Multi-working-condition scene-oriented coal mill fault root cause rapid identification method

Abstract

The invention discloses a coal mill fault root cause identification method oriented to multiple working condition scenes, belongs to the field of coal mill system fault monitoring and traceability, and aims at the multiple working condition scenes, overcomes the defects of an existing method from the perspective of root cause identification efficiency, and provides a new thought for solving similar root cause identification problems. The invention designs a set of working condition-model mutual matching method framework. In the offline stage, a corresponding model is trained for each working condition, and a model corresponding to the working condition to which the online data belongs is matched for the online data, so that the root cause identification problem in a multi-working-condition scene is effectively solved. According to the method, a multi-sensor causal discovery model and a time sequence feature interference method are combined, so that the causal relationship among a plurality of sensors can be extracted at the same time, pairwise modeling for each pair of sensors is avoided, and the complexity of causal modeling is reduced. The fault-related sensors are directly screened out by using the online deviation score, and an additional causal analysis process is not needed, so that the root cause identification speed is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial process fault root cause tracing, and in particular to a method for quickly identifying the root cause of a coal mill fault in multiple operating scenarios. Background Art

[0002] The power system is an important part of modern society. Whether it operates normally is directly related to the efficiency and quality of people's daily production and life. Thermal power, especially coal-fired power, still accounts for more than 60% of my country's total power generation and plays a major role in the power system. Coal mill is a device that breaks coal blocks and grinds them into coal powder. It is a key component of the coal-fired power system. According to the speed, coal mills can be divided into three types: low speed, medium speed and high speed. According to the working principle, they can be divided into ball mills, roller mills, fan mills and other categories. Because there are many components and sensors inside the coal mill and its working conditions are complex, even if a minor fault occurs, it will cause serious consequences due to the propagation of the fault among multiple components, which brings difficulties to the relevant personnel in troubleshooting and unit maintenance.

[0003] As an important means to ensure process safety, the root cause identification method has attracted the attention of a large number of researchers due to its intuitiveness and explainability. The root cause identification method constructs the causal relationship between each component and sensor after the fault to characterize the propagation path of the fault in the entire industrial process, forms a fault causal graph, and locates the root cause of the fault by finding the root node, thereby providing a reference for relevant personnel to maintain the system and reduce the loss caused by the fault. Early root cause identification research mostly relied on expert experience and process mechanism knowledge, but in actual environments, due to the complexity of the process and the high cost of acquiring knowledge, such methods are often difficult to implement. With the development of computer and sensor technology, the acquisition and storage of process data has become more and more convenient, so data-driven methods are gradually dominating the field of root cause identification. Some widely used methods such as Granger causality test and transfer entropy have been proven to be effective in root cause identification.

[0004] However, existing root cause identification methods still have shortcomings: (1) Existing methods often require the collection of complete fault data first, and then the corresponding analysis and processing of the fault data to obtain a fault cause-effect diagram. This means that when the root cause is identified, the fault has already spread and destroyed, so these methods are not efficient in fault identification. (2) In actual industrial production, due to the constant changes in product specifications and output with market and seasonal demand, as well as the frequent fluctuations in process parameters such as product composition, setting values, and feed ratio, the production process generally has multiple working modes, that is, multi-condition industrial processes. Data from different working conditions have different statistical characteristics, and the causal relationship between working conditions will also be different. Achieving root cause identification of faults in multi-condition processes is a major challenge faced in actual processes. Summary of the invention

[0005] One object of the present invention is to provide a method for quickly identifying the root cause of a coal mill failure in multiple operating scenarios, and the present invention improves the speed of identifying the root cause.

[0006] A method for quickly identifying the root cause of a coal mill fault in multiple operating scenarios according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect historical time series data of the target coal mill when it is operating normally under various working conditions from the monitoring sensors of the target coal mill to form a multi-condition operation offline data set;

[0008] S2. Standardize the offline data set of multiple operating conditions to eliminate the scale differences between the data of different monitoring sensors, and divide the offline data set of multiple operating conditions into multiple segments using a sliding window of length T;

[0009] S3. Use the preprocessed multi-operating condition offline data set to train the preset multi-monitoring sensor causal discovery model to obtain multiple sub-models, each of which corresponds to one operating condition;

[0010] S4. Based on the prediction modules of the trained multiple sub-models, the causal relationship between the monitoring sensors is quantified using the time series feature interference method to form a causal topological network of the coal mill under various working conditions, and the mean and variance of the prediction error for each monitoring sensor are calculated respectively;

[0011] S5. Based on the reconstruction module of the trained multiple sub-models, calculate the reconstruction error control limit for monitoring faults under each working condition;

[0012] S6. using clustering algorithm and Euclidean distance to determine the working condition to which the online sample belongs, and inputting the sample into the sub-model of the corresponding working condition;

[0013] S7. Calculate the reconstruction error and prediction error of the online sample, and determine whether a fault occurs based on the control limit. Once a fault occurs, calculate the deviation score of each monitoring sensor based on the prediction error, and screen out the monitoring sensors related to the fault;

[0014] S8. Combine the causal topology network and the screened fault-related monitoring sensors to identify the root cause node of the fault.

