A wind speed abnormal fluctuation detection method based on SSA and LSTM network
By combining singular spectrum analysis and LSTM network, the method of decomposing and predicting mine wind speed data solves the problem of detecting abnormal wind speed fluctuations, realizes efficient identification of abnormal wind speed fluctuations, and improves the early warning capability of mine ventilation system.
Patent Information
- Application Number
- CN202310893023.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-20
- Publication Date
- 2025-11-04
- Estimated Expiration
- 2043-07-20
AI Technical Summary
Existing technologies cannot effectively detect abnormal fluctuations in wind speed in mine ventilation systems caused by the opening and closing of air doors, cages, and the movement of mine cars, resulting in high rates of missed and false alarms, which makes early warning and control difficult.
Singular Spectral Analysis (SSA) is used to decompose and denoise wind speed data. Long Short-Term Memory (LSTM) neural network is used to detect abnormal wind speed fluctuations. An LSTM wind speed reconstruction model is constructed to train and predict wind speed data. A threshold is set using the logarithmic probability density function to detect anomalies.
It effectively identifies short-term wind speed fluctuations caused by the opening and closing of air doors, cages, and the operation of mine cars, reducing the false alarm rate and improving the reliability of coal mine early warning and control.
Smart Images
Figure BDA0004349743970000021 
Figure BDA0004349743970000022 
Figure BDA0004349743970000026
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of mine ventilation, and particularly relates to a wind speed abnormal fluctuation detection method based on SSA and an LSTM network. BACKGROUND
[0002] The opening and closing of the air door will cause the fluctuation of the air flow of the ventilation system and the associated roadway. When the air door is in a static state (remains closed or remains open), that is, under the condition of no disturbance, the wind speed sensor monitoring data presents aperiodic, nonlinear, small amplitude fluctuation under the action of turbulent pulsation. However, in the process of opening and closing the air door, the wind speed sensor monitoring data of the associated roadway will appear short-time abnormal fluctuation. Due to the action of turbulent pulsation, there is a large amount of noise in the wind speed sensor monitoring data. The traditional statistical method mainly relying on manual monitoring cannot accurately detect the abnormal fluctuation in the wind speed sensor monitoring data, and there is a problem of high false negative rate and false positive rate, which brings great difficulty to early warning and control, and therefore it is necessary to detect the wind speed abnormal fluctuation as soon as possible.
[0003] The air flow of the roadway in the mine is a dynamic flow process. When the mine ventilation system fails or a mine car, cage or other conditions that break the dynamic balance occur in the roadway, there will be insufficient air volume and other abnormal conditions, and the data monitored by the monitor will appear abnormal fluctuation. The turbulent data of the roadway wind speed collected by the monitor is a kind of aperiodic, nonlinear time series data. At present, anomaly detection and identification are widely researched and applied in the fields of power quality signals and speech signals, and the neural network artificial intelligence method is often used to detect and identify the abnormality.
[0004] At present, the singular spectrum analysis method can suppress the noise of the nonlinear non-stationary time series data. For this abnormal data fluctuation phenomenon, the detection method of machine learning and neural network can be used to realize data anomaly detection and identification. The long short-term memory neural network (LSTM) has superior performance in this respect, solves the problem of long-term dependence of data, and compared with the recurrent neural network (RNN), the LSTM can effectively solve the problem of gradient disappearance or gradient explosion caused by too large data. Therefore, it is necessary to propose a method combining singular spectrum analysis and LSTM for wind speed abnormal fluctuation detection. SUMMARY
[0005] In view of the above technical problems, the present application provides a wind speed abnormal fluctuation detection method based on SSA and an LSTM network, which can effectively monitor the short-time wind speed fluctuation caused by the opening and closing of the air door, the operation of the cage and the mine car, and can also monitor the air flow in the intelligent mine in real time.
