A mine ventilation fan anomaly detection system and method based on inverse learning

Abstract

This invention discloses a mine ventilation fan anomaly detection system and method based on inverse learning. The data acquisition module records the ventilation fan's operating status in real time at a set period. The data processing module preprocesses and standardizes the operating data. The model training and inference module uses the preprocessed data to train the model and generate pseudo-labels, while simultaneously establishing a multivariate Gaussian distribution reconstruction error statistical model. The logical inference module generates logical labels and performs consistency checks with the pseudo-labels. When inconsistencies are found, inverse correction is performed, generating corrected labels to filter training data and guide the iterative training of the autoencoder and error modeling process, effectively mitigating bias caused by class imbalance. The switching state recognition module distinguishes between ventilation fan anomalies and switching states. The result output module labels the ventilation fan status at each time point, accurately detecting three different operating conditions of the mine ventilation fan: normal, abnormal, and switching.

Description

Technical Field

[0001] This invention belongs to the field of mine ventilation fan anomaly detection, specifically relating to a mine ventilation fan anomaly detection system and method based on inverse learning. Background Technology

[0002] Coal mining largely relies on underground operations. Due to the enclosed environment underground, toxic, harmful, flammable, and explosive gases easily accumulate, making ventilation a crucial aspect of ensuring safe production. As the core equipment of the ventilation system, mine ventilation fans bear the important responsibility of expelling toxic and harmful gases and maintaining air circulation underground. Therefore, conducting research on abnormal detection of mine ventilation fans is an urgent practical need to ensure their long-term stable operation and prevent major safety accidents.

[0003] Existing fault detection methods for mine ventilation fans have three limitations: First, most studies focus on specific faults or single signal sources, lacking the ability to comprehensively model multimodal and multi-state operating data. However, the vibration, temperature, pressure, and other heterogeneous features of mine ventilation fans are coupled together, and relying on only a single or few features makes it difficult to capture potential anomalies. Second, the data collected during the actual operation of ventilation fans in coal mines consists mostly of normal samples, with very few abnormal samples, resulting in a serious class imbalance problem. This makes traditional machine learning models prone to overfitting, while unsupervised learning methods based on cluster analysis and density estimation often lack targeted optimization and are difficult to effectively distinguish between multiple states. Third, ventilation fans switch states during operation (such as "reverse ventilation"), and the changes in detection data caused by state switching are similar to abnormal behavior, making it impossible for current fault detection methods to distinguish between switching states and fault states. This leads to false alarms and ultimately fails to accurately detect anomalies in ventilation fans. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a mine ventilation fan anomaly detection system and method based on inverse learning, which can effectively solve the above-mentioned technical problems, realize the anomaly detection of ventilation fans, and achieve high accuracy.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a mine ventilation fan anomaly detection system based on inverse learning, comprising a data acquisition module, a data transmission and storage module, a data processing module, a model training and inference module, a logical inference module, a switching state recognition module, and a result output module; the data acquisition module is used to collect mine ventilation fan operation data in real time from sensors deployed around and inside the mine ventilation fan according to a set sampling period; the data transmission and storage module is used to transmit the collected data to the server for storage in real time through an industrial internet platform, ensuring the stability and reliability of data transmission, and then store the data in a database, and establish an indexing mechanism and redundant backup for the data in the server; the data processing module is used to preprocess the collected data and standardize all feature values ​​in the preprocessed data; the model training and inference module is used to generate pseudo-labels during the model training stage, and The training data is filtered based on the corrected labels obtained from the logical reasoning module. The filtered data is used to train the model and reconstruct the error statistics model. During the inference stage, the abnormality score of the data sample to be detected is calculated. The logical reasoning module generates logical labels based on preset domain knowledge rules and performs consistency checks between the logical labels and the pseudo labels generated by the model training and inference module. When inconsistencies are found, the labels are corrected to generate inverse corrected labels, which are used to guide the iterative training of the model by the model training and inference module. The switching state recognition module determines whether the data sample with a high abnormality score calculated by the model training and inference module is in a switching or abnormal state. The result output module marks the ventilation fan status of the data sample to be detected based on the comprehensive judgment results of the model training and inference module and the switching state recognition module, and outputs it to the monitoring platform and server to realize graphical status display and abnormal warning.

