi-time-scale feature vector.

[0060] The SHAP value enables dynamic weighted calculation of reconstruction error, quantifies contribution, and accurately measures the marginal contribution of each parameter to fault detection; it also achieves physical interpretability, with weight allocation consistent with equipment mechanism; and the weights can be adaptively adjusted, automatically updating the weight distribution as operating conditions change.

[0061] Step S104: Determine the fault judgment threshold based on the sliding window statistic of the reconstruction error of normal samples, and determine whether the operating status of the coal mill is abnormal based on the fault judgment threshold and the reconstruction error.

[0062] In some embodiments, determining the fault judgment threshold based on the sliding window statistic of normal sample reconstruction error includes: Within a sliding time window of the first preset time period, calculate the mean and standard deviation of the reconstruction error of normal samples; The fault judgment threshold is determined based on the mean, standard deviation, and adaptive sensitivity coefficient, where the adaptive sensitivity coefficient is determined according to the stability of the unit's operating conditions.

[0063] To adapt to changes in operating conditions, the threshold is dynamically updated based on sliding window statistics, continuously recording recent reconstruction error sequences, calculating their mean and standard deviation, and characterizing the current normal operating range.

[0064] Fault detection threshold τ t The calculation formula is as follows:

[0065] Among them, the mean μ t Standard deviation σ t and sensitivity coefficient K The calculation formula is:

[0066]

[0067] In the formula, E is the reconstruction error, and T is the time window, which can optionally be 300.

[0068] K automatically adjusts the alarm margin based on the stability of the operating conditions—tightening the threshold during the stable period to improve sensitivity, and loosening the threshold during the fluctuating period to suppress false alarms.

[0069] Through the above steps, a fault diagnosis process of "time-shift feature construction—stacked noise reduction encoding—dynamic threshold judgment—memory matching" is formed. Time-shift features solidify the fault propagation chain, breaking through the perception bottleneck of time delay effects in traditional single-point sampling and significantly extending the early warning window. The incremental learning mechanism enables the model to continuously adapt to equipment aging, and the fault prototypes accumulated in the memory bank enable accurate diagnosis of small samples. Optimization requires no manual intervention and achieves autonomous evolution. At the same time, the lightweight model and hardware acceleration work together to achieve millisecond-level response throughout the entire process in industrial controller-level equipment, completely solving the pain points of high latency and poor reliability of cloud solutions. Finally, an industrial-grade intelligent maintenance system with both early accurate diagnosis and autonomous evolution characteristics is built.

[0070] An autoencoder (AE) is a neural network used to learn feature representations of input data. Its structure typically includes an input layer, hidden layers, and an output layer. This network maps input data to the hidden layer feature space through an encoding process, and then maps the feature space back to the reconstructed data through a decoding process. It optimizes feature extraction capabilities by minimizing the error between the input data and the reconstructed data (reconstruction error). Model training uses gradient descent to optimize the reconstruction error function to obtain the best feature representation of the data. To avoid the model learning an identity mapping, a derivative autoencoder with specific functions can be constructed by introducing a constraint term into the loss function. Figure 2 This is a network structure diagram of an autoencoder according to an embodiment of this application.

[0071] Suppose the input sample set is {x1, x2, ..., x} n First, the encoding function is used to extract the feature representation of the training samples. Then, the decoding function maps the feature representation to the output y. Finally, the similarity between the input and output is characterized by the loss function.

[0072] The encoding process function of the autoencoder is:

[0073] The decoding process function of the autoencoder is:

[0074] The loss function of the autoencoder is:

[0075] The cost function J is:

[0076] in, θ is a non-linear function, usually the sigmoid function is used; θ is the parameter matrix {W, b} of the weights W and biases b of the encoding network; y = g ( f ( x i )) is the input x i The output predicted value; n l s represents the number of network layers. l For the first l Number of units in the layer; W l ji For the first l Layered network j × i The first is the order of the encoding weight matrix; m is the number of samples; λ is the weight decay coefficient, which controls the weight magnitude and prevents overfitting.

