Method for predicting blast furnace TRT equipment maintenance time based on LSTM model

Abstract

The invention discloses a method for predicting blast furnace TRT equipment maintenance time based on an LSTM model, and the method comprises the steps: constructing a sliding window adaptive algorithm based on vibration signal change characteristics, and dynamically adjusting the size of an input data sliding window of the LSTM model according to the fluctuation amplitude and change rate of a vibration signal; adopting an adaptive sliding window to obtain TRT equipment training data, and training the LSTM model; applying the trained LSTM model to the vibration signals acquired in real time, predicting the vibration trend of the TRT equipment in the future n days at the current point, and outputting a predicted value of the vibration signals in a future period of time; and according to the prediction result of the vibration signal, in combination with a preset threshold value of shutdown maintenance, judging the time when the TRT equipment reaches the threshold value, and further obtaining the remaining days of the TRT equipment away from the maintenance. According to the method, the future development trend of the TRT equipment vibration signal is accurately predicted, the maintenance plan is reasonably arranged, and the TRT equipment operation reliability and production efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of data application technology, and in particular to a method for predicting the maintenance time of blast furnace TRT equipment based on an LSTM model. Background Art

[0002] In the modern industrial field, the stable operation of equipment is a key factor in ensuring production efficiency and safety. With the improvement of industrial automation, the remote intelligent operation and maintenance platform of equipment has gradually become an important tool for enterprises to optimize equipment status management. The huge amount of data on the intelligent operation and maintenance platform provides a rich data foundation for monitoring the operating status of equipment and fault diagnosis. However, it is difficult to fully realize its potential value by simply storing and simply analyzing massive data. How to effectively use this data for intelligent prediction and provide a scientific basis for predictive maintenance and fault diagnosis of equipment is an important challenge currently faced.

[0003] The steel industry uses a large number of mechanical equipment, and the health status of these equipment is usually measured by vibration signals. The state degradation of mechanical equipment goes through a period of development, which is manifested as trend changes in vibration signals. Trend prediction of vibration signals is of great significance in equipment management. In the blast furnace gas recovery turbine power generation device (TRT, Top Gas Recovery Turbine), the state of the equipment is mainly judged by the amplitude of the vibration signal to determine whether the equipment needs to be shut down for maintenance. If the future development direction can be predicted from the trend change of the vibration signal and it can be judged when the vibration amplitude of the equipment reaches the threshold line, the maintenance plan can be arranged in advance and shutdown maintenance can be carried out in combination with production scheduling, thereby avoiding unplanned shutdowns caused by sudden equipment failures and reducing unnecessary production capacity losses of enterprises. The working conditions in the production process will change, and the vibration signals of the equipment show complex nonlinear characteristics. For such complex and changeable nonlinear changes, traditional prediction methods seem to be unable to cope with them.

[0004] In recent years, with the rapid development of deep learning technology, time series prediction methods based on long short-term memory networks (LSTM) have shown great potential in processing complex nonlinear data. Due to its special memory structure, the LSTM model can effectively capture the dependencies in long time series and has been widely used in various time series prediction tasks. However, in the actual application of TRT equipment, the speed and amplitude of the change of the equipment vibration signal may vary greatly. When the LSTM model processes these changes, a single fixed window often fails to achieve ideal results. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for predicting the maintenance time of blast furnace TRT equipment based on the LSTM model. The method can accurately predict the future development trend of the vibration signal of the TRT equipment, reasonably arrange the maintenance plan, avoid unplanned downtime caused by sudden failure of the TRT equipment, and improve the operating reliability and production efficiency of the TRT equipment.

[0006] In order to solve the above technical problems, the method for predicting the maintenance time of blast furnace TRT equipment based on LSTM model of the present invention comprises the following steps:

[0007] Step 1: Build a sliding window adaptive algorithm based on the vibration signal change characteristics. This algorithm dynamically adjusts the input data sliding window size of the LSTM model according to the fluctuation amplitude and change rate of the vibration signal. When the vibration signal fluctuates, the sliding window is automatically reduced to allow the LSTM model to capture subtle changes in the vibration signal in a timely manner. When the vibration signal tends to be stable, the sliding window is automatically expanded to increase the coverage of the training data.