[0006] Furthermore, the operating data of the mine ventilation fan includes six monitoring parameters: vibration signal, current, voltage, temperature, air pressure, and air volume.

[0007] Furthermore, in order to effectively improve the ability to identify abnormal states in the context of data class imbalance, the logical label generation strategy in the logical reasoning module includes binary alarm rules, interval constraint rules, and Z-score statistical rules.

[0008] The above-mentioned detection method for an anomaly detection system for mine ventilation fans based on inverse learning includes the following steps:

[0009] Step 1: Data Acquisition and Transmission.

[0010] S11: Acquire the operating data of the mine ventilation fan in real time according to the set sampling period to form the raw dataset. ,in Indicates time of The feature vector is generated and transmitted to the server in real time through the industrial internet platform to complete data reception and storage.

[0011] Step 2: Data preprocessing.

[0012] S21: For the received raw dataset Preprocessing is performed.

[0013] S22: Standardize all feature values ​​in the preprocessed data, converting them to a form with zero mean and unit standard deviation; for each dimension... Its standardization is transformed into:

[0014]

[0015] In the formula, For the first Time of the first One original feature value, and These are the mean and standard deviation of the feature in the training set, respectively. This is a standardized result.

[0016] S23: Standardize the data Divide the window into sections of length using a sliding window mechanism. Step size is The set of subsequences:

[0017]

[0018] In the formula, This represents the set of all subsequences obtained by sliding window partitioning. This indicates the total number of subsequences generated. Indicates the first Subsequence samples.

[0019] Step 3: Model training and error modeling.

[0020] S31: Using Subsequence Sets The training data reconstructs the model, and the data reconstruction model is composed of an encoder. and decoder Composition, the objective function is to minimize the input With reconstruction output The average reconstruction error between them (using mean square error) :

[0021]

[0022]

[0023] In the formula, This represents the standardized input vector. This indicates the output of the corresponding data reconstruction model.

[0024] S32: Calculate the reconstruction error vector for all training samples, and... The error matrix is ​​expanded as follows -Dimensional vector:

[0025]

[0026] In the formula, Indicates the first 1 input subsequence This indicates the results of the reconstruction. This is the reconstruction error vector for the subsequence. For vectorization operations, the two-dimensional matrix is ​​flattened into a one-dimensional column vector.

[0027] S33: Generate pseudo-labels for each training sample. .

[0028] S34: For all error vectors Establish a multivariate Gaussian distribution model and estimate the mean vector. Covariance Matrix Construct the reconstruction error probability density function:

[0029]

[0030] In the formula, The input is the reconstruction error vector. To find the mean of the error vector in the training set, Let covariance matrix be the variance matrix. The determinant of the covariance matrix. This represents the inverse of the covariance matrix.

[0031] Step 4: Logical reasoning and reverse correction.

[0032] S41: Generate logical labels for training samples based on domain knowledge rules. .

[0033] S42: For pseudo-tags and logical tags Perform a consistency check.

[0034] S43: Reverse-correct inconsistent label pairs to generate corrected labels. .

[0035] S44: Correction label Select training data and retain a portion of the samples to form a new training set. The model and the multivariate Gaussian distribution model are then iteratively trained using this data.

[0036] Step 5: Model Inference and Anomaly Score Calculation.

[0037] S51: For the test set subsequence Perform reconstruction and calculate its error vector. Then, the log probability density is calculated using the multivariate Gaussian model established in step four, and its anomaly score is defined. for:

[0038]

[0039] In the formula, Indicates the test sample number The reconstruction error vector of each subsequence This represents the Gaussian probability density function constructed during the training phase; the larger the score, the more likely the sample is to deviate from the normal distribution, i.e., the more likely it is to be an anomaly.