[0015] Optionally, the detection targets of the monitoring sensor in S1 are coal mill current, coal feeder coal quantity, coal mill inlet air pressure, coal mill outlet air and powder pressure, coal mill inlet air temperature, coal mill outlet air and powder temperature, coal mill inlet air volume, coal mill angular primary wind speed, coal mill bearing temperature, coal mill bearing X-direction vibration, coal mill motor winding temperature, coal mill primary air and closed air differential pressure, and roller loading oil pressure;

[0016] The multi-operating condition offline data set is expressed as:

[0017] {X1,X2,...,X M};

[0018] Where M represents the number of operating conditions, X M Represents the data of the Mth working condition. The data of a certain working condition is expressed as:

[0019]

[0020] Among them, L is the length of the time series data, and N is the number of monitoring sensors.

[0021] Optionally, the specific method of the standardized preprocessing of the multi-operating condition offline data set in S2 is:

[0022]

[0023] in, represents the raw data of the i-th monitoring sensor, and Represent their mean and variance respectively, x i Represents the preprocessed data of the i-th monitoring sensor. After being processed by a sliding window of length T, each sample is in the form of There are L-T+1 samples in total.

[0024] Optionally, the multi-monitoring sensor causal discovery model in S3 allocates a reconstruction and prediction module to each monitoring sensor to form a parallel structure. The reconstruction module provides an LSTM autoencoder for the input data of each monitoring sensor. The LSTM autoencoder accepts input samples and uses the output of the hidden layer of the LSTM unit to map the input samples into a vector of fixed dimension:

[0025] h i (t),t=1,2,...,T; i=1,2,...,N1;

[0026] The output of the last hidden layer The dynamic features of the samples are used for reconstruction and prediction: N1 LSTM decoders receive the dynamic features of their corresponding monitoring sensors and reversely reconstruct the original input samples moment by moment; N1 multi-layer perceptrons receive the outputs and predict the values ​​of the corresponding monitoring sensors at the next moment;

[0027] The loss function of the multi-monitoring sensor causal discovery model includes reconstruction loss and prediction loss:

[0028] Loss=Loss r +Loss p ;

[0029]

[0030]

[0031] Among them, Loss r Represents the reconstruction loss term, Loss p represents the prediction loss term, represents the reconstructed value, Represents the predicted value. The data of each working condition is used to train a corresponding multi-monitoring sensor causal discovery sub-model. Each sub-model is responsible for the subsequent fault monitoring and root cause identification tasks of its corresponding working condition.

[0032] Optionally, in S4, the causal relationship between the monitoring sensors is quantified by using the time series feature interference technology. i and monitoring sensors x j , the timing characteristics are disturbed by randomly arranging the monitoring sensors x i Disrupt the timing characteristics to make the monitoring sensor x i No longer includes predictive monitoring sensors x j Valuable information from monitoring sensors x i To monitoring sensor x j The causal strength is calculated as the interference monitoring sensor x i Front and interference monitoring sensors x i After that, the target monitoring sensor x j The increase in prediction error is:

[0033]

[0034]

[0035] Among them, ε j and ε j(i) Respectively represent the disruption of monitoring sensor x i Front and rear model pair monitoring sensors x j The prediction error of ij Indicates that from the monitoring sensor x i To monitoring sensor x j the strength of the causal relationship;

[0036] The complete causal relationship topology network is obtained by executing between each pair of monitoring sensors. Each working condition corresponds to a causal relationship topology network. The mean and variance of the prediction error of the model for each monitoring sensor are calculated without interference:

[0037] [μ(ε1),μ(ε2),...,μ(ε N )];

[0038] [σ(ε1),σ(ε2),...,σ(ε N )];

[0039] Among them, μ(ε i ) and σ(ε i ) are the mean and variance of the prediction error of the model for the training data of the monitoring sensor.

[0040] Optionally, in S5, all training samples under a certain working condition are input into the trained multi-monitoring sensor causal discovery model, and the average reconstruction error is calculated, and the control limit α of the reconstruction error is calculated using kernel density estimation:

[0041]

[0042]

[0043]

[0044] Where MRE(s) represents the mean reconstruction error of the sth sample, is the reconstruction result of the model for the input sample xi(s), f is the kernel density estimation function, S is the number of normal samples used for training, K(·) and h are the Gaussian kernel function and bandwidth respectively, θ is the significance level, and the control limit α is obtained by selecting an appropriate significance level θ, and each operating condition corresponds to a control limit.