[0077] A denoising autoencoder (DAE) is a neural network model based on unsupervised learning. Its core idea is to reconstruct the original data by applying controllable random noise interference to the input data, thereby learning the essential features of the data. This model effectively achieves feature learning, noise removal, and data dimensionality reduction through a three-stage mechanism of "noise injection - feature extraction - data reconstruction." Compared with traditional autoencoders, DAE significantly improves the robustness of the learned features by forcing the model to recover the original signal from contaminated data, thus enhancing the model's generalization ability. In the modeling process of coal mill systems in thermal power plants, due to the nonlinear characteristics of the coal mill itself, such as large inertia and strong coupling, its operating data is often affected by various uncertainties such as random interference and sensor errors in the industrial environment, exhibiting significant multidimensionality and strong noise characteristics. Specifically, random noise is introduced in two main ways: 1) adding random noise conforming to a Gaussian distribution; 2) using random masking technology, i.e., randomly setting some input features to zero according to a preset ratio. By minimizing reconstruction error, DAE can extract discriminative low-dimensional feature representations from noisy data. It is particularly important to note that the intensity of the added noise is a key parameter affecting model performance. Too little noise is insufficient for regularization, while excessive noise can corrupt the original data structure; both reduce the accuracy of the final model. Therefore, in practical applications, meticulous noise intensity tuning is necessary based on the data characteristics.

[0078] Assuming the input parameter x∈R, adding random noise to x yields corrupted input data elements. The hidden layer features h of these corrupted input data elements, obtained through the encoding process, are represented as follows:

[0079] The decoding process for the hidden layer is the same as that for a regular autoencoder:

[0080] Where z is the decoding feature.

[0081] To prevent overfitting during DAE model training, a regularization term is typically introduced into the loss function. The loss function for DAE is expressed as:

[0082] In the formula, the first term of the loss function is the mean squared error between the original and reconstructed data; the second term is the weight decay term; x and z are the original input value and the reconstructed value of the i-th sample, respectively; λ is the weight decay coefficient; and W represents the weight parameters between neurons in each layer. Gradient descent is usually used to optimize network parameters, and robust feature representation of noisy input data is obtained by minimizing the loss function.

[0083] Stacked Denoising Autoencoder (SDAE) is a deep feature learning model built by stacking multiple DAE units layer by layer. Its core advantage lies in its ability to extract multi-level feature representations from data through a hierarchical abstraction mechanism: shallow networks capture local detailed features, while deep networks learn higher-order global features. This hierarchical feature extraction structure enables SDAE to effectively mine deep-seated patterns hidden in the data, thereby obtaining more discriminative feature representations. During model construction, each DAE unit forces the network to recover the original data from distorted inputs through a "noise injection-feature denoising-data reconstruction" learning mechanism. This training method not only enhances the model's robustness to noise and redundant information but also significantly improves the generalization ability of features. Through layer-by-layer stacking, the cascaded network structure formed by SDAE can perform multi-level abstraction of complex nonlinear relationships, where network depth is positively correlated with the degree of feature abstraction. However, the choice of network depth requires weighing the following key factors: (1) the drawbacks of overly deep networks: exponential growth in computational complexity, gradient decay during error backpropagation, and overfitting when training data is insufficient; (2) the limitations of overly shallow networks: insufficient feature abstraction ability, significantly increased reconstruction error, and difficulty in capturing deep dependencies in the data. SDAE training adopts a strategy combining layered pre-training and global fine-tuning. This training method can ensure the effectiveness of feature extraction at each layer and optimize the overall network performance through end-to-end fine-tuning. In practical applications, the optimal network depth needs to be determined through experiments based on constraints such as data complexity, sample size, and computational resources.

[0084] SDAE Model Construction: First, the lowest-level DAE module is trained in an unsupervised manner, using the feature output of the previous DAE layer as the input data for the next layer. Random noise is injected into the input layer to generate noisy samples, and the network parameters of the current layer are optimized with the goal of minimizing the reconstruction error. After completing layer-by-layer pre-training of all DAE layers, a global fine-tuning strategy is used to optimize the overall network. At this point, the original, undamaged data is used as input, and the parameters of each layer of the network are adjusted end-to-end through the backpropagation algorithm, ultimately achieving the goal of minimizing the overall reconstruction error.

[0085] In this embodiment, a time-shifted memory stacked denoising autoencoder (TS-MSDAE model) is constructed based on time-shifted feature vectors and stacked denoising autoencoders, and multivariate data is reconstructed through the TS-MSDAE model.

