Method for intelligently identifying working state of hydraulic support

Abstract

The invention discloses a method for intelligently recognizing the working state of a hydraulic support. The method is a method for recognizing the working state of the hydraulic support based on an improved full convolutional network of multi-scale feature fusion. By collecting and preprocessing hydraulic support stand column pressure data and utilizing the multi-scale feature extraction and self-adaptive feature fusion technology, accurate recognition of the working state of a hydraulic support is achieved, and the state conversion period is effectively processed. A soft label mode is adopted for data labeling, and the recognition precision in the state conversion period is improved. The method has the advantages of high practicability, high identification accuracy and high automation level, and is suitable for hydraulic support state monitoring in coal mine safety production.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring, and particularly to an intelligent method and device for identifying the working state of hydraulic supports in mining engineering. Background Art

[0002] In mining engineering, the identification of the state of hydraulic supports is crucial for ensuring the safe production of coal mines. At present, the identification of the working state of hydraulic supports mainly relies on manual experience judgment, lacking an automated judgment method. The existing methods have the following problems: Manual judgment is easily affected by subjective factors, and the judgment results are not accurate enough; it is impossible to monitor the working state of hydraulic supports in real time, making it difficult to detect abnormal situations in a timely manner; it is impossible to provide accurate data support for subsequent fault diagnosis, state prediction, etc. In the prior art:

[0003] Patent CN 112302720 B discloses a method for identifying the initial support force and the cyclic end resistance of hydraulic supports in a working face. This method judges the initial support force and the end resistance through the methods of difference and extreme value identification. However, in practical applications, it is easily affected by abnormal values, resulting in a problem of decreased accuracy. At the same time, this method cannot identify each stage of the working cycle of hydraulic supports.

[0004] Patent CN 110929384 A discloses a real-time analysis system and method for mine pressure big data based on a fully mechanized coal mining face. This method involves a working cycle determination module for supports. However, this module mainly relies on manual optimization and confirms the working cycle by inputting production shifts, which is easily affected by subjective factors, resulting in inaccurate judgment results.

[0005] Patent CN 111911214 B discloses a method for judging the working state of a safety valve by monitoring the pressure change of the hydraulic support column. This method determines the working state of the safety valve by real-time monitoring of the pressure change data of the hydraulic support column and combining with the working characteristics of the safety valve. However, this method mainly focuses on the judgment of the opening and closing states of the safety valve and fails to provide a comprehensive identification of the working state of the hydraulic support, such as the accurate identification of different working states (such as lifting the support, supporting, and lowering the support) and the processing of the state transition period.

[0006] At present, the identification of the working state of hydraulic supports mainly relies on manual experience judgment, lacking an automated judgment method. The existing methods have the following problems: ① Manual judgment is easily affected by subjective factors, and the judgment results are not accurate enough; ② It is impossible to monitor the working state of hydraulic supports in real time, making it difficult to detect abnormal situations in a timely manner; ③ It is impossible to provide accurate data support for subsequent fault diagnosis, state prediction, etc. Summary of the Invention

[0007] Technical problem: The objective of the present invention is to provide a method for intelligently identifying the working state of a hydraulic support, aiming to solve the problems of how to automatically, accurately, and real-time identify the working state of the hydraulic support using pressure data, and how to effectively handle the state transition period.

[0008] Technical solution: A method for intelligently identifying the working state of a hydraulic support according to the present invention includes the following steps:

[0009] (1) Data acquisition: Collect the pressure data of the hydraulic support columns based on the data directly exported from the electro-hydraulic controlled hydraulic support or the data generated by the pressure sensors of the hydraulic support columns. Resample the collected original data at a fixed frequency for the pressure data to ensure the consistency of the time series of the data; for missing values, use linear interpolation to fill the missing values to ensure the continuity of the data.

[0010] (2) Data annotation: Identify the working state of the hydraulic support at each moment based on the collected original pressure data of the columns, and use an encoded form to make the label vector l t ;

[0011] (3) Data preprocessing: Correspond the resampled pressure data with the annotated label data one by one to form a data set. Divide the data set into a training set, a validation set, and a test set, and perform feature scaling on the pressure data respectively, and select maximum-minimum normalization or standardization.