[0045] Optionally, in step S6, the offline samples of the two working conditions are clustered into M clusters using a K-means clustering algorithm, and the centroid of each cluster (c1, c2, ..., c M ), by calculating the Euclidean distance between the online data and each centroid, the working condition corresponding to the nearest centroid is determined as the working condition to which the current data belongs, and the data is input into the sub-model corresponding to the working condition to which it belongs.

[0046] Optionally, in step S7, the reconstruction error and prediction error of the online data are calculated. If the reconstruction error is lower than the control limit, the data is normal data. If the reconstruction error is higher than the control limit, it indicates that a fault has occurred.

[0047] To perform root cause identification, filter the set of monitoring sensors affected by the fault: Calculate the deviation score for each monitoring sensor:

[0048]

[0049] Among them, D i represents the deviation fraction of monitoring sensor i, ε' i represents the prediction error of the model for the online fault data of monitoring sensor i;

[0050] Quantitative criteria are used to further identify the monitoring sensors affected by the fault:

[0051] if The monitoring sensor i is affected by the fault, otherwise it is not affected by the fault.

[0052] Optionally, in S8, monitoring sensors affected by the fault are marked in the corresponding causal topology network, unmarked monitoring sensors are removed and a fault directed graph is constructed, and nodes without incoming edges are determined as the root causes of the fault under this working condition.

[0053] The beneficial effects of the present invention are:

[0054] 1. The present invention proposes a method for quickly identifying the root cause of a coal mill failure for multiple operating scenarios. This method addresses the problem of low root cause identification efficiency of existing methods for multiple operating scenarios and provides a new approach to solving similar root cause identification problems.

[0055] 2. The present invention designs a method framework for matching working conditions and models. In the offline stage, the corresponding model is trained for each working condition, and the model corresponding to the working condition to which the online data belongs is matched, which effectively solves the root cause identification problem in multi-working condition scenarios.

[0056] 3. The present invention combines the multi-sensor causal discovery model and the time series feature interference method to simultaneously extract the causal relationship between multiple sensors, avoiding modeling for each pair of sensors, and reducing the complexity of causal modeling. On this basis, the online deviation score is used to directly screen out the fault-related sensors without the need for additional causal analysis, thereby improving the speed of root cause identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0058] Figure 1 It is a flowchart of the execution of a method for quickly identifying the root cause of a coal mill fault for multiple operating conditions proposed by the present invention;

[0059] Figure 2 It is a schematic diagram of a framework of a method for quickly identifying the root cause of a coal mill failure in multiple operating scenarios proposed by the present invention;

[0060] Figure 3 It is a structural schematic diagram of a medium-speed roller coal mill provided by an embodiment of the present invention;

[0061] Figure 4 It is a schematic diagram of the multi-sensor causal discovery model proposed by the present invention;

[0062] Figure 5 This is a graph of coal blocking fault detection results of a medium-speed coal mill under multiple working conditions provided by an embodiment of the present invention;

[0063] Figure 6 is a sensor deviation score result diagram provided by an embodiment of the present invention;

[0064] Figure 7 This is a diagram showing the root cause identification results of a coal blocking failure in a medium-speed coal mill under multiple operating conditions provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0066] refer to Figure 1 A method for quickly identifying the root cause of coal mill failure in multiple operating scenarios includes the following steps:

[0067] S1. Collect historical time series data of the target coal mill when it is operating normally under various working conditions from the monitoring sensors of the target coal mill to form a multi-condition operation offline data set;

[0068] S2. Standardize the offline data set of multiple operating conditions to eliminate the scale differences between the data of different monitoring sensors, and divide the offline data set of multiple operating conditions into multiple segments using a sliding window of length T;

[0069] S3. Use the preprocessed multi-operating condition offline data set to train the preset multi-monitoring sensor causal discovery model to obtain multiple sub-models, each of which corresponds to one operating condition;

[0070] S4. Based on the prediction modules of the trained multiple sub-models, the causal relationship between the monitoring sensors is quantified using the time series feature interference method to form a causal topological network of the coal mill under various working conditions, and the mean and variance of the prediction error for each monitoring sensor are calculated respectively;

[0071] S5. Based on the reconstruction module of the trained multiple sub-models, calculate the reconstruction error control limit for monitoring faults under each working condition;

[0072] S6. using clustering algorithm and Euclidean distance to determine the working condition to which the online sample belongs, and inputting the sample into the sub-model of the corresponding working condition;

[0073] S7. Calculate the reconstruction error and prediction error of the online sample, and determine whether a fault occurs based on the control limit. Once a fault occurs, calculate the deviation score of each monitoring sensor based on the prediction error, and screen out the monitoring sensors related to the fault;

[0074] S8. Combine the causal topology network and the screened fault-related monitoring sensors to identify the root cause node of the fault.