[0006] To achieve the above purpose, the technical scheme adopted by the present application is as follows:
[0007] A wind speed abnormal fluctuation detection method based on SSA and LSTM network, comprising the following steps:
[0008] Step 1: Collect wind speed data monitored by a wind speed sensor, decompose the obtained data into trend components, periodic components and noise components by using singular spectrum analysis method SSA, and reconstruct the trend components and periodic components to achieve the purpose of noise reduction;
[0009] Step 1.1: Collect wind speed fluctuation data with and without wind door opening and closing respectively, and regard the wind speed data without wind door opening and closing as undisturbed data; and regard the wind speed data with wind door opening and closing as disturbed data;
[0010] Step 1.2: Denoise the collected wind speed data;
[0011] Map one-dimensional original data [x1, x2…x N ] with length N to a time window with length L to form K vectors with length L, wherein K=N-L+1, and the trajectory matrix is as follows:
[0012]
[0013] Step 1.3: Singular value decomposition SVD is performed on the obtained trajectory matrix:
[0014]
[0015] In the formula, d is the number of non-zero singular values, d=rank(x)≤min(L,K); λ1, λ2, λ3,…, λ d are singular values of the trajectory matrix X arranged in descending order, U i , V i represent the left and right singular vectors of X;
[0016] Step 1.4: In order to remove redundant noise points in the data, the trajectory matrix X is regarded as the sum of useful signal S and noise E, that is, X=S+E, then S is the effective data after removing noise data E; and it is considered that the data composed of the first r singular values corresponds to the useful signal, and the data corresponding to d-r is identified as the noise component;
[0017] Step 1.5: Diagonal averaging is used to convert the matrix composed of the useful signal in step 1.4 into the required sequence RC with length N, and let Y∈R L×K represent any matrix obtained after grouping, y ij (1≤i≤L,1≤j≤K) are elements in the matrix; let L * =min(L,K), K * =max(L,K), when L , otherwise, Then the matrix Y is transformed into a sequence of desired length N by diagonal averaging The correlation formula is as follows:
[0018]
[0019] In the formula, is the kth diagonal-averaged transformed sequence, is each element in the matrix;
[0020] Step 2: Construct an LSTM wind speed reconstruction model, use the denoised undisturbed data in step 1 as the training set to train the model, use the disturbed data as the test set to realize the prediction of the wind speed data at each time point, and obtain the undisturbed data prediction data set and the disturbed data prediction data set composed of each time point;
[0021] Step 2.1: The input layer pre-processes the time series data denoised in step 1, and then divides it into a training set and a test set, and takes the obtained training set and test set as input;
[0022] Step 2.2: Train the hidden layer based on the training set data, update the LSTM neural network weight value through the Adam optimization algorithm, continuously adjust the network layer to obtain the optimal network structure with the best fitting effect, and add a Dropout layer to prevent overfitting;
[0023] Step 2.3: The output layer predicts the input data according to the model trained in the hidden layer, and outputs the predicted wind speed data;
[0024] Step 3: Calculate the difference between the t-1 time point historical data and the t time point wind speed prediction value and the actual wind speed at t time point, take the logarithmic probability density function value of the error as the abnormal score, divide the data by the abnormal score of the overall data, and then set a threshold to divide the abnormal fluctuation interval;
[0025] Step 3.1: Calculate the difference between the t-1 time point historical data and the t time point wind speed prediction value and the actual wind speed at t time point;
[0026] y t = |x t -x′ t |
[0027] In the formula, x t is the actual wind speed at t time, x t ' is the wind speed prediction value at t time based on the historical data before t-1, y t is the prediction error at t time;
[0028] Step 3.2: Calculate the logarithmic probability density of the difference value and take it as an abnormal score, set a threshold value Tau for determining abnormal wind speed, since the training set is no abnormal disturbance data, the abnormal detection threshold is selected as the highest value of the abnormal score in the training set; the abnormal score is:
[0029]
[0030] In the formula, mu represents the mean of the data sample wind speed, and sigma represents the standard deviation of the data sample wind speed;
[0031] Step 3.3: Abnormal detection is performed on the abnormal score data set calculated in step 3.2, since the smaller the value calculated by the logarithmic probability density function, the stronger the fluctuation of the data corresponding to the point, the minimum abnormal score Tau in the no disturbance abnormal score data set is taken as the threshold value for abnormal division of the disturbance abnormal score data set, then the fluctuation amplitude of the original wind speed data corresponding to the abnormal score less than Tau in the disturbance abnormal score data set is greater than the fluctuation amplitude of the wind speed value in the no disturbance data set; that is, the part less than the threshold value in the disturbance abnormal score data set is the abnormal fluctuation section or the damper opening and closing stage.