[0008] Step 2: Adopt adaptive sliding window to obtain TRT equipment training data and train the LSTM model. The LSTM model captures the complex time dependency in the vibration signal through the gating mechanism, including long-term trends, periodic fluctuations and short-term anomalies. Through the training data adaptively adjusted by the sliding window, the LSTM model continuously learns the vibration mode of the equipment under different working conditions. The training process adopts batch gradient descent method and is optimized in combination with the self-learning mechanism. When new training data enters, the LSTM model gradually updates parameters to adapt to the latest equipment operation status to ensure that the prediction results are always accurate.

[0009] Step 3: Apply the trained LSTM model to the real-time collected vibration signal, predict the vibration trend of the TRT equipment in the next n days at the current point, and output the predicted value of the vibration signal in the future period of time;

[0010] Step 4: Based on the prediction results of the vibration signal and the preset shutdown maintenance threshold, determine the time when the TRT equipment reaches the threshold. For the first time point when the threshold is reached, the LSTM model calculates the time difference between this point and the current time point to obtain the remaining days before the TRT equipment needs maintenance.

[0011] Furthermore, during the operation of the TRT equipment, the LSTM model is regularly updated, and the latest vibration signal data is used to fine-tune the LSTM model. Through incremental training, the LSTM model is adapted to the changes in the working conditions of the TRT equipment, reducing the prediction deviation caused by changes in data distribution.

[0012] Furthermore, the actual maintenance data of TRT equipment is compared with the prediction results of the LSTM model, and a feedback mechanism is established to analyze the error between the actual TRT equipment failure time and the LSTM model prediction time. The adaptive algorithm of the sliding window, the structure and hyperparameters of the LSTM model are dynamically adjusted to ensure that the LSTM model maintains stability and accuracy in long-term predictions.

[0013] Furthermore, the vibration signal is collected from the intelligent operation and maintenance platform to obtain time series vibration data, with one data point per minute. After accumulating historical data for a period of time, it is used as a data set for initial training and verification of the LSTM model; the vibration data is preprocessed, and the vibration data during the shutdown of the TRT equipment is identified and removed by setting a speed threshold to avoid interference to the LSTM model caused by abnormal fluctuations in the vibration signal during the shutdown period. The outliers in the vibration data are detected and removed using a box plot to prevent extreme data from affecting the training of the LSTM model. The preprocessed vibration data is used to calculate an average value every day as a data point, and the aggregated vibration data is used for subsequent fitting and prediction.

[0014] Furthermore, in the sliding window adaptive algorithm, the sliding window size is represented by the number of aggregated points, the time series length of the vibration signal is N aggregated points, and the initial sliding window size is set to W. 0 The size of the sliding window is dynamically adjusted according to the real-time changes of the vibration signal, including:

[0015] Fluctuation amplitude detection, within the current sliding window, calculates the standard deviation of the vibration signal as a measure of the fluctuation amplitude. The formula is as follows:

[0016] σ t =std(X t-W+1 ,…,X t ) (5)

[0017] Among them, σ t is the standard deviation of the vibration signal in the sliding window at time t, X i is the vibration signal value at the i-th moment;

[0018] Change rate detection uses the difference in standard deviations between adjacent sliding windows to calculate the change rate. The formula is as follows:

[0019]

[0020] Among them, R t is the rate of change in the sliding window at time t, σ t-1 is the standard deviation in the previous sliding window, and W is the size of the current sliding window;

[0021] Adjust the sliding window size according to the fluctuation amplitude and change rate.

[0022] Set the sliding window reduction condition. When the fluctuation amplitude exceeds the preset threshold A or the change rate R t When the preset threshold C is exceeded, the vibration signal is considered to be in a state of violent fluctuation. At this time, the size of the sliding window is W t Reduce it by the following formula:

[0023] W t =max(W min ,W t-1 -△W) (7)

[0024] Among them, W min is the minimum sliding window size, △W is the sliding window reduction step size;

[0025] Set the sliding window expansion condition. When the fluctuation amplitude is lower than the preset threshold B or the change rate R t If the value is lower than the preset threshold D, the vibration signal is considered to be in a stable state. At this time, the size of the sliding window is W t Expand according to the following formula:

[0026] W t =min(W max ,W t-1 +△W) (8)

[0027] Among them, W max is the maximum sliding window size, △W is the window expansion step;

[0028] After each adjustment of the sliding window size, the sliding window moves forward one time point to obtain the data segment in the next sliding window, continue to perform fluctuation amplitude and change rate detection, and adjust the sliding window.