[0075] In this embodiment, the detection targets of the monitoring sensor in S1 are coal mill current, coal feeder coal quantity, coal mill inlet air pressure, coal mill outlet air and powder pressure, coal mill inlet air temperature, coal mill outlet air and powder temperature, coal mill inlet air volume, coal mill angular primary wind speed, coal mill bearing temperature, coal mill bearing X-direction vibration, coal mill motor winding temperature, coal mill primary air and closed air differential pressure, and roller loading oil pressure;

[0076] The offline data set of multiple operating conditions is expressed as:

[0077] {X1,X2,...,X M};

[0078] Where M represents the number of operating conditions, X M Represents the data of the Mth working condition. The data of a certain working condition is expressed as:

[0079]

[0080] Among them, L is the length of the time series data, and N is the number of monitoring sensors.

[0081] In this implementation manner, the specific method of standardizing and preprocessing the multi-operating-condition offline data set in S2 is as follows:

[0082]

[0083] in, represents the raw data of the i-th monitoring sensor, and Represent their mean and variance respectively, x i Represents the preprocessed data of the i-th monitoring sensor. After being processed by a sliding window of length T, each sample is in the form of There are L-T+1 samples in total.

[0084] In this embodiment, the multi-monitoring sensor causal discovery model in S3 allocates a reconstruction and prediction module to each monitoring sensor to form a parallel structure. The reconstruction module provides an LSTM autoencoder for the input data of each monitoring sensor. The LSTM autoencoder accepts input samples and uses the output of the hidden layer of the LSTM unit to map the input samples into a vector of fixed dimension:

[0085] h i (t),t=1,2,...,T; i=1,2,...,N1;

[0086] The output of the last hidden layer The dynamic features of the samples are used for reconstruction and prediction: N1 LSTM decoders receive the dynamic features of their corresponding monitoring sensors and reversely reconstruct the original input samples moment by moment; N1 multi-layer perceptrons receive the outputs and predict the values ​​of the corresponding monitoring sensors at the next moment;

[0087] The loss function of the multi-monitoring sensor causal discovery model includes reconstruction loss and prediction loss:

[0088] Loss=Loss r +Loss p ;

[0089]

[0090]

[0091] Among them, Loss r Represents the reconstruction loss term, Loss p represents the prediction loss term, represents the reconstructed value, Represents the predicted value. The data of each working condition is used to train a corresponding multi-monitoring sensor causal discovery sub-model. Each sub-model is responsible for the subsequent fault monitoring and root cause identification tasks of its corresponding working condition.

[0092] In this implementation, S4 uses the time series feature interference technology to quantify the causal relationship between monitoring sensors. i and monitoring sensors x j , the timing characteristics are disturbed by randomly arranging the monitoring sensors x i Disrupt the timing characteristics to make the monitoring sensor x i No longer includes predictive monitoring sensors x j Valuable information from monitoring sensors x i To monitoring sensor x j The causal strength is calculated as the interference monitoring sensor x i Front and interference monitoring sensors x i After that, the target monitoring sensor x j The increase in prediction error is:

[0093]

[0094]

[0095] Among them, ε j and ε j(i) Respectively represent the disruption of monitoring sensor x i Front and rear model pair monitoring sensors x j The prediction error of ij Indicates that from the monitoring sensor x i To monitoring sensor x j the strength of the causal relationship;

[0096] The complete causal relationship topology network is obtained by executing between each pair of monitoring sensors. Each working condition corresponds to a causal relationship topology network. The mean and variance of the prediction error of the model for each monitoring sensor are calculated without interference:

[0097] [μ(ε1),μ(ε2),...,μ(ε N )];

[0098] [σ(ε1),σ(ε2),...,σ(ε N )];

[0099] Among them, μ(ε i ) and σ(ε i ) are the mean and variance of the prediction error of the model for the training data of the monitoring sensor.

[0100] In this implementation, in S5, all training samples under a certain working condition are input into the trained multi-monitoring sensor causal discovery model, and the average reconstruction error is calculated. The control limit α of the reconstruction error is calculated using kernel density estimation:

[0101]

[0102]

[0103]

[0104] Where MRE(s) represents the mean reconstruction error of the sth sample, is the model's response to the input sample x i (s), f is the kernel density estimation function, S is the number of normal samples used for training, K(·) and h are the Gaussian kernel function and bandwidth, respectively, and θ is the significance level. The control limit α is obtained by selecting an appropriate significance level θ, and each operating condition corresponds to a control limit.

[0105] In this embodiment, the K-means clustering algorithm is used in step S6 to cluster the offline samples of the two working conditions into M clusters, and the centroid of each cluster (c1, c2, ..., c M ), by calculating the Euclidean distance between the online data and each centroid, the working condition corresponding to the nearest centroid is determined as the working condition to which the current data belongs, and the data is input into the sub-model corresponding to the working condition to which it belongs.