[0032] The beneficial effects produced by the above technical scheme are:
[0033] The present application provides a kind of based on SSA and LSTM network's wind speed abnormal fluctuation detection method, which is based on the monitoring data of wind speed sensor, reconstructs and detects anomaly by LSTM to data, can effectively identify the short-time abnormal wind speed fluctuation caused by damper opening and closing, cage, mine car operation etc., effectively solve the problem of high false alarm rate and false alarm rate, improve the reliability of coal mine early warning and control. BRIEF DESCRIPTION OF DRAWINGS
[0034] Figure 1 It is a flow chart of the wind speed abnormal fluctuation detection method based on SSA and LSTM network in the present application example.
[0035] Figure 2 It is a wind speed monitor curve graph of no disturbance data in the present application example.
[0036] Figure 3 It is a wind speed sensor monitoring no disturbance data decomposition into trend component, periodic component and noise component and trend component and periodic component reconstruction graph in the present application example.
[0037] Figure 4 It is a wind speed sensor monitoring no disturbance data residual error, abnormal score graph in the present application example.
[0038] Figure 5 It is a wind speed sensor monitoring curve graph of disturbance data in the present application example.
[0039] Figure 6 For the wind speed sensor monitoring disturbance data in the examples of the present application, the trend component, periodic component and noise component are decomposed, and the trend component and periodic component are reconstructed.
[0040] Figure 7 For the wind speed sensor monitoring disturbance data in the examples of the present application, the residual, anomaly score graph is shown.
[0041] Figure 8 For the wind speed sensor monitoring disturbance data in the examples of the present application, the anomaly detection graph is shown. DETAILED DESCRIPTION
[0042] The application is further illustrated below in combination with the drawings and specific implementation examples:
[0043] A wind speed abnormal fluctuation detection method based on SSA and LSTM network, as shown in the accompanying drawings, comprising the following steps: Figure 1
[0044] Step 1: Collect wind speed data monitored by wind speed sensor, and decompose the obtained data into trend component, periodic component and noise component by singular spectrum analysis method SSA, and then reconstruct the trend component and periodic component to achieve the purpose of noise reduction;
[0045] Step 1.1: Collect wind speed fluctuation data at different time nodes, under different wind speeds, and with and without wind door opening and closing, respectively. The wind speed data without wind door opening and closing is regarded as undisturbed data; the wind speed data with wind door opening and closing is regarded as disturbed data; as shown in Figure 2 、 Figure 5 , Figure 2 is the undisturbed wind speed fluctuation data, Figure 5 is the wind door opening and closing data, i.e. the disturbed wind speed fluctuation data.
[0046] Step 1.2: Noise reduction is performed on the collected wind speed data;
[0047] Map the one-dimensional original data [x1, x2…x N ] of length N to a time window of length L to form K vectors of length L, where K=N-L+1, and the trajectory matrix composed of the vectors is as follows:
[0048]
[0049] Step 1.3: Singular value decomposition SVD is performed on the obtained trajectory matrix:
[0050]
[0051] In the formula, d is the number of non-zero singular values, d=rank(x)≤min(L,K); λ1, λ2, λ3,…, λ d are singular values of trajectory matrix X in descending order, U i , V i represent left and right singular vectors of X;
[0052] Step 1.4: In order to remove the redundant noise in the data, the trajectory matrix X is regarded as a sum of useful signal S and noise E, i.e. X = S + E, S is the effective data after removing the noise data E; and it is considered that the data contribution rate of the signal corresponding to the first r singular values accounts for the main component of the data, so it is determined that this part of the data is the useful signal, and the signal corresponding to d-r is determined as the noise component; therefore, selecting a suitable r value can achieve the effect of data and noise separation.
[0053] Step 1.5: Diagonal averaging is used to convert the matrix composed of the useful signal in step 1.4 into a sequence RC of the required length N, let Y ∈ R L×K represents any matrix obtained after grouping, y ij (1≤i≤L,1≤j≤K) are elements in the matrix; let L * = min(L,K), K * = max(L,K), when L < K, otherwise, Then the matrix Y is transformed into a sequence RC of the required length N by diagonal averaging The relevant calculation formula is as follows:
[0054]
[0055] In the formula, is the kth sequence converted by diagonal averaging, is each element in the matrix;
[0056] As shown in Figure 3 , Figure 6 : Figure 3 a is the trend component decomposed from the undisturbed wind speed fluctuation data in the formula; Figure 3 b is the periodic component decomposed from the undisturbed wind speed fluctuation data in the formula; Figure 3 c is the noise component decomposed from the undisturbed wind speed fluctuation data in the formula; Figure 3 d is the wind speed fluctuation reconstruction curve after noise reduction of the undisturbed wind speed fluctuation data in the formula; Figure 6 a is the trend component decomposed from the disturbed wind speed fluctuation data in the formula; Figure 6 b is the periodic component decomposed from the disturbed wind speed fluctuation data in the formula; Figure 6 c is the noise component decomposed from the disturbed wind speed fluctuation data in the formula; Figure 6 d is the wind speed fluctuation reconstruction curve after noise reduction of the undisturbed wind speed fluctuation data in the formula.