[0029] Furthermore, the preset threshold A is twice the standard deviation of the vibration signal, the preset threshold C=0.1, and the minimum sliding window size W min =10, the sliding window reduction step size and expansion step size △W=5, the preset threshold B is 1 times the standard deviation of the vibration signal, the preset threshold D=0.05, the maximum sliding window size W max =100.

[0030] Furthermore, the LSTM model training process inputs the training set data into the LSTM model in batches for iterative training. At the end of each training, the validation set is used to evaluate the LSTM model performance, the validation loss and evaluation indicators are recorded, and the training process is dynamically adjusted according to the early stopping strategy and the learning rate adjustment strategy. During the training process, the LSTM model weights with the smallest validation loss are saved as the final model parameters, and the LSTM model structure and weights are saved as files for subsequent loading and use.

[0031] Furthermore, the early stopping strategy is to monitor the validation set loss, and when the validation loss does not decrease after 10 consecutive trainings, stop training to prevent overfitting; the learning rate adjustment strategy is to reduce the learning rate by 50% when the validation set loss does not decrease after 5 consecutive trainings.

[0032] Since the method for predicting the maintenance time of blast furnace TRT equipment based on LSTM model of the present invention adopts the above technical scheme, that is, the method constructs a sliding window adaptive algorithm based on the change characteristics of vibration signal, and the algorithm dynamically adjusts the input data sliding window size of LSTM model according to the fluctuation amplitude and change rate of vibration signal; adopts adaptive sliding window to obtain TRT equipment training data and trains LSTM model; applies the trained LSTM model to the vibration signal collected in real time, predicts the vibration trend of TRT equipment in the next n days at the current point, and outputs the predicted value of vibration signal in the future period; according to the prediction result of vibration signal, combined with the preset shutdown maintenance threshold, judges the time when TRT equipment reaches the threshold, and for the first time point reaching the threshold, the LSTM model calculates the time difference between the point and the current time point, and obtains the remaining days of TRT equipment maintenance. The method can accurately predict the future development trend of TRT equipment vibration signal, reasonably arrange maintenance plan, avoid unplanned shutdown caused by sudden failure of TRT equipment, and improve the operation reliability and production efficiency of TRT equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments:

[0034] Figure 1 is a flowchart of the method;

[0035] Figure 2 is a flowchart of the sliding window adaptive algorithm in this method;

[0036] Figure 3 This is a schematic diagram for comparing the original data and preprocessed data in this method;

[0037] Figure 4 This is a schematic diagram of the prediction results when the sliding window is too small in this method;

[0038] Figure 5 This is a schematic diagram of the prediction results when the sliding window is too large in this method;

[0039] Figure 6 Schematic diagram of suitable prediction results of sliding window in this method;

[0040] Figure 7 This is a schematic diagram of the prediction results of a trend that rises rapidly starting from a certain day in this method;

[0041] Figure 8 This is a schematic diagram of the forecast results of a rapidly rising trend predicted from another day in this method;

[0042] Fig. 9 This is a schematic diagram of the prediction results of trend mutation in this method;

[0043] Fig.10 Schematic diagram of the prediction results with a flat trend in this method. DETAILED DESCRIPTION

[0044] Implementation example Figure 1 As shown, the method for predicting the maintenance time of blast furnace TRT equipment based on the LSTM model of the present invention comprises the following steps:

[0045] Step 1: Build a sliding window adaptive algorithm based on the vibration signal change characteristics. This algorithm dynamically adjusts the input data sliding window size of the LSTM model according to the fluctuation amplitude and change rate of the vibration signal. When the vibration signal fluctuates, the sliding window is automatically reduced to allow the LSTM model to capture subtle changes in the vibration signal in a timely manner. When the vibration signal tends to be stable, the sliding window is automatically expanded to increase the coverage of the training data.