[0012] (4) Feature engineering: Use sliding windows with multiple set radii to extract statistical features, trend features, shape features, and relative position features at each time point in the training set, validation set, and test set respectively, and then use sample cutting windows with a specific length to divide the training set, validation set, and test set into multiple samples respectively.

[0013] (5) Model construction: Construct an improved fully convolutional network, including an input layer, a multi-scale feature extraction layer, an adaptive feature fusion layer, a convolutional network layer, and an output layer.

[0014] (6) Model training: Use the training set processed by feature engineering to train the constructed model, and use the cross-entropy loss function and the Adam optimizer to train the model, and monitor the performance on the validation set, save the optimal model, and use the test set to verify the generalization ability of the model.

[0015] (7) Model inference: After deploying the trained model to the production environment, preprocess the pressure data collected in real-time or extracted from the historical database and perform feature engineering using the same steps, then input the preprocessed data into the trained model for inference to obtain the probability distribution of the state at each time point, and select the state with the highest probability as the prediction result.

[0016] In step (2), the label vector l is made in an encoded form t , for the moments in a single working state, the one-hot encoding is used to label the label vector; for the moments in the state transition period, the linear interpolation method is used to label the label vector, and its formula is:

[0017]

[0018] where the label vector a t in the state transition period depends on the length of the converter. For the state transition period [t s , t s+k , it is calculated by the following formula:

[0019]

[0020] where t represents the current moment, t s represents the start moment of the state transition period, t s+k represents the end moment of the state transition period, and k represents the length of the state transition period.

[0021] In step (3), the formula for the maximum-minimum normalization is:

[0022]

[0023] where x is a data point in the dataset X, and x' is the normalized data.

[0024] The formula for the standardization is:

[0025]

[0026] where μ and σ are the mean and standard deviation of the pressure data in the dataset X respectively.

[0027] In step (4), the sliding window with multiple radii includes two sliding windows: the feature extraction window and the sample cutting window

[0028] The feature extraction window is used to extract the statistical features, trend features, shape features and relative position features of each time point; the sample cutting window is used to divide the time series into multiple samples.

[0029] In step (4), the statistical features of each time point are the mean, standard deviation, and range; the trend features are the first-order difference and the second-order difference; the shape features are the number of wave peaks and the fluctuation intensity; the relative position feature is the relative mean deviation.

[0030] In step (5), the input layer receives input data with a shape of (T, d), where T is the length of the time series and d is the number of features at each time step. The features include the preprocessed pressure data and the features extracted after the feature engineering of the preprocessed pressure data in step (4), and the two jointly form d-dimensional features.

[0031] In step (5), the multi-scale feature extraction layer contains multiple parallel convolutional branches. Each branch uses convolutional kernels of different sizes to extract short-term, medium-term, and long-term time features respectively. The output shape of each parallel convolutional branch is (T, D), where D is the number of convolutional kernels of each branch; the different sizes of the convolution are 3, 7, and 15.

[0032] In step (5), the adaptive feature fusion layer fuses by calculating the weights of each scale feature, and adaptively fuses the multi-scale features into a feature representation with a shape of (T, D) according to the weights. The weights of the fused features are calculated by the following formula:

[0033]

[0034] where, α i is the weight of the i-th scale feature; w i and w j are the learned weight vectors of the i-th and j-th scales respectively, H i and H j represent the features extracted by the i-th and j-th parallel convolutional branches in the multi-scale feature extraction layer respectively; n is the number of parallel convolutional branches; exp() represents the exponential function, which is used to convert the linear weights into positive values for normalization.

[0035] In step (5), the convolutional network layer consists of multiple convolutional layers. It further extracts deep features by shortening the sequence length and increasing the number of feature maps, and then uses transposed convolution to restore the sequence length, so that the output shape of the last convolutional layer is (T, C), where C is the number of working state categories of the hydraulic support.

[0036] In step (5), the output layer uses the Softmax activation function to convert the network output into the probability distribution of the state categories at each time point, realizing the accurate identification of the working state of the hydraulic support.