[0057] Step 2: Construct the LSTM wind speed reconstruction model, use the denoised undisturbed data in step 1 as the training set to train the model, use the disturbed data as the test set to realize the prediction of the wind speed data at each time point, and obtain the undisturbed data prediction data set and the disturbed data prediction data set composed of each time point;
[0058] Step 2.1: The input layer pre-processes the time series data denoised in step 1, and then divides it into a training set and a test set, and the obtained training set and test set are used as inputs;
[0059] Step 2.2: Train the hidden layer based on the training set data, update the LSTM neural network weight value through the Adam optimization algorithm, continuously adjust the network layer to obtain the optimal network structure with the best fitting effect, and add a Dropout layer to prevent overfitting;
[0060] Step 2.3: The output layer predicts the input data according to the trained model in the hidden layer, and outputs the predicted wind speed data;
[0061] Step 3: Calculate the difference between the predicted value of the wind speed at time t and the actual wind speed at time t based on the historical data before time t-1, use the log probability density function value of the error as the anomaly score, divide the data according to the anomaly score of the overall data, and then set a threshold to divide the abnormal fluctuation interval;
[0062] Step 3.1: Calculate the difference between the predicted value of the wind speed at time t and the actual wind speed at time t based on the historical data before time t-1;
[0063] y t =|x t -x′ t |
[0064] In the formula, x t is the actual wind speed at time t, x t ' is the predicted value of the wind speed at time t based on the historical data before time t-1, and y t is the prediction error at time t;
[0065] Step 3.2: Calculate the log probability density of the difference as the anomaly score, and set a threshold τ to determine the abnormal wind speed. Since the training set is undisturbed data, the anomaly detection threshold is selected as the highest value of the anomaly score in the training set. The anomaly score is:
[0066]
[0067] In the formula, μ represents the mean of the data sample wind speed, and σ represents the standard deviation of the data sample wind speed;
[0068] The calculation result is as follows: Figure 4 ,Figure 7 As shown: Figure 4 In the middle, 'a' represents a schematic diagram of the residuals between the undisturbed wind speed fluctuation denoised data and the LSTM prediction data; Figure 4 In the middle b, it is a schematic diagram of the anomaly score obtained by calculating the undisturbed wind speed fluctuation data using the log probability density function; Figure 7 In the middle, 'a' represents a schematic diagram of the residuals between the denoised data with disturbed wind speed fluctuations and the LSTM prediction data. Figure 7 In the middle b, it is a schematic diagram of the anomaly score obtained by calculating the disturbed wind speed fluctuation data using the logarithmic probability density function.
[0069] Step 3.3: Perform anomaly detection on the anomaly score dataset calculated in Step 3.2. Since the smaller the value calculated by the logarithmic probability density function, the stronger the data fluctuation corresponding to that point, the minimum anomaly score value in the undisturbed anomaly score dataset is used as the threshold τ to classify the perturbed anomaly score dataset. The original wind speed data fluctuation amplitude corresponding to values less than τ in the perturbed anomaly score dataset is greater than the wind speed value with the largest fluctuation amplitude in the undisturbed dataset; that is, the portion of the perturbed anomaly score dataset less than the threshold is the anomaly fluctuation segment or the damper opening and closing stage. For example... Figure 8 As shown, the shaded area represents the detected abnormal fluctuation segment (damper opening and closing phase).