[0106] In this embodiment, the reconstruction error and prediction error of the online data are calculated in step S7. If the reconstruction error is lower than the control limit, the data is normal data. If the reconstruction error is higher than the control limit, it indicates that a fault has occurred.

[0107] To perform root cause identification, filter the set of monitoring sensors affected by the fault: Calculate the deviation score for each monitoring sensor:

[0108]

[0109] Among them, D i represents the deviation fraction of monitoring sensor i, ε' i represents the prediction error of the model for the online fault data of monitoring sensor i;

[0110] Quantitative criteria are used to further identify the monitoring sensors affected by the fault:

[0111] if The monitoring sensor i is affected by the fault, otherwise it is not affected by the fault.

[0112] In this implementation, in S8, the monitoring sensors affected by the fault are marked in the corresponding causal topology network, the unmarked monitoring sensors are removed and a fault directed graph is constructed, and the nodes without incoming edges are determined as the root causes of the fault under this working condition.

[0113] Embodiment 1:

[0114] S1: MP265G medium-speed roller coal mill, the actual structure is as follows Figure 3 As shown. The working conditions of the coal mill can be divided into four types according to its workload: 0% to 25%, 25% to 50%, 50% to 75%, and 75% to 100%. This embodiment selects two working conditions of 0% to 25% and 25% to 50% to verify the proposed method. The complete data set comes from 1000 normal sample points with a sensor sampling frequency of 1 second under two working conditions and 600 fault sample points recorded by coal blocking faults. The complete coal mill system includes coal mill, induced fan, air supply fan, air conditioner and other equipment with a total of 172 sensors. In order to ensure the pertinence, this embodiment selects thirteen key sensors: coal mill current, coal feeder coal quantity, coal mill inlet air pressure, coal mill outlet air powder pressure, coal mill inlet air temperature, coal mill outlet air powder temperature, coal mill inlet air volume, coal mill angular primary wind speed, coal mill bearing temperature, coal mill bearing X-direction vibration, coal mill motor winding temperature, coal mill primary air and closed air differential pressure and roller loading oil pressure. The detailed information of the selected sensors is shown in Table 1:

[0115] Table 1 Detailed information of selected key sensors

[0116]

[0117]

[0118] S2: Standardize the sample data of each sensor to eliminate the influence of data scale. Taking sensor i as an example, its data standardization calculation formula is:

[0119]

[0120] in represents the raw data collected from sensor i, x i Represents the preprocessed data. Since the data distribution of the same sensor under different working conditions is different, it is necessary to perform the above standardization process on the sensor data under each working condition.

[0121] In order to construct a vector autoregression model to capture the causal relationship between sensors, the standardized data set is processed into samples in time units with a sliding window of length 10 and step size 1. Each sample is in the form of The length of the sliding window should be able to cover the maximum causal lag between sensors. In this embodiment, T=10 is determined based on expert experience.

[0122] S3: Use normal samples of two working conditions to train the multi-sensor causal discovery model. Figure 4 As shown in the figure, the multi-sensor causal discovery model assigns a reconstruction and prediction module to each sensor, thus forming a parallel structure. The reconstruction module provides an LSTM structured autoencoder for each sensor input data. The LSTM encoder accepts the input data and maps it into a fixed-dimensional vector h using the output of the hidden layer of the LSTM unit. i (t), t=1,2,...,T;i=1,2,...,N。 The output of the last hidden layer H(T)=[h1(t),h2(t),...,h N (t)] is used as the dynamic features of the sample in the corresponding time period for reconstruction and prediction: N LSTM decoders receive the dynamic features of their corresponding sensors and reversely reconstruct the original input samples moment by moment; N multi-layer perceptrons (MLPs) receive H(T) and predict the value of their corresponding sensors at the next moment. Therefore, the loss function of the multi-sensor causal discovery model contains reconstruction loss and prediction loss:

[0123] Loss=Loss r +Loss p

[0124]

[0125]

[0126] Among them, Loss r Represents the reconstruction loss term, Loss p represents the prediction loss term, represents the reconstructed value, Represents the predicted value. The model's encoding, decoding, and prediction processes for each sensor data are independent, and each sensor does not interfere with each other. The essence of the model's prediction process is to construct a separate multivariate nonlinear regression model for each sensor, thereby satisfying the definition of conditional Granger causality:

[0127]

[0128] in Represents the value of sensor i at time t.

[0129] The essence of the model reconstruction process is to ensure that the model has the ability to represent each sensor, so that fault data can be identified in the subsequent online monitoring link. In this embodiment, the parameter settings of the multi-sensor causal discovery model are shown in Table 2.