[0086] In some embodiments, the TS-MSDAE model structure is a stacked noise-reducing autoencoder that includes a multi-layer encoder and a multi-layer decoder. The encoder consists of a bottom layer, a middle layer, and a high layer. The bottom layer is used to capture instantaneous parameter fluctuations, the middle layer is used to fuse cross-parameter nonlinear correlations, and the high layer is used to model long-range causal chains, compressing parameter features into a low-dimensional latent space.

[0087] Figure 3 This is a flowchart of a state monitoring process based on a multi-source data reconstruction according to an embodiment of this application. The mathematical representation of the TS-MSDAE model is as follows: coding:

[0088] decoding:

[0089] Reconstruction error E t :

[0090] in, x TS This is the concatenated multi-timescale feature vector. ε It is Gaussian noise. f θ For encoder functions, θ Let z be the set of weights and biases of the encoder, and z be the encoding result; g Φ For decoder functions, Φ This is the set of weights and biases for the decoder. x ^TS For the decoding result, ω i This is the result of dynamic weight calculation based on SHAP values.

[0091] After real-time monitoring data from the coal mill is input, the current state is first fused with historical key nodes through a time-shifted feature construction module to form a high-dimensional feature vector, thus solidifying the dynamic characteristics of fault propagation. This vector is then injected with adaptive Gaussian noise to improve robustness and input into a symmetric stacked encoder-decoder architecture: the encoder abstracts features in three layers progressively; the bottom layer captures instantaneous parameter fluctuations, the middle layer fuses cross-parameter nonlinear correlations, and the top layer models long-range causal chains, ultimately compressing them into a low-dimensional latent space; the decoder symmetrically reconstructs the input features and outputs predicted values. The reconstruction error is weighted by physical rules, and the dynamic threshold module determines the operating status—if an anomaly is triggered, a fault memory matching mechanism is activated, identifying known fault types through latent vector similarity comparison; unmatched anomaly patterns are manually confirmed and updated to the memory, forming a closed-loop learning system. The entire process, through hierarchical feature decoupling and dynamic causal embedding, achieves early fault warning and accurate classification.

[0092] In some embodiments, the method further includes: If the average reconstruction error of the collected normal samples exceeds twice the baseline error within the sliding time window of the second preset time period, the TS-MSDAE model is considered to need to perform incremental learning. The incremental learning process includes: locking the encoder's weight parameters and adjusting the decoder's weight parameters based on the incremental learning model.

[0093] Considering the continuous changes in the operating characteristics of coal mills due to equipment wear, coal type compatibility, and seasonal conditions, a fixed model is difficult to adapt to long-term baseline drift. Therefore, an incremental learning strategy is adopted to dynamically evolve the model parameters, ensuring the accuracy of the diagnostic system throughout its entire lifecycle. During incremental learning, the encoder is frozen to retain the learned fault feature extraction capabilities, and then the decoder is fine-tuned to adapt to the data reconstruction patterns under new operating conditions.

[0094] The parameter freezing mechanism includes encoder weight locking and decoder weight updating.

[0095] The encoder weight W e locking:

[0096] L inc This is the output of the incremental learning model.

[0097] Decoder weights W d The updated formula is:

[0098] The incremental learning model is:

[0099] In the formula, N new Let x be the number of newly added normal samples, λ be the strength of historical knowledge retention, and x be the number of samples added. i This represents the i-th input data sample (i.e., a sample from the newly added dataset). x ^ i The model represents the first i The reconstructed output of each sample, θ j This represents the j-th parameter in the model. θ j * Here, Ω represents the historically optimal parameters, and F represents the protected parameter set. j Let β be the diagonal term of the Fisher information matrix, η be the sparse penalty coefficient, and η be the learning rate. W d ||1 represents the decoder weights W d L1 regularization. λ, Fj The calculation methods for η are as follows:

[0100] In the formula, N old The number of normal samples before the update. L old The result is the incremental model calculation result before the update; η0 is the standard learning rate, and k is the learning rate update correction coefficient, which is determined according to the actual operating status of the unit and is set to 1 by default.

[0101] The update trigger condition is:

[0102] Where the window T = 24 hours, and the threshold τ inc =2×baseline error τ t (See step 104 for the fault judgment threshold calculation method).

[0103] In some embodiments, after determining whether the operating status of the coal mill is abnormal, the method further includes: When the reconstruction error exceeds the fault judgment threshold and triggers an abnormal state signal, feature vectors are extracted from the latent space of the TS-MSDAE model.