[0037] Beneficial effects: Due to the adoption of the above technical solutions, the present invention solves the problems in the prior art: ① Manual judgment is easily affected by subjective factors, and the judgment results are not accurate enough; ② It is impossible to monitor the working state of hydraulic supports in real time, and it is difficult to detect abnormal situations in a timely manner; ③ It is impossible to provide accurate data support for subsequent fault diagnosis, state prediction, etc. It can not only accurately identify different working states of hydraulic supports, but also effectively handle the state transition period, greatly improving the accuracy and real-time performance of identification. Compared with the prior art, it has the following advantages:

[0038] (1) Improve identification accuracy and eliminate the influence of subjective factors: Through the improved fully convolutional network with multi-scale feature fusion and the soft label technology, this method can objectively and accurately identify the working state of hydraulic supports, effectively overcoming the inaccurate problem caused by the influence of subjective factors in traditional manual judgment.

[0039] (2) Realize real-time monitoring and abnormal situation discovery: This method supports the real-time identification of the working state of hydraulic supports, can perform state judgment while collecting data, and can quickly discover abnormal states by comparing with the normal working mode, providing decision-making support for taking safety measures in a timely manner.

[0040] (3) Provide high-quality data support and empower intelligent decision-making: This method not only accurately identifies the working state of hydraulic supports, but also precisely depicts the state transition process through the soft label technology, providing a high-quality data basis for the fault diagnosis, life prediction and preventive maintenance of hydraulic supports, and providing reliable support for the intelligent decision-making system.

[0041] (4) Strong practicability and easy to implement in engineering: This method can be realized only with pressure data, greatly reducing the complexity and cost of data collection, and is easy to be deployed and applied in actual engineering. Once the system is trained, it can run fully automatically without human intervention, and is suitable for large-scale coal mine site applications.

[0042] (5) Precise processing of the state transition period: By adopting the unique soft label technology, the gradual change process of the working state is precisely depicted, avoiding the sudden change phenomenon of state switching, which is more in line with the actual physical process of the working state change of hydraulic supports, and improving the stability and reliability of the system. Brief Description of the Drawings

[0043] Figure 1 It is a flowchart of the method for identifying the working state of hydraulic supports of the present invention.

[0044] Figure 2 It is a schematic diagram of the overall network architecture in the method for identifying the working state of hydraulic supports of the present invention.

[0045] Figure 3 It is a schematic diagram of the working state identification result of the method for identifying the working state of hydraulic supports of the present invention. Specific implementation mode

[0046] The present invention will be further described below with reference to the embodiments in the accompanying drawings:

[0047] A method for intelligently identifying the working state of a hydraulic support according to the present invention includes the following steps:

[0048] (1) Data acquisition: According to the data directly exported from the electro-hydraulic controlled hydraulic support or the data generated by the pressure sensor of the hydraulic support column, collect the pressure data of the hydraulic support column, resample the collected original data at a fixed frequency for the pressure data to ensure the consistency of the time series of the data; for missing values, use linear interpolation to fill in the missing values to ensure the continuity of the data;

[0049] (2) Data annotation: Identify the working state of the hydraulic support at each moment according to the original column pressure data collected, and make a label vector l in the form of encoding t ; said making the label vector l in the form of encoding t , for the moment in a single working state, use one-hot encoding to label the label vector; for the moment in the state transition period, use linear interpolation to label the label vector, and its formula is:

[0050]

[0051] Among them, the label vector α t of the state transition period depends on the length of the converter. For the state transition period [t s , t s+k , calculate through the following formula:

[0052]

[0053] Among them, t represents the current moment, t s represents the start moment of the state transition period, t s+k represents the end moment of the state transition period, and k represents the length of the state transition period.

[0054] (3) Data preprocessing: Correspond the resampled pressure data with the labeled label data one by one to form a data set, divide the data set into a training set, a validation set and a test set, and perform feature scaling on the pressure data respectively, and select maximum-minimum normalization or standardization; the formula for selecting maximum-minimum normalization is:

[0055]

[0056] Among them, x is a data point in the data set X, and x' is the normalized data.