[0130] Table 2 Parameter settings of multi-sensor causal discovery model

[0131] Parameter name Setting Value LSTM hidden layer dimensions 4 MLP number of layers and dimensions (4*13，8)，(8，1) Training cycle epochs 400 Activation Function ReLu Learning rate 0.001 Batch size 32

[0132] S4: After completing the training of the sub-models corresponding to each working condition, the time series feature interference technology is used to quantify the causal relationship between variables. According to the definition of Granger causality, if a variable A helps to predict another variable B, then variable A is the Granger cause of variable B. Therefore, to explore the causal relationship from A to B, it is necessary to train two regression models for variable B, one of which contains variable A and the other does not contain variable A, and then evaluate the causal relationship based on the prediction errors of the two models for B. i and x j For example, the temporal feature interference is achieved by randomly arranging x i to disrupt its time series features, thus ensuring that it no longer contains the predicted x j Valuable information from sensor x i to x j The causal strength of the interference x is calculated as i Front and rear target sensors x j The degree to which the prediction error increases:

[0133]

[0134]

[0135] where ε j and ε j(i) Represents unshuffled and shuffled x respectively i After that, the model j The prediction error of ij Indicates that from x i to x j strength of causal relationship.

[0136] Thanks to the parallel structure of the model, applying time series feature interference to a certain sensor data will not affect other sensors. Therefore, the causal relationship between multiple sensors can be efficiently captured through one model, avoiding modeling of sensors two by two, which greatly reduces the complexity of the model. For this embodiment, each working condition needs to perform 10 times of time series feature interference and take the average value of the results to avoid the influence of randomness on the results. This embodiment performs the above steps under the two selected working conditions respectively to obtain the causal topological structure of the coal mill under the two working conditions.

[0137] In order to facilitate the subsequent screening of sensors affected by the fault, the mean and variance of the model's prediction error for each sensor are calculated without applying interference:

[0138] [μ(ε1),μ(ε2),...,μ(ε N )]

[0139] [σ(ε1),σ(ε2),...,σ(ε N )]

[0140] where μ(ε i ) and σ(ε i ) are the mean and variance of the model’s prediction error for the training data of sensor i.

[0141] S5: Based on the offline data samples and the reconstruction module of the causal discovery model, the reconstruction error control limit for monitoring whether a fault occurs under each working condition is calculated. It should be emphasized that the control limit of fault monitoring is for all sensors, so each time a sample of all sensors is input, the average value of the reconstruction output of the model is calculated, and the average reconstruction error of each input sample is counted:

[0142]

[0143] Where MRE(s) represents the mean reconstruction error of the sth sample, x i (s) is in the form of a time period of length T = 10, is the model's reconstruction of the input sample. Then the kernel density estimation method is used to estimate the control limit α:

[0144]

[0145]

[0146] Where S represents the number of samples used for training, f is the kernel density estimation function, K(·) and h represent the Gaussian kernel function and bandwidth respectively, and θ represents the significance level. By selecting an appropriate significance level θ, the control limit α can be obtained, and each operating condition corresponds to a control limit. In this embodiment, the significance level is 95%, and the control limits of the two operating conditions are α1=7.57 and α2=8.18 respectively.

[0147] S6: Use K-means clustering algorithm to cluster the offline samples of the two working conditions into two clusters, and record the centroids of the clusters as c1 and c2 respectively. The online data at time t is recorded as Calculate the distance between the data at that moment and the two centroids and If d1>d2, the data belongs to working condition 1, otherwise it belongs to working condition 2. The working condition to which the online data belongs at each moment is calculated in real time, and then the data is divided into time period samples using a sliding window with a length of 10 and a step size of 1, and input into the sub-model of the corresponding working condition.

[0148] S7: Use the sub-model to monitor whether the input online sample is a fault. The average reconstruction error of the t-th input sample is recorded as MRE(t). If MRE(t) is greater than the control limit corresponding to the working condition to which it belongs, the sample is fault data. The fault monitoring results provided in this embodiment are as follows: Figure 5 As shown in the figure, the first 1000 samples are normal data for training, and the last 600 samples are fault data for testing. The fault detection rate FDR and false alarm rate FAR are used to evaluate the fault detection effect of the method:

[0149]

[0150]

[0151] Where TP represents the number of fault samples that are correctly detected, FN represents the number of fault samples that are not detected, FP represents the number of normal samples that are falsely detected as faults, and TN represents the number of normal samples that are correctly identified. The fault detection rate of the proposed method in working condition 1 of this embodiment is 99.5%, and the false alarm rate is 0.2%. The fault detection rate in working condition 2 is 100%, and the false alarm rate is 0.7%.

[0152] Calculate the prediction error of the sub-model for the input online fault sample, and further calculate the deviation score of each sensor:

[0153]

[0154] Where D i represents the deviation fraction of sensor i, ε' i represents the prediction error of the model for the online fault data of sensor i. Then sensor i is affected by the fault, otherwise it is not affected by the fault. The calculation results of the deviation scores of each sensor under the two working conditions provided in this embodiment are as follows: Figure 6 As shown, the dotted line in the figure represents the average value of the deviation scores of all sensors under this condition. Sensors with deviation scores greater than the average value are considered to be affected by the fault, that is, the fault behavior of the sensor is related to the root cause.