[0040] Step 6: Switch status recognition.

[0041] S61: Center time point of subsequence with high outlier scores Nearby vibration parameters were used to construct three equal-length time windows, namely the first window... Central window and back window .

[0042] S62: The sample corresponding to the time point is determined to be in a switching state when the following three conditions are met:

[0043]

[0044] In the formula, , These represent the average vibration signal values ​​of the front and rear windows, respectively. , , These represent the standard deviations of the three windows, , Upper and lower thresholds for the standard deviation of the two windows are set based on experience; This is the threshold for the mean jump.

[0045] Step 7: Status labeling and result output.

[0046] S71: Combining the inference results of the integrated model with the switching state identification results, each time point is ultimately marked as one of three states: normal, abnormal, or switching, and output to the monitoring platform and server.

[0047] Furthermore, to ensure the accuracy of subsequent data analysis, the preprocessing includes merging duplicate records, removing data samples with too many missing parameters, and filling in missing values.

[0048] Furthermore, in order to effectively fill missing parameters in the data samples, the missing values ​​are filled using a bidirectional nearest neighbor method.

[0049] Furthermore, in order to effectively improve the ability to identify abnormal states in the context of data class imbalance, the data reconstruction model adopts an autoencoder model, and the reconstruction error statistical model adopts a multivariate Gaussian distribution statistical model.

[0050] Compared with existing technologies, this invention utilizes a data acquisition module to record the operating status of the ventilation fan in real time at a set period, ensuring the comprehensiveness and timeliness of data acquisition. The acquired data is uploaded to the server, and a redundant backup mechanism is established to ensure the reliability of data transmission and the security of storage. Next, the data processing module preprocesses and standardizes the raw operating data. The standardization strategy not only improves the model's learning ability across different parameter scales but also significantly accelerates model convergence and enhances the system's adaptability to multi-source heterogeneous data. Subsequently, the model training and inference module uses the preprocessed data to train a data reconstruction model. The data reconstruction model learns deep feature representations under normal operating conditions using an autoencoder structure and generates pseudo-labels accordingly. By calculating the reconstruction error of each sample point and establishing a multivariate Gaussian distribution reconstruction error statistical model, the quantitative calculation of anomaly scores is achieved. Based on this, a logical inference module is introduced, generating logical labels for training samples based on preset domain knowledge rules and performing consistency checks with the pseudo-labels generated by the model. When a contradiction is found between the pseudo-label and the logical label, the system performs a reverse correction to generate a corrected label, ensuring that the label conforms to both data statistical characteristics and domain knowledge constraints. The corrected labels are used to filter training data and guide the iterative training of the autoencoder and error modeling process, thereby effectively mitigating the bias caused by class imbalance and improving the robustness and accuracy of the model. Considering that the "reversing fan" switching operation, which frequently occurs in mines, may cause sudden changes in vibration values ​​and be misjudged as abnormal, the system is designed with a switching status recognition module to avoid misidentifying normal fan rotation as a malfunction. Finally, the results output module integrates the inference results of the autoencoder model with the analysis and judgment of the switching recognition module to accurately label the fan status at each time point and output it to the monitoring platform and server.

[0051] In summary, this invention requires no additional equipment and can be seamlessly integrated into the existing industrial internet and data processing platforms in mines. It also effectively improves the detection capability under conditions of scarce abnormal samples and unbalanced data categories. Furthermore, it can accurately detect three different operating conditions of mine ventilation fans: normal, abnormal, and switching, significantly reducing misjudgments caused by "reverse ventilation fans." Ultimately, this ensures the practicality and stability of this invention in actual operating environments. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the system structure in this invention.

[0053] Figure 2 This is a flowchart of the method in this invention. Detailed Implementation

[0054] The present invention will be further described below.