[0046] Its core goal is to strike a balance between capturing the changes in equipment vibration signals affected by working conditions and reducing environmental noise interference, ensuring that the model can accurately predict under different working conditions; reducing the sliding window helps to improve the sensitivity of the model, especially to short-term signal fluctuations, thereby effectively improving the accuracy of predictions; expanding the sliding window helps to reduce the impact of short-term noise, allowing the model to learn a wider range of time series trends, thereby improving the robustness of the model and avoiding overfitting short-term noise data; the sliding window adjustment mechanism is based on real-time vibration signal changes, and dynamically controls the size of the sliding window by calculating the fluctuation amplitude, change rate and other parameters of the vibration signal; the sliding window adjustment rules can pre-set a set of thresholds and conditions to distinguish different states of the signal, and flexibly adjust the sliding window according to these conditions;

[0047] Step 2: Adopt adaptive sliding window to obtain TRT equipment training data and train the LSTM model. The LSTM model captures the complex time dependency in the vibration signal through the gating mechanism, including long-term trends, periodic fluctuations and short-term anomalies. Through the training data adaptively adjusted by the sliding window, the LSTM model continuously learns the vibration mode of the equipment under different working conditions. The training process adopts batch gradient descent method and is optimized in combination with the self-learning mechanism. When new training data enters, the LSTM model gradually updates parameters to adapt to the latest equipment operation status to ensure that the prediction results are always accurate.

[0048] Step 3: Apply the trained LSTM model to the real-time collected vibration signal, predict the vibration trend of the TRT equipment in the next n days at the current point, and output the predicted value of the vibration signal in the future period of time;

[0049] Step 4: Based on the prediction results of the vibration signal and the preset shutdown maintenance threshold, determine the time when the TRT equipment reaches the threshold. For the first time point that reaches the threshold, the LSTM model calculates the time difference between this point and the current time point to obtain the remaining days until the TRT equipment is maintained. This provides a reference for the reasonable arrangement of equipment maintenance time and avoids unplanned shutdowns and production interruptions caused by equipment damage.

[0050] Preferably, during the operation of the TRT equipment, the LSTM model is updated regularly, and the latest vibration signal data is used to fine-tune the LSTM model. Through incremental training, the LSTM model is adapted to changes in the working conditions of the TRT equipment, reducing prediction deviations caused by changes in data distribution.

[0051] Preferably, the actual maintenance data of the TRT equipment is compared with the prediction results of the LSTM model, a feedback mechanism is established, the error between the actual TRT equipment failure time and the LSTM model prediction time is analyzed, and the adaptive algorithm of the sliding window, the structure and hyperparameters of the LSTM model are dynamically adjusted to ensure that the LSTM model maintains stability and accuracy in long-term predictions and optimize the LSTM model performance.

[0052] Preferably, the vibration signal is collected from the intelligent operation and maintenance platform to obtain time series vibration data, with one data point per minute. After accumulating historical data for a period of time, the data set is used as the initial training and verification data set of the LSTM model; the vibration data is preprocessed to improve the efficiency of model training and the accuracy of prediction. By setting a speed threshold, the vibration data when the TRT equipment is shut down is identified and removed to avoid interference to the LSTM model caused by abnormal fluctuations in the vibration signal during the shutdown period. A box plot is used to detect and eliminate outliers in the vibration data to prevent extreme data from affecting the LSTM model training. The preprocessed vibration data is used to calculate an average value every day as a data point, and the aggregated vibration data is used for subsequent fitting and prediction.

[0053] The raw data and pre-processed data of TRT equipment vibration are as follows: Figure 3 As shown in the figure, by removing noise points and aggregating data, the processed data is smoother, which can effectively reduce the impact of near-zero or individual noise points on data statistical results and improve the stability and reliability of the prediction model.

[0054] Preferably, Figure 2As shown in the sliding window adaptive algorithm, the sliding window size is represented by the number of aggregated points, the time series length of the vibration signal is N aggregated points, and the size of the initial sliding window is set to W 0 The size of the sliding window is dynamically adjusted according to the real-time changes of the vibration signal, including:

[0055] Fluctuation amplitude detection, within the current sliding window, calculates the standard deviation of the vibration signal as a measure of the fluctuation amplitude. The formula is as follows:

[0056] σ t =std(X t-W+1 ,…,X t ) (9)

[0057] Among them, σ t is the standard deviation of the vibration signal in the sliding window at time t, X i is the vibration signal value at the i-th moment;

[0058] Change rate detection uses the difference in standard deviations between adjacent sliding windows to calculate the change rate. The formula is as follows:

[0059]

[0060] Among them, R t is the rate of change in the sliding window at time t, σ t-1 is the standard deviation in the previous sliding window, and W is the size of the current sliding window;

[0061] Adjust the sliding window size according to the fluctuation amplitude and change rate.