[0155] S8: Mark the sensors affected by the fault in the causal topology graph of the corresponding working condition to construct a fault directed graph. The sensor set corresponding to the nodes in the causal topology graph is V n , the set of sensors affected by the fault is V f , the construction criteria of the fault directed graph are as follows:

[0156]

[0157] Among them G f represents the fault directed graph, G n represents the initial causal topology graph, v i Represents the node corresponding to sensor i. Figure 7 The fault directed graphs of the two working conditions provided by this embodiment are shown. Since there is only variable X2 without an incoming edge in both graphs, the sensor corresponding to X2 (coal quantity of the coal feeder) is the root cause of the fault, and the conclusion is consistent with the preset coal blocking fault. The basis of this method is that regardless of whether a fault occurs or not, the causal relationship structure formed by the propagation of material flow and information flow between sensors should remain unchanged, that is, after a fault occurs, it will still propagate along the existing causal path. Compared with using fault data for causal analysis and then identifying the root cause, this method can determine the root cause immediately after the fault is detected, thereby achieving rapid positioning of the root cause of the multi-condition coal mill system failure.

[0158] The present invention proposes a method for quickly identifying the root cause of a coal mill failure for multiple operating scenarios. The method is targeted at multiple operating scenarios, fills the deficiency of low efficiency in identifying the root cause of existing methods, and provides a new idea for solving similar root cause identification problems.

[0159] The present invention designs a method framework for matching working conditions and models. In the offline stage, the corresponding model is trained for each working condition, and the model corresponding to the working condition to which the online data belongs is matched, which effectively solves the root cause identification problem in multiple working condition scenarios.

[0160] The present invention combines the multi-sensor causal discovery model and the time series feature interference method to simultaneously extract the causal relationship between multiple sensors, avoiding modeling for each pair of sensors, and reducing the complexity of causal modeling. On this basis, the online deviation score is used to directly screen out the fault-related sensors without the need for additional causal analysis, thereby improving the speed of root cause identification.

[0161] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for quickly identifying the root cause of coal mill failure in multiple operating scenarios, characterized in that: The steps include: S1. Collect historical time series data of the target coal mill when it is operating normally under various working conditions from the monitoring sensors of the target coal mill to form a multi-condition operation offline data set; S2. Standardize the offline data set of multiple operating conditions to eliminate the scale differences between the data of different monitoring sensors, and divide the offline data set of multiple operating conditions into multiple segments using a sliding window of length T; S3. Use the preprocessed multi-operating condition offline data set to train the preset multi-monitoring sensor causal discovery model to obtain multiple sub-models, each of which corresponds to one operating condition; S4. Based on the prediction modules of the trained multiple sub-models, the causal relationship between the monitoring sensors is quantified using the time series feature interference method to form a causal topological network of the coal mill under various working conditions, and the mean and variance of the prediction error for each monitoring sensor are calculated respectively; S5. Based on the reconstruction module of the trained multiple sub-models, calculate the reconstruction error control limit for monitoring faults under each working condition; S6. using clustering algorithm and Euclidean distance to determine the working condition to which the online sample belongs, and inputting the sample into the sub-model of the corresponding working condition; S7. Calculate the reconstruction error and prediction error of the online sample, and determine whether a fault occurs based on the control limit. Once a fault occurs, calculate the deviation score of each monitoring sensor based on the prediction error, and screen out the monitoring sensors related to the fault; S8. Combine the causal topology network and the screened fault-related monitoring sensors to identify the root cause node of the fault.

2. The method for rapid identification of root causes of coal mill faults in multiple operating scenarios according to claim 1 is characterized in that: The monitoring sensors in S1 detect targets such as coal mill current, coal feeder coal quantity, coal mill inlet air pressure, coal mill outlet air and powder pressure, coal mill inlet air temperature, coal mill outlet air and powder temperature, coal mill inlet air volume, coal mill angular primary wind speed, coal mill bearing temperature, coal mill bearing X-direction vibration, coal mill motor winding temperature, coal mill primary air and closed air differential pressure, and roller loading oil pressure; The multi-operating condition offline data set is expressed as: {X1,X2,...,X M }; Where M represents the number of operating conditions, X M Represents the data of the Mth working condition. The data of a certain working condition is expressed as: Among them, L is the length of the time series data, and N is the number of monitoring sensors.

3. The method for rapid identification of root causes of coal mill faults in multiple operating scenarios according to claim 1 is characterized in that: The specific method of the standardized preprocessing of the multi-operating condition offline data set in S2 is: in, represents the raw data of the i-th monitoring sensor, and Represent their mean and variance respectively, x i Represents the preprocessed data of the i-th monitoring sensor. After being processed by a sliding window of length T, each sample is in the form of There are L-T+1 samples in total.