[0057] The selected standardized formula is as follows:

[0058]

[0059] where μ and σ are the mean and standard deviation of the pressure data in the dataset X, respectively.

[0060] (4) Feature engineering: Use sliding windows with multiple set radii to extract statistical features, trend features, shape features, and relative position features at each time point in the training set, validation set, and test set respectively, and then use sample cutting windows with a specific length to divide the training set, validation set, and test set into multiple samples; the sliding windows with multiple set radii include two sliding windows: a feature extraction window and a sample cutting window:

[0061] The feature extraction window is used to extract statistical features, trend features, shape features, and relative position features at each time point; the sample cutting window is used to divide the time series into multiple samples.

[0062] The statistical features at each time point are the mean, standard deviation, and range; the trend features are the first-order difference and second-order difference; the shape features are the number of wave peaks and the fluctuation intensity; the relative position feature is the relative mean deviation.

[0063] (5) Model construction: Construct an improved fully convolutional network, including an input layer, a multi-scale feature extraction layer, an adaptive feature fusion layer, a convolutional network layer, and an output layer; the input layer receives input data with a shape of (T, d), where T is the time series length and d is the number of features at each time step. The features include the preprocessed pressure data and the features extracted after the pressure data is preprocessed through the feature engineering in step (4). The two together form d-dimensional features.

[0064] The multi-scale feature extraction layer contains multiple parallel convolutional branches. Each branch uses convolutional kernels of different sizes to extract short-term, medium-term, and long-term time features respectively. The output shape of each parallel convolutional branch is (T, D), where D is the number of convolutional kernels in each branch; the different sizes of the convolution are 3, 7, and 15. The adaptive feature fusion layer fuses by calculating the weights of each scale feature and adaptively fuses the multi-scale features into a feature representation with a shape of (T, D). The weights of the fused features are calculated by the following formula:

[0065]

[0066] where α i is the weight of the i-th scale feature; w i and w j are the learned weight vectors of the i-th and j-th scales respectively, H i and Hj respectively represent the features extracted by the i-th and j-th parallel convolutional branches in the multi-scale feature extraction layer; n is the number of parallel convolutional branches; exp() represents the exponential function, which is used to convert the linear weights into positive values for normalization.

[0067] The convolutional network layer is composed of multiple convolutional layers, which further extract deep features by shortening the sequence length and increasing the number of feature maps, and then use transposed convolution to restore the sequence length, so that the output shape of the last convolutional layer is (T, C), where C is the number of working state categories of the hydraulic support.

[0068] The output layer uses the Softmax activation function to convert the network output into the probability distribution of the state category at each time point, realizing the accurate identification of the working state of the hydraulic support.

[0069] (6) Model training: Use the training set processed by feature engineering to train the constructed model, adopt the cross-entropy loss function and the Adam optimizer to train the model, monitor the performance on the validation set, save the optimal model, and use the test set to verify the generalization ability of the model;

[0070] (7) Model inference: After deploying the trained model to the production environment, preprocess the pressure data collected in real time or extracted from the historical database and perform feature engineering using the same steps, then input the trained model for inference to obtain the probability distribution of the state at each time point, and select the state with the highest probability as the prediction result.

[0071] The following is a detailed description of the specific implementation of the invention, covering key steps such as data collection, preprocessing, feature engineering, model training and inference.

[0072] (1) Data collection

[0073] Collect the pressure data of the hydraulic support columns, and the data collection frequency is set according to actual needs (such as 1Hz). The data source can be the data directly exported from the electro-hydraulic controlled hydraulic support or the data generated by the pressure sensors of the hydraulic support columns.

[0074] (2) Data resampling and interpolation

[0075] Resample the collected original data at a fixed frequency (such as 1 minute or 5 minutes) to ensure the consistency of the time series of the data. For missing values, use the linear interpolation method to fill them to ensure the continuity of the data.

[0076] (3) Dataset division

[0077] The dataset is divided into a training set, a validation set, and a test set in sequence to prevent data leakage during the preprocessing process. The dataset ratio can be selected according to the actual task, such as 7:2:1.

[0078] (4) Data Scaling

[0079] Select the maximum-minimum normalization or standardization method to process the pressure data according to the application scenario.