Claims
1. A method for detecting abnormal wind speed fluctuations based on SSA and LSTM networks, characterized in that, Includes the following steps: Step 1: Collect wind speed data monitored by wind speed sensors, decompose the obtained data into trend components, periodic components and noise components using singular spectrum analysis (SSA), and then reconstruct the trend components and periodic components to achieve noise reduction. Step 2: Construct an LSTM wind speed reconstruction model. Use the undisturbed data after noise reduction in Step 1 as the training set to train the model, and use the perturbed data as the test set to predict the wind speed data at each time point, and obtain the undisturbed data prediction dataset and the perturbed data prediction dataset composed of each time point. Step 3: Calculate the difference between the predicted wind speed at time t and the actual wind speed at time t based on historical data before time t-1. Use the log probability density function value of this difference as the outlier score. Divide the data into segments based on the outlier scores of the overall data, and then set a threshold to divide the abnormal fluctuation range. Step 3.1: Calculate the difference between the predicted wind speed at time t based on historical data up to time t-1 and the actual wind speed at time t; y t =|x t -x′ t | In the formula, x t Let x′ be the actual wind speed at time t. t The wind speed prediction at time t is based on historical data prior to t-1. t Let be the prediction error at time t; Step 3.2: Calculate the logarithmic probability density of the difference and use it as the anomaly score. Set a threshold τ to determine abnormal wind speeds. Since the training set consists of data without abnormal disturbances, the anomaly detection threshold is selected as the highest anomaly score in the training set. The anomaly score is: In the formula, μ represents the mean wind speed of the data sample, and σ represents the standard deviation of the wind speed of the data sample; Step 3.3: Perform anomaly detection on the anomaly score dataset calculated in Step 3.
2. Since the smaller the value calculated by the log probability density function, the stronger the data fluctuation corresponding to that point, use the smallest anomaly score τ in the undisturbed anomaly score dataset as the threshold to divide the disturbed anomaly score dataset into anomalies. Then, the original wind speed data fluctuation amplitude corresponding to the value less than τ in the disturbed anomaly score dataset is greater than the wind speed value with the largest fluctuation amplitude in the undisturbed dataset; that is, the part of the disturbed anomaly score dataset less than the threshold is the abnormal fluctuation segment or the opening and closing stage of the damper.
2. The wind speed anomaly fluctuation detection method based on SSA and LSTM networks according to claim 1, characterized in that, Step 1 specifically involves: Step 1.1: Collect wind speed fluctuation data when the damper is open and closed and when the damper is not open and closed. The wind speed data when the damper is not open and closed is considered as undisturbed data; the wind speed data when the damper is open and closed is considered as disturbed data. Step 1.2: Noise reduction is applied to the collected wind speed data; Given a one-dimensional original data of length N [x1, x2, ... x...] N Mapping is performed using a time window of length L, resulting in K vectors of length L, where K = N - L + 1. The resulting trajectory matrix is as follows: Step 1.3: Perform Singular Value Decomposition (SVD) on the obtained trajectory matrix: In the formula, d represents the number of non-zero singular values, d = rank(x) ≤ min(L,K); λ1, λ2, λ3, ..., λ d U is the singular value of the trajectory matrix X arranged in descending order. i V i Let X represent the left and right singular vectors; Step 1.4: To remove redundant noise from the data, the trajectory matrix X is considered to be composed of the useful signal S and the noise E, i.e., X = S + E. Then S is the effective data after removing the noise data E. Furthermore, the data composed of the signals corresponding to the first r singular values is considered to be the useful signal, and the signal corresponding to dr is considered to be the noise component. Step 1.5: Transform the matrix of useful signals described in Step 1.4 into a sequence RC of length N using diagonal averaging, let Y∈R L×K y represents any matrix obtained after grouping. ij (1≤i≤L, 1≤j≤K) are the elements of the matrix; let L * =min(L,K), K * =max(L,K), when L < K, otherwise, Then, by diagonal averaging, the matrix Y is transformed into the desired sequence y of length N. rc1 y rc2 …y rcN The relevant calculation formulas are as follows: In the formula, Let k be the sequence that has undergone diagonal averaging transformation. represents the elements in the matrix.
3. The wind speed anomaly fluctuation detection method based on SSA and LSTM networks according to claim 1, characterized in that, Step 2 specifically involves: Step 2.1: The input layer preprocesses the time series data after noise reduction in Step 1, and then divides it into training set and test set, using the resulting training set and test set as input; Step 2.2: Train the hidden layer based on the training set data, update the weights of the LSTM neural network through the Adam optimization algorithm, continuously adjust the number of network layers to obtain the optimal network structure with the best fitting effect, and add a Dropout layer to prevent overfitting. Step 2.3: The output layer predicts the input data based on the model trained in the hidden layer and outputs the predicted wind speed data.