[0104] Calculate the similarity between the feature vector and the feature vectors of all known fault prototypes stored in the fault memory.

[0105] If the maximum similarity exceeds the preset matching threshold, the fault type of the current feature vector is considered to be the fault type corresponding to the prototype feature vector with the highest similarity.

[0106] If the maximum similarity does not exceed the preset matching threshold, the current feature vector will be temporarily stored and added to the fault memory bank as a new fault prototype feature vector after manual confirmation.

[0107] The fault memory database addresses the challenge of scarce rare fault samples in industrial scenarios by systematically accumulating and utilizing the feature knowledge of historical fault cases. When a novel anomaly is detected, the latent spatial feature vector generated by the autoencoder is extracted, and rapid classification is performed through similarity matching.

[0108] Optionally, the construction process of the fault memory bank includes: Step S201: When an anomaly is detected (reconstruction error exceeds a threshold), extract the latent vector z. t .

[0109] Step S202, calculate z t With all prototypes z in the memory bank (k) cosine similarity simk .

[0110]

[0111] Step S203, if max k (sim k If the value is greater than or equal to 0.85, the corresponding fault type will be output.

[0112] Step S204, otherwise, z t Store temporarily, and add to memory after manual confirmation.

[0113] Traditional fault diagnosis relies on a fixed sample set, making it unable to handle novel anomaly patterns. The fault memory continuously accumulates the "fault fingerprints" of historical cases, forming an evolving diagnostic knowledge system. The update formula for the fault memory is as follows:

[0114] Where M represents the set of prototype feature vectors, in the current feature vector z new If the maximum cosine similarity with all prototypes is less than 0.85, then z new Add to collection M.

[0115] When a new fault is manually confirmed, the latent spatial feature vector is extracted, redundant prototypes with similarity greater than a preset threshold are merged, and the decision priority of low-frequency matching prototypes is reduced through a weight decay mechanism, forming a fault knowledge graph that grows over time, so that the system's diagnostic coverage increases linearly with runtime.

[0116] The fault memory database optimizes its storage structure by automatically integrating similar cases through clustering algorithms, while also periodically discarding outdated prototypes. Over time, it gradually forms a comprehensive knowledge system covering all fault modes, providing decision support for preventative maintenance and driving a paradigm shift in equipment health management from reactive response to proactive prevention. Figure 4 This is a flowchart of a method for detecting operational faults in a coal mill of a thermal power unit according to an embodiment of this application.

[0117] It should be noted that the steps shown in the above process or in the flowchart of the accompanying figures can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0118] This embodiment also provides a fault detection system for coal mills in thermal power units based on multi-source data reconstruction. This system is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0119] Figure 5 This is a structural block diagram of a coal mill fault detection system for thermal power units based on multi-source data reconstruction, according to an embodiment of this application. Figure 5 As shown, the system includes: The data acquisition module 51 is used to acquire time-series data and historical data of multiple operating parameters of the coal mill during operation. The operating parameters include material flow parameters, equipment status parameters, and air volume and air temperature parameters. The feature extraction module 52 is used to construct time-shift feature vectors for each operating parameter based on the principle of maximizing mutual information and according to time series data and historical data. The data reconstruction module 53 is used to input the time-shifted feature vector into the pre-trained TS-MSDAE model to obtain the reconstruction values ​​of each running parameter; calculate the contribution of each running parameter to the reconstruction error based on the SHAP value to obtain the dynamic weight of each running parameter; and determine the reconstruction error by combining the dynamic weight with the reconstruction value and time series data. The status judgment module 54 is used to determine the fault judgment threshold based on the sliding window statistics of the reconstruction error of normal samples, and to determine whether the operating status of the coal mill is abnormal based on the fault judgment threshold and the reconstruction error.

[0120] In some embodiments, the feature extraction module 52 includes: The curve construction module is used to construct the time-shift mutual information curve for each operating parameter based on time-series data and historical data.

[0121] The time delay determination module is used to determine the target time delay when the mutual information between different operating parameters is maximized, based on the time shift mutual information curve.

[0122] The vector generation module is used to concatenate time-series data and historical data based on the target time delay to obtain a time-shifted feature vector.

[0123] In some embodiments, the state determination module 54 includes: The statistical analysis module is used to calculate the mean and standard deviation of the reconstruction error of normal samples within a sliding time window of a first preset time period. The threshold determination module is used to determine the fault judgment threshold based on the mean, standard deviation, and adaptive sensitivity coefficient, wherein the adaptive sensitivity coefficient is determined according to the stability of the unit's operating conditions.