[0062] Set the sliding window reduction condition. When the fluctuation amplitude exceeds the preset threshold A or the change rate R t When the preset threshold C is exceeded, the vibration signal is considered to be in a state of violent fluctuation. At this time, the size of the sliding window is W t Reduce it by the following formula:

[0063] W t =max(W min ,W t-1 -△W) (11)

[0064] Among them, W min is the minimum sliding window size, △W is the sliding window reduction step size;

[0065] Set the sliding window expansion condition. When the fluctuation amplitude is lower than the preset threshold B or the change rate R t If the value is lower than the preset threshold D, the vibration signal is considered to be in a stable state. At this time, the size of the sliding window is W t Expand according to the following formula:

[0066] W t =min(W max ,W t-1 +△W) (12)

[0067] Among them, W max is the maximum sliding window size, △W is the window expansion step;

[0068] After each adjustment of the sliding window size, the sliding window moves forward one time point to obtain the data segment in the next sliding window, continue to perform fluctuation amplitude and change rate detection, and adjust the sliding window.

[0069] The initial parameters of the sliding window are set based on the historical data of the long-term operation of the blast furnace TRT equipment, verifying the dynamic adjustment effect of the window size under different vibration signal fluctuations, and combining the training and prediction performance of the LSTM model to finally determine the optimal parameter combination. If the data characteristics change too much, causing the parameter combination to no longer be applicable, the parameter optimization mechanism is triggered to automatically find the optimal parameter combination.

[0070] The initial length of the sliding window can be determined based on the statistical characteristics of the vibration signal, such as mean, variance, etc., to ensure that sufficient information can be covered in the initial stage. According to the characteristics of the TRT equipment vibration historical data, 60 days of historical data are selected for initialization.

[0071] Preferably, the preset threshold A is twice the standard deviation of the vibration signal, the preset threshold C=0.1, and the minimum sliding window size W min =10, the sliding window reduction step size and expansion step size △W=5, the preset threshold B is 1 times the standard deviation of the vibration signal, the preset threshold D=0.05, the maximum sliding window size W max =100.

[0072] Preferably, the LSTM model training process inputs the training set data into the LSTM model in batches for iterative training. At the end of each training, the validation set is used to evaluate the LSTM model performance, the validation loss and evaluation indicators are recorded, and the training process is dynamically adjusted according to the early stopping strategy and the learning rate adjustment strategy. During the training process, the LSTM model weights with the smallest validation loss are saved as the final model parameters, and the LSTM model structure and weights are saved as files for subsequent loading and use.

[0073] Preferably, the early stopping strategy is to monitor the validation set loss, and when the validation loss does not decrease after 10 consecutive trainings, stop training to prevent overfitting; the learning rate adjustment strategy is to reduce the learning rate by 50% when the validation set loss does not decrease after 5 consecutive trainings.

[0074] This method implements model performance evaluation and retraining in actual applications. When the true value arrives, the model prediction performance is continuously evaluated, and the performance of the model in actual applications is monitored by calculating key indicators, such as mean absolute error and root mean square error. According to the difference between the actual value and the predicted value, a comparison curve is drawn to intuitively display the deviation of the model, ensuring that the on-site operators have a comprehensive understanding of the model performance and give correct feedback. Based on the feedback from the on-site operators, the source of the model's prediction error is analyzed, and according to the actual operation situation, it is determined whether the model needs to be retrained or adjusted. If the decision mechanism determines that the model needs to be retrained, it will be incrementally retrained using the latest real data to maintain the model's sensitivity and adaptability to changes in the equipment's operating status. The retrained model will be re-evaluated on the test set to ensure its effectiveness and reliability in the new data environment.