[0055] like Figure 1As shown, a mine ventilation fan anomaly detection system based on inverse learning includes a data acquisition module, a data transmission and storage module, a data processing module, a model training and inference module, a logic inference module, a switching state recognition module, and a result output module. The data acquisition module is used to collect mine ventilation fan operation data in real time from sensors deployed around and inside the mine ventilation fan according to a set sampling period. The data acquisition module includes six monitoring parameters: vibration signal, current, voltage, temperature, wind pressure, and air volume. The data transmission and storage module transmits the collected data to the server for storage in real time via the industrial internet platform, ensuring the stability and reliability of data transmission. The data is then stored in a database, and an indexing mechanism and redundant backups are established within the server. The data processing module preprocesses the collected data and standardizes all feature values ​​in the preprocessed data. The model training and inference module generates pseudo-labels during model training and filters training data based on corrected labels obtained from the logical inference module. The filtered data is used to train the model, reconstruct the error statistics model, and calculate anomaly scores for the data samples to be detected during the inference phase. The logical inference module generates logical labels based on preset domain knowledge rules and performs consistency checks between the logical labels and the pseudo-labels generated by the model training and inference module. If inconsistencies are found, the labels are corrected to generate inverse corrected labels, which are then used to guide the iterative training of the model by the model training and inference module. To guide the iterative training of the model, the logical label generation strategy includes binary alarm rules, interval constraint rules, and Z-score statistical rules. The switching state identification module is used to determine whether data samples with high abnormal scores calculated by the model training and inference module are in a switching or abnormal state. The result output module marks the ventilator status of the data sample to be detected based on the comprehensive judgment results of the model training and inference module and the switching state identification module, and outputs it to the monitoring platform and server to realize graphical status display and abnormal warning.

[0056] like Figure 2 As shown, the detection method of the above-mentioned mine ventilation fan anomaly detection system based on inverse learning includes the following steps:

[0057] Step 1: Data Acquisition and Transmission.

[0058] S11: Acquire the operating data of the mine ventilation fan in real time according to the set sampling period to form the raw dataset. ,in Indicates time of The feature vector is generated and transmitted to the server in real time through the industrial internet platform to complete data reception and storage.

[0059] Step 2: Data preprocessing.

[0060] S21: For the received raw dataset Preprocessing is performed, including merging duplicate records, removing data samples with too many missing parameters, and filling missing values, where the missing values ​​are filled using a bidirectional nearest neighbor method.

[0061] S22: Standardize all feature values ​​in the preprocessed data, converting them to a form with zero mean and unit standard deviation; for each dimension... Its standardization is transformed into:

[0062]

[0063] In the formula, For the first Time of the first One original feature value, and These are the mean and standard deviation of the feature in the training set, respectively. This is a standardized result.

[0064] S23: Standardize the data Divide the window into sections of length using a sliding window mechanism. Step size is The set of subsequences:

[0065]

[0066] In the formula, This represents the set of all subsequences obtained by sliding window partitioning. This indicates the total number of subsequences generated. Indicates the first Subsequence samples.

[0067] Step 3: Model training and error modeling.

[0068] S31: Using Subsequence Sets The training data reconstruction model employs an autoencoder model, which consists of an encoder... and decoder Composition, the objective function is to minimize the input With reconstruction output The average reconstruction error between them (using mean square error) :

[0069]

[0070]

[0071] In the formula, This represents the standardized input vector. This indicates the output of the corresponding data reconstruction model.

[0072] S32: Calculate the reconstruction error vector for all training samples, and... The error matrix is ​​expanded as follows -Dimensional vector:

[0073]

[0074] In the formula, Indicates the first 1 input subsequence This indicates the results of the reconstruction. This is the reconstruction error vector for the subsequence. For vectorization operations, the two-dimensional matrix is ​​flattened into a one-dimensional column vector.

[0075] S33: Generate pseudo-labels for each training sample. .

[0076] S34: For all error vectors Establish a multivariate Gaussian distribution model and estimate the mean vector. Covariance Matrix Construct the reconstruction error probability density function:

[0077]

[0078] In the formula, The input is the reconstruction error vector. To find the mean of the error vector in the training set, Let covariance matrix be the variance matrix. The determinant of the covariance matrix. This represents the inverse of the covariance matrix.