4. The method for rapid identification of root causes of coal mill faults in multiple operating scenarios according to claim 1 is characterized in that: The multi-monitoring sensor causal discovery model in S3 assigns a reconstruction and prediction module to each monitoring sensor to form a parallel structure. The reconstruction module provides an LSTM autoencoder for the input data of each monitoring sensor. The LSTM autoencoder accepts input samples and uses the output of the hidden layer of the LSTM unit to map the input samples into a vector of fixed dimension: h i (t),t=1,2,...,T;i=1,2,...,N1; The output of the last hidden layer The dynamic features of the samples are used for reconstruction and prediction: N1 LSTM decoders receive the dynamic features of their corresponding monitoring sensors and reversely reconstruct the original input samples moment by moment; N1 multi-layer perceptrons receive the outputs and predict the values ​​of the corresponding monitoring sensors at the next moment. The loss function of the multi-monitoring sensor causal discovery model includes reconstruction loss and prediction loss: Loss=Loss r +Loss p ; Among them, Loss r Represents the reconstruction loss term, Loss p represents the prediction loss term, represents the reconstructed value, Represents the predicted value. The data of each working condition is used to train a corresponding multi-monitoring sensor causal discovery sub-model. Each sub-model is responsible for the subsequent fault monitoring and root cause identification tasks of its corresponding working condition.

5. The method for rapid identification of root causes of coal mill faults in multiple operating scenarios according to claim 1 is characterized in that: In S4, the causal relationship between the monitoring sensors is quantified by using the time series feature interference technology. i and monitoring sensors x j , the timing characteristics are disturbed by randomly arranging the monitoring sensors x i Disrupt the timing characteristics to make the monitoring sensor x i No longer includes predictive monitoring sensors x j Valuable information from monitoring sensors x i To monitoring sensor x j The causal strength is calculated as the interference monitoring sensor x i Front and interference monitoring sensors x i After that, the target monitoring sensor x j The increase in prediction error is: Among them, ε j and ε j(i) Respectively represent the disruption of monitoring sensor x i Front and rear model pair monitoring sensors x j The prediction error of ij Indicates that from the monitoring sensor x i To monitoring sensor x j the strength of the causal relationship; The complete causal relationship topology network is obtained by executing between each pair of monitoring sensors. Each working condition corresponds to a causal relationship topology network. The mean and variance of the prediction error of the model for each monitoring sensor are calculated without interference: [μ(ε1),μ(ε2),...,μ(ε N )]; [σ(ε1),σ(ε2),...,σ(ε N )]; Among them, μ(ε i ) and σ(ε i ) are the mean and variance of the prediction error of the model for the training data of the monitoring sensor.

6. The method for rapid identification of root causes of coal mill faults in multiple operating scenarios according to claim 1 is characterized in that: In S5, all training samples under a certain working condition are input into the trained multi-monitoring sensor causal discovery model, and the average reconstruction error is calculated. The control limit α of the reconstruction error is calculated using kernel density estimation: Where MRE(s) represents the mean reconstruction error of the sth sample, is the model's response to the input sample x i (s), f is the kernel density estimation function, S is the number of normal samples used for training, K(·) and h are the Gaussian kernel function and bandwidth, respectively, and θ is the significance level. The control limit α is obtained by selecting an appropriate significance level θ, and each operating condition corresponds to a control limit.

7. The method for rapid identification of root causes of coal mill faults in multiple operating scenarios according to claim 1 is characterized in that: In step S6, the offline samples of the two working conditions are clustered into M clusters using the K-means clustering algorithm, and the centroid of each cluster (c1, c2, ..., c M ), by calculating the Euclidean distance between the online data and each centroid, the working condition corresponding to the nearest centroid is determined as the working condition to which the current data belongs, and the data is input into the sub-model corresponding to the working condition to which it belongs.

8. The method for rapid identification of root causes of coal mill faults in multiple operating scenarios according to claim 1 is characterized in that: In step S7, the reconstruction error and prediction error of the online data are calculated. If the reconstruction error is lower than the control limit, the data is normal data. If the reconstruction error is higher than the control limit, it indicates that a fault has occurred. To perform root cause identification, filter the set of monitoring sensors affected by the fault: Calculate the deviation score for each monitoring sensor: Among them, D i represents the deviation fraction of monitoring sensor i, ε′ i represents the prediction error of the model for the online fault data of monitoring sensor i; Quantitative criteria are used to further identify the monitoring sensors affected by the fault: if The monitoring sensor i is affected by the fault, otherwise it is not affected by the fault.

9. The method for rapid identification of root causes of coal mill faults in multiple operating scenarios according to claim 1 is characterized in that: In the S8, the monitoring sensors affected by the fault are marked in the corresponding causal topology network, the unmarked monitoring sensors are removed and a fault directed graph is constructed, and the nodes without input edges are determined as the root causes of the fault under this working condition.

Images