[0075] In this method, the combination of LSTM model and sliding window adaptive algorithm can effectively predict the development trend of vibration signals. In the process of sliding window optimization, the improper setting of sliding window is checked to show the influence of sliding window size on the prediction accuracy of the model, such as Figure 4 and Figure 5 shown.

[0076] Figure 4 In the example, since the sliding window is set too small, about 10 days, the model is too sensitive to noise fluctuations, the predicted trend changes too quickly, the vibration trend quickly reaches the equipment maintenance threshold, the predicted remaining days are small, and the equipment maintenance is recommended too early, which increases the burden of equipment maintenance.

[0077] Figure 5 In the case of a large window setting of about 90 days, the model reacts slowly to changes in signal trends and cannot accurately capture future trends. The predicted value may reach the equipment maintenance threshold later than the true value, and the time for recommending equipment maintenance is too late, which cannot reserve enough preparation time for the equipment maintenance team, thus losing the meaning of predictive maintenance.

[0078] Therefore, this method can further introduce a parameter optimization mechanism to optimize the sliding window size and other model parameters to ensure the improvement of prediction performance.

[0079] (1) Optimization goal: By optimizing the window reduction step size △W and the minimum sliding window W during dynamic adjustment of the sliding window min and the maximum sliding window W max ,Therefore, in the process of adaptive adjustment of the sliding window, the changes of the vibration signal can be captured more accurately, and the accuracy of the model prediction can be improved.,The evaluation criteria for the accuracy of the prediction model,In the optimization test, the average absolute error between the predicted value and the measured value is less than 2, which is judged to be high in accuracy and the accuracy of the prediction model meets the standard.

[0080] (2) Optimization algorithm selection: In the process of dynamic adjustment of window size, the particle swarm optimization algorithm is used to reduce the sliding window step size △W and the minimum sliding window W. min and the maximum sliding window W max Optimize to find the most suitable parameter combination to achieve the best prediction accuracy of the LSTM model.

[0081] (3) Optimization steps:

[0082] Initialization parameters: Set the initial sliding window reduction step size △W, the minimum sliding window W min and the maximum sliding window W max For example, the initial value of the sliding window reduction step is 10, and the minimum sliding window W min The initial value is 5, and the maximum sliding window W max The initial value is 20.

[0083] Objective function definition: Taking the model prediction accuracy as the objective function, the size of the sliding window is optimized. The objective function can be expressed as:

[0084] Maximize Accuracy(△W,W min ,W max ) subject to W min ≤△W≤W max (5)

[0085] Optimization process: Use the optimization algorithm to gradually adjust the length of the sliding window. After each adjustment, retrain the LSTM model and calculate the new prediction accuracy. The optimization algorithm determines the adjustment value of the next sliding window size based on the direction of change in accuracy until the optimal window length is found to achieve the model prediction accuracy.

[0086] (4) Optimization result verification: Once the optimal parameter combination (sliding window reduction step size △W, minimum sliding window W) is determined through the optimization algorithm, min , maximum sliding window W max ), the system will readjust the dynamic change process of the sliding window based on these parameters and verify it on the LSTM model. If the accuracy of the model on the verification set reaches the preset target, the set of parameters will be used for subsequent vibration trend prediction. Figure 6 As shown in the figure, only when the sliding window is selected appropriately, about 60 days, can the forecast result be closer to the true value, thereby effectively reflecting the trend change of the signal and being used to predict future trends.

[0087] This method was applied to the historical data of vibration parameters of TRT equipment in an ironmaking plant in the past two years, and the prediction accuracy of the LSTM model and sliding window adaptation under different working conditions, including rapid trend increase, sudden trend change, and gentle trend, was verified.

[0088] (1) Rapidly rising trend: Figure 7 and Figure 8 The forecast results starting from December 30, 2023 and January 10, 2024 are shown respectively. The two forecasts reflect the forecast results of future trends at different time points. Figure 7 As shown in the figure, the data in this group increased by more than 100 in 2 months, indicating that the equipment status deteriorated rapidly. The predicted value in the figure is close to the measured value, indicating that the predicted value follows the change law of the measured value and can be used to predict future trends.