[0080] (5) Feature Engineering

[0081] For each dataset, feature extraction is performed for each time point. Multidimensional features are extracted from a feature extraction window with a radius of w (such as 4), such as: statistical features such as mean, standard deviation, maximum value, minimum value, range, and quantiles (25%, 50%, 75%); trend features such as first-order difference and second-order difference; shape features such as the number of peaks, the number of valleys, and the fluctuation intensity; relative position features such as relative mean deviation and relative extreme deviation.

[0082] The feature vector for each time point t consists of the features extracted above and the pressure value, and is used as the input to the model.

[0083] The calculation formulas for some features are as follows:

[0084] Mean:

[0085] Standard Deviation:

[0086] Maximum and Minimum: Max t = max(X t ), Min t = min(X t )

[0087] Range: t = Max t - Min t

[0088] First-order Difference:

[0089] Second-order Difference:

[0090] Number of Peaks:

[0091] Number of Valleys:

[0092] Fluctuation Intensity:

[0093] Relative Mean Deviation:

[0094] Relative extreme value deviation:

[0095] (6) Model architecture

[0096] ① Input layer: The shape of the input data is (T, d), where T is the length of the time series and d is the feature dimension.

[0097] ② Multi-scale feature extraction layer: Multiple parallel convolutional layers are used for feature extraction. For example, three parallel convolutional layers with kernel sizes of 3, 7, and 15 are used to extract short-term, medium-term, and long-term features respectively. The output shape of each parallel convolutional layer is (T, D), and the output shape of the multi-scale feature extraction layer is (T, 3*D), where D is the number of convolutional kernels in each parallel convolutional layer, and can be selected as 32 or 64 for example.

[0098] ③ Adaptive feature fusion layer: Calculate the weights of features at each scale and fuse the features according to the weights, with the output shape of (T, D).

[0099] ④ Convolutional network layer: It contains multiple convolutional layers. First, deep features are further extracted by shortening the sequence length and increasing the number of feature maps. Then, transposed convolution is used to restore the sequence length, and the output shape of the last convolutional layer is (T, C), where C is the number of state categories, which is 3 in this embodiment.

[0100] ⑤ Output layer: Use the Softmax activation function to output the state probability distribution at each time point.

[0101] (7) Model training

[0102] The cross-entropy loss function is used, and Adam is used as the optimizer for model training. Monitor the model performance (such as accuracy, recall, and F1 score, etc.) on the validation set to prevent overfitting, and save the model with the best performance on the validation set.

[0103] (8) Model inference

[0104] The newly collected pressure data is preprocessed in the same way as the training data, including resampling, interpolation, and scaling. The preprocessed data is input into the trained model, and the model outputs the state probability distribution at each time point. The state with the highest probability is selected as the prediction result. Output the predicted state at each time point, which can be used for subsequent anomaly detection, cycle identification, and other tasks.

Claims

1. A method for intelligently identifying the working status of a hydraulic support, characterized in that: The following steps are involved: (1) Data acquisition: The pressure data of the hydraulic support column is collected based on the data directly exported by the electro-hydraulic hydraulic support or the data generated by the pressure sensor of the hydraulic support column. The collected original data is resampled according to a fixed frequency to ensure the time series consistency of the data; for missing values, linear interpolation is used to fill the missing values ​​to ensure data continuity; (2) Data labeling: Identify the working state of the hydraulic support at each moment based on the collected column pressure raw data, and use the encoding form to create a label vector l t ; (3) Data preprocessing: The resampled pressure data is matched with the annotated label data one by one to form a data set, which is divided into a training set, a validation set, and a test set. The pressure data is feature scaled and the maximum and minimum normalization or standardization is selected. (4) Feature engineering: Use sliding windows with set radii to extract statistical features, trend features, shape features, and relative position features at each time point in the training set, validation set, and test set. Then use a sample cutting window of a specific length to divide the training set, validation set, and test set into multiple samples. (5) Model construction: construct an improved full convolutional network, including input layer, multi-scale feature extraction layer, adaptive feature fusion layer, convolutional network layer and output layer; (6) Model training: Use the training set that has been processed by feature engineering to train the constructed model, use the cross entropy loss function and Adam optimizer to train the model, monitor the performance on the validation set, save the optimal model, and use the test set to verify the generalization ability of the model; (7) Model reasoning: After the trained model is deployed to the production environment, the pressure data collected in real time or extracted from the historical database is preprocessed and feature engineered using the same steps. The trained model is then input for reasoning to obtain the probability distribution of the state at each time point, and the state with the highest probability is selected as the prediction result.