[0079] Step 4: Logical reasoning and reverse correction.

[0080] S41: Generate logical labels for training samples based on domain knowledge rules. .

[0081] S42: For pseudo-tags and logical tags Perform a consistency check.

[0082] S43: Reverse-correct inconsistent label pairs to generate corrected labels. .

[0083] S44: Correction label Select training data and retain a portion of the samples to form a new training set. The model and the multivariate Gaussian distribution model are then iteratively trained using this data.

[0084] Step 5: Model Inference and Anomaly Score Calculation.

[0085] S51: For the test set subsequence Perform reconstruction and calculate its error vector. Then, the log probability density is calculated using the multivariate Gaussian model established in step four, and its anomaly score is defined. for:

[0086]

[0087] In the formula, Indicates the test sample number The reconstruction error vector of each subsequence This represents the Gaussian probability density function constructed during the training phase; the larger the score, the more likely the sample is to deviate from the normal distribution, i.e., the more likely it is to be an anomaly.

[0088] Step 6: Switch status recognition.

[0089] S61: Center time point of subsequence with high outlier scores Nearby vibration parameters were used to construct three equal-length time windows, namely the first window... Central window and back window .

[0090] S62: The sample corresponding to the time point is determined to be in a switching state when the following three conditions are met:

[0091]

[0092] In the formula, , These represent the average vibration signal values ​​of the front and rear windows, respectively. , , These represent the standard deviations of the three windows, , Upper and lower thresholds for the standard deviation of the two windows are set based on experience; This is the threshold for the mean jump.

[0093] Step 7: Status labeling and result output.

[0094] S71: Combining the inference results of the integrated model with the switching state identification results, each time point is ultimately marked as one of three states: normal, abnormal, or switching, and output to the monitoring platform and server.

[0095] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A mine fan abnormality detection system based on inductive learning, characterized in that, It includes a data acquisition module, a data transmission and storage module, a data processing module, a model training and inference module, a logical inference module, a switching state recognition module, and a result output module; The data acquisition module is used to collect real-time operating data of the mine ventilation fan from sensors deployed around and inside the mine ventilation fan according to a set sampling period. The data transmission and storage module is used to transmit the collected data to the server for storage in real time through the industrial internet platform, and to establish an indexing mechanism and redundant backup for the data within the server. The data processing module is used to preprocess the collected data and standardize all feature values ​​in the preprocessed data. The model training and inference module is used to generate pseudo-labels during the model training phase, and to filter training data based on the corrected labels obtained by the logical inference module, so as to use the filtered data to train the data to reconstruct the model and rebuild the error statistics model, and to calculate the abnormality score of the data sample to be detected during the inference phase. The logical reasoning module is used to generate logical labels based on preset domain knowledge rules, and to perform consistency checks between the logical labels and the pseudo labels generated by the model training and reasoning module. When inconsistencies are found, the labels are corrected to generate inverse corrected labels, and the corrected labels are used to guide the model training and reasoning module to iteratively train the model. The switching state identification module is configured to determine that a data sample with a high abnormal score after calculation by the model training and reasoning module is in a switching or abnormal state, and specifically, the center time point of a subsequence with a high abnormal score is determined as follows Vibration parameters near the center time point are determined Three equal-length time windows are constructed, namely a front window , a center window , and a rear window When the following three conditions are met, the sample corresponding to the time point is determined to be in a switching state, and the determination conditions are as follows: In the formula, , These represent the average vibration signal values ​​of the front and rear windows, respectively. , , These represent the standard deviations of the three windows, , Upper and lower thresholds for the standard deviation of the two windows are set based on experience; The threshold for the mean jump; The result output module, based on the combined judgment results of the model training and inference module and the switching state recognition module, marks the ventilator status of the data sample to be detected and outputs it to the monitoring platform and server to realize graphical status display and abnormal warning.