[0124] In some embodiments, the TS-MSDAE model structure is a stacked noise-reducing autoencoder containing multiple layers of encoders and multiple layers of decoders.

[0125] The encoder consists of a bottom layer, a middle layer, and a high layer. The bottom layer is used to capture instantaneous parameter fluctuations, the middle layer is used to fuse cross-parameter nonlinear correlations, and the high layer is used to model long-range causal chains, compressing parameter features into a low-dimensional latent space.

[0126] In some embodiments, the system further includes: The judgment module is used to determine whether the TS-MSDAE model needs to undergo incremental learning if the average reconstruction error of the collected normal samples exceeds twice the baseline error within a sliding time window of the second preset time period.

[0127] The incremental learning module, used in the incremental learning process, includes: locking the encoder's weight parameters and adjusting the decoder's weight parameters based on the incremental learning model.

[0128] In some embodiments, the system further includes: The vector extraction module is used to extract feature vectors in the latent space of the TS-MSDAE model when the reconstruction error exceeds the fault judgment threshold and triggers an abnormal state signal.

[0129] The similarity calculation module is used to calculate the similarity between the feature vector and the feature vectors of all known fault prototypes stored in the fault memory.

[0130] The fault type determination module is used to determine the fault type of the current feature vector as the fault type corresponding to the prototype feature vector with the highest similarity if the maximum similarity exceeds the preset matching threshold.

[0131] The memory update module is used to temporarily store the current feature vector if the maximum similarity does not exceed the preset matching threshold, and add it to the fault memory as a new fault prototype feature vector after manual confirmation.

[0132] In some embodiments, the data acquisition module 51 includes: The first preprocessing module is used to handle missing values ​​in time series data and historical data using the sliding window mean method.

[0133] The second preprocessing module is used to eliminate the dimensional differences between time-series data and historical data based on standardization processing.

[0134] The third preprocessing module is used to remove outliers from time-series and historical data by filtering out transient interference from sensors.

[0135] Through the aforementioned system, the TS-MSDAE model employs a dual-scale historical state splicing architecture, extracting and fusing features from time-series data at different granularities. This simultaneously captures local details and global trends, effectively addressing the complex fault propagation mechanisms in equipment such as coal mills. Multi-scale feature fusion improves the model's adaptability to noise and operating condition changes, reduces the risk of overfitting, and solves the problem of poor adaptability in traditional coal mill fault diagnosis methods.

[0136] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0137] This embodiment also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0138] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0139] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program: S1 acquires time-series and historical data of multiple operating parameters of the coal mill during operation. The operating parameters include material flow parameters, equipment status parameters, and air volume and temperature parameters.

[0140] S2, based on the principle of maximizing mutual information, constructs time-shift feature vectors for each operating parameter according to time series data and historical data.

[0141] S3. Input the time-shifted feature vector into the pre-trained TS-MSDAE model to obtain the reconstruction values ​​of each running parameter; calculate the contribution of each running parameter to the reconstruction error based on the SHAP value to obtain the dynamic weight of each running parameter; combine the dynamic weight with the reconstruction value and time series data to determine the reconstruction error.

[0142] S4. Determine the fault judgment threshold based on the sliding window statistic of the reconstruction error of normal samples. Based on the fault judgment threshold and the reconstruction error, determine whether the operating status of the coal mill is abnormal.

[0143] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0144] In one embodiment, Figure 6 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application, such as... Figure 6 As shown, an electronic device is provided, which can be a server, and its internal structure diagram can be as follows. Figure 6 As shown, this electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a fault detection method for coal mills in thermal power units based on multi-source data reconstruction.

[0145] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0146] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0147] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0148] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for fault detection of coal mill in thermal power units based on multivariate data reconstruction, characterized in that, The method includes: The system acquires time-series and historical data of multiple operating parameters of the coal mill during operation, including material flow parameters, equipment status parameters, and air volume and temperature parameters. Based on the principle of maximizing mutual information, a time-shift feature vector of each of the operating parameters is constructed according to the time-series data and the historical data. The time-shifted feature vector is input into the pre-trained TS-MSDAE model to obtain the reconstructed values ​​of each of the running parameters; the contribution of each running parameter to the reconstruction error is calculated based on the SHAP value to obtain the dynamic weight of each running parameter; the reconstruction error is determined by combining the dynamic weight with the reconstructed values ​​and the time series data. The fault judgment threshold is determined based on the sliding window statistics of the reconstruction error of normal samples. Based on the fault judgment threshold and the reconstruction error, it is determined whether the operating status of the coal mill is abnormal.