[0089] Comparing the prediction results at two time points, it can be seen that the prediction results updated daily introduce the recent data change patterns. When the data trend changes faster, the remaining days of the prediction also decrease rapidly, instead of decreasing linearly with the increase of the date. This can more accurately remind equipment maintenance personnel to schedule downtime.

[0090] As time goes by, the prediction results are updated every day, including the remaining days to reach the equipment maintenance threshold. See Table 1 for details of the remaining days predicted each day. It can be seen that the remaining days predicted each day change with the latest data patterns, and the recommendations for equipment maintenance are more accurate.

[0091] Table 1

[0092]

[0093] (2) Sudden trend change: Fig. 9 The prediction effect of this method when the vibration trend of the equipment changes suddenly is demonstrated. At the prediction start time, December 11, 2023, the vibration signal undergoes a small mutation, and the vibration value changes from 10 to 25, and then drops back to 10 four days later. By combining the LSTM model with the sliding window adaptive algorithm, this mutation can be accurately captured and an accurate prediction can be made to avoid misjudgment of the trend. The optimized sliding window mechanism significantly improves the response speed of the model, and can quickly adjust the window size when the vibration signal changes rapidly, thereby improving the model's sensitivity to sudden changes in trends. The prediction results show that in the next 60 days, the vibration signal will rise and then fall back, and then remain stable, and will not reach the equipment maintenance threshold. The predicted value follows the change law of the measured value, which further verifies the prediction accuracy and reliability of this method under complex working conditions.

[0094] (3) Stable trend: Fig.10The prediction effect of this method is demonstrated when the equipment is running smoothly and the equipment status is stable. The overall change of the vibration signal in the figure is relatively stable, without obvious mutations or large fluctuations. In this case, the sliding window adaptive algorithm can effectively expand the sliding window range and reduce noise interference, thereby improving the prediction accuracy of the model under a stable trend. The prediction results show that the vibration signal will not reach the equipment maintenance threshold in the next 60 days, which verifies that this method has good robustness and accuracy under stable conditions and can provide effective status monitoring for equipment operation.

[0095] Through the prediction demonstration of these three working conditions, this method shows good prediction accuracy and reliability under different working conditions.

[0096] This method is used to predict the vibration value in the future and the remaining time for the vibration value to reach the equipment maintenance threshold. By combining the sliding window adaptive algorithm with the LSTM model, the accuracy, flexibility and robustness of equipment vibration trend prediction are significantly improved.

[0097] 1. Improved accuracy of TRT equipment vibration trend prediction: By introducing a sliding window adaptive algorithm, this method can dynamically adjust the sliding window size of the training data according to the real-time changes of the vibration signal. When the vibration fluctuates violently, the sliding window is reduced to capture subtle changes, ensuring that the LSTM model can accurately identify and learn the changing characteristics of the vibration signal, thereby improving the accuracy of the prediction. When the signal is stable, the sliding window is automatically expanded to increase the range of training data, reduce noise interference, and further improve the accuracy of the prediction.

[0098] 2. Estimation of the remaining days to reach the equipment maintenance threshold: Combining the prediction results of the vibration signal with the preset equipment maintenance threshold, this method can estimate the remaining time to reach the equipment maintenance threshold, providing a reference for equipment maintenance personnel to reasonably arrange downtime. This early warning mechanism helps enterprises take preventive maintenance measures before equipment failure occurs, thereby reducing sudden equipment downtime and improving production continuity and safety.

[0099] This method significantly improves the effect of TRT equipment vibration signal trend prediction and provides a quantitative reference for predictive maintenance based on equipment status, which has important practical value and promotion significance.

[0100] This method is suitable for the condition monitoring of industrial equipment such as blast furnace TRT equipment. By intelligently predicting the vibration trend of equipment and the remaining time to reach the maintenance threshold, it helps equipment managers take preventive maintenance measures, reduce equipment failure rate, extend equipment service life, and reduce operation and maintenance costs.

[0101] Through the combination of LSTM model training and adaptive sliding window, this method can effectively solve the shortcomings of existing vibration trend prediction methods, and provide a more flexible, accurate and adaptable equipment vibration signal trend prediction method, which is suitable for intelligent monitoring and management of industrial equipment status, and provides more forward-looking decision support for equipment management, and finally realizes intelligent prediction of equipment operation status.