2. The mine ventilation fan anomaly detection system based on inverse learning according to claim 1, characterized in that, The operating data of the mine ventilation fan includes six monitoring parameters: vibration signal, current, voltage, temperature, air pressure, and air volume.

3. The mine ventilation fan anomaly detection system based on inverse learning according to claim 1, characterized in that, The logical label generation strategy in the logical reasoning module includes binary alarm rules, interval constraint rules, and Z-score statistical rules.

4. A detection method using the inverse learning-based mine ventilation fan anomaly detection system as described in any one of claims 1 to 3, characterized in that, Includes the following steps: Step 1: Data Acquisition and Transmission; S11: Acquire the operating data of the mine ventilation fan in real time according to the set sampling period to form the raw dataset. ,in Indicates time of The feature vector is generated and transmitted to the server in real time through the industrial internet platform to complete data reception and storage. Step 2: Data preprocessing; S21: For the received raw dataset Preprocessing is required; S22: Standardize all feature values ​​in the preprocessed data, converting them to a form with zero mean and unit standard deviation; for each dimension... Its standardization is transformed into: In the formula, For the first Time of the first One original feature value, and These are the mean and standard deviation of the feature in the training set, respectively. For standardized results; S23: Standardize the data Divide the window into sections of length using a sliding window mechanism. Step size is The set of subsequences: In the formula, This represents the set of all subsequences obtained by sliding window partitioning. This indicates the total number of subsequences generated. Indicates the first Subsequence samples; Step 3: Model training and error modeling; S31: Using Subsequence Sets The training data reconstructs the model, and the data reconstruction model is composed of an encoder. and decoder Composition, the objective function is to minimize the input With reconstruction output Average reconstruction error between : In the formula, This represents the standardized input vector. This represents the reconstruction output of the corresponding data reconstruction model; S32: Calculate the reconstruction error vector for all training samples, and... The error matrix is ​​expanded as follows Dimensional vector: In the formula, Indicates the first input subsequences, This indicates the results of the reconstruction. This is the reconstruction error vector for the subsequence. For vectorization operations, the two-dimensional matrix is ​​flattened into a one-dimensional column vector; S33: Generate pseudo-labels for each training sample. ; S34: For all error vectors Establish a multivariate Gaussian distribution model and estimate the mean vector. Covariance Matrix Construct the reconstruction error probability density function: In the formula, The input is the reconstruction error vector. To find the mean of the error vector in the training set, Let covariance matrix be the variance matrix. The determinant of the covariance matrix. The inverse matrix of the covariance matrix; Step Four: Logical Reasoning and Reversal Correction; S41: Generate logical labels for training samples based on domain knowledge rules. ; S42: For pseudo-tags and logical tags Perform a consistency check; S43: Reverse-correct inconsistent label pairs to generate corrected labels. ; S44: Correction label Select training data and retain a portion of the samples to form a new training set. This is used to iteratively train data to reconstruct the model and the multivariate Gaussian distribution model; Step 5: Model Inference and Anomaly Score Calculation; S51: For the test set subsequence Perform reconstruction and calculate its error vector. Then, the log probability density is calculated using the multivariate Gaussian model established in step four, and its anomaly score is defined. for: In the formula, Indicates the test sample number The reconstruction error vector of each subsequence This represents the Gaussian probability density function constructed during the training phase. Step Six: Switch Status Recognition; Step 7: Status labeling and result output; S 71: Combining the inference results of the integrated model with the switching state identification results, each time point is ultimately marked as one of three states: normal, abnormal, or switching, and output to the monitoring platform and server.

5. The detection method according to claim 4, characterized in that, The preprocessing includes merging duplicate records, removing data samples with too many missing parameters, and filling in missing values.

6. The detection method according to claim 5, characterized in that, The missing values ​​are filled using a two-way nearest neighbor method.

7. The detection method according to claim 4, characterized in that, The data reconstruction model adopts an autoencoder model, and the reconstruction error statistical model adopts a multivariate Gaussian distribution statistical model.

Images