2. The method according to claim 1, characterized in that, The construction of time-shift feature vectors for each operating parameter based on the principle of maximizing mutual information, according to the time-series data and the historical data, includes: Based on the time-series data and the historical data, construct the time-shift mutual information curve for each operating parameter; Based on the time-shift mutual information curve, determine the target time delay when the mutual information between different operating parameters is maximized; Based on the target time delay, the time-series data and the historical data are concatenated to obtain the time-shift feature vector.

3. The method according to claim 1, characterized in that, The method for determining the fault judgment threshold based on the sliding window statistic of normal sample reconstruction error includes: Within a sliding time window of the first preset time period, calculate the mean and standard deviation of the reconstruction error of normal samples; The fault judgment threshold is determined based on the mean, the standard deviation, and the adaptive sensitivity coefficient, wherein the adaptive sensitivity coefficient is determined based on the stability of the unit's operating conditions.

4. The method according to claim 1, characterized in that, The TS-MSDAE model structure is a stacked noise-reducing autoencoder containing a multi-layer encoder and a multi-layer decoder; The encoder comprises a bottom layer, a middle layer, and a high layer. The bottom layer is used to capture instantaneous parameter fluctuations, the middle layer is used to fuse cross-parameter nonlinear correlations, and the high layer is used to model long-range causal chains and compress parameter features into a low-dimensional latent space.

5. The method according to claim 4, characterized in that, The method also includes: If the average reconstruction error of the collected normal samples exceeds twice the baseline error within the sliding time window of the second preset time period, the TS-MSDAE model is considered to need to perform incremental learning. The incremental learning process includes: locking the weight parameters of the encoder and adjusting the weight parameters of the decoder based on the incremental learning model.

6. The method according to claim 1, characterized in that, After determining whether the operating status of the coal mill is abnormal, the method further includes: When the reconstruction error exceeds the fault judgment threshold and triggers a state abnormality signal, the feature vector in the latent space of the TS-MSDAE model is extracted. Calculate the similarity between the feature vector and the feature vectors of all known fault prototypes stored in the fault memory bank; If the maximum similarity exceeds the preset matching threshold, the fault type of the current feature vector is considered to be the fault type corresponding to the prototype feature vector with the highest similarity. If the maximum similarity does not exceed the preset matching threshold, the current feature vector is temporarily stored and added to the fault memory bank as a new fault prototype feature vector after manual confirmation.

7. The method according to claim 1, characterized in that, After acquiring time-series and historical data of multiple operating parameters of the coal mill during operation, the method further includes: The sliding window mean method is used to handle missing values ​​in the time series data and the historical data; The dimensional differences between the time-series data and the historical data are eliminated through standardization processing; By filtering out transient interference from sensors, outliers are removed from the time-series data and the historical data.

8. A fault detection system for coal mills in thermal power units based on multivariate data reconstruction, characterized in that, The system includes: The data acquisition module is used to acquire time-series and historical data of multiple operating parameters of the coal mill during operation. The operating parameters include material flow parameters, equipment status parameters, and air volume and air temperature parameters. The feature extraction module is used to construct time-shift feature vectors for each of the operating parameters based on the principle of maximizing mutual information, according to the time-series data and the historical data. The data reconstruction module is used to input the time-shifted feature vector into a pre-trained TS-MSDAE model to obtain the reconstruction values ​​of each of the running parameters; calculate the contribution of each running parameter to the reconstruction error based on the SHAP value to obtain the dynamic weight of each running parameter; and determine the reconstruction error by combining the dynamic weight with the reconstruction value and the time series data. The status judgment module is used to determine the fault judgment threshold based on the sliding window statistics of the reconstruction error of normal samples, and to determine whether the operating status of the coal mill is abnormal according to the fault judgment threshold and the reconstruction error.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the coal mill fault detection method for thermal power units based on multi-source data reconstruction as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the coal mill fault detection method for thermal power units based on multivariate data reconstruction as described in any one of claims 1 to 7.

Images