2. The method for intelligently identifying the working status of a hydraulic support according to claim 1, characterized in that: In step (2), the label vector l is generated in the form of encoding. t , for the moment in a single working state, the label vector is annotated using one-hot encoding; for the moment in the state transition period, the label vector is annotated using linear interpolation, and the formula is: Among them, the label vector α of the state transition period t The value of depends on the length of the converter. s ,t s+k ], calculated by the following formula: Among them, t represents the current time, t s Indicates the starting time of the state transition period, t s+k represents the end time of the state transition period, and k represents the length of the state transition period.

3. The method for intelligently identifying the working status of a hydraulic support according to claim 1, characterized in that: In step (3), the formula for selecting the maximum and minimum normalization is: Where x is a data point in the dataset X and x' is the normalized data. The formula for selecting standardization is: Where μ and σ are the mean and standard deviation of the pressure data in the dataset X, respectively.

4. The method for intelligently identifying the working status of a hydraulic support according to claim 1, characterized in that: In step (4), the sliding windows with multiple radii are set to include two sliding windows: a feature extraction window and a sample cutting window: The feature extraction window is used to extract statistical features, trend features, shape features and relative position features at each time point; The sample cutting window is used to divide the time series into multiple samples.

5. A method for intelligently identifying the working status of a hydraulic support according to claim 1 or 4, characterized in that: In step (4), the statistical characteristics of each time point are mean, standard deviation, and range; the trend characteristics are first-order difference and second-order difference; the shape characteristics are the number of peaks and fluctuation intensity; and the relative position characteristics are relative mean deviation.

6. The method for intelligently identifying the working status of a hydraulic support according to claim 1, characterized in that: In step (5), the input layer receives input data of the shape of (T, d), where T is the length of the time series and d is the number of features for each time step. The features include the preprocessed pressure data and the features extracted from the preprocessed pressure data after feature engineering in step (4), and the two together constitute a d-dimensional feature.

7. The method for intelligently identifying the working status of a hydraulic support according to claim 1, characterized in that: In step (5), the multi-scale feature extraction layer includes multiple parallel convolution branches, each branch uses convolution kernels of different sizes to extract short-term, medium-term and long-term time features respectively, and the output shape of each parallel convolution branch is (T, D), where D is the number of convolution kernels of each branch; the different sizes of convolutions are 3, 7, and 15.

8. The method for intelligently identifying the working status of a hydraulic support according to claim 1, characterized in that: In step (5), the adaptive feature fusion layer performs fusion by calculating the weights of the features at each scale, and adaptively fuses the multi-scale features into a feature representation of shape (T, D) according to the weights. The weights of the fused features are calculated by the following formula: Among them, α i is the weight of the i-th scale feature; w i and w j are the weight vectors of the i-th and j-th scales learned, respectively, and H i and h j They respectively represent the features extracted by the i-th and j-th parallel convolution branches in the multi-scale feature extraction layer; n is the number of parallel convolution branches; exp() represents the exponential function, which is used to convert linear weights into positive values ​​for normalization.

9. The method for intelligently identifying the working state of a hydraulic support according to claim 1, characterized in that: In step (5), the convolutional network layer is composed of multiple convolutional layers, which further extracts deep features by shortening the sequence length and increasing the number of feature maps, and then uses transposed convolution to restore the sequence length, so that the output shape of the last convolution layer is (T, C), where C is the number of hydraulic support working state categories.

10. The method for intelligently identifying the working status of a hydraulic support according to claim 1, characterized in that: In step (5), the output layer uses the Softmax activation function to convert the network output into a state category probability distribution at each time point, thereby achieving accurate identification of the working state of the hydraulic support.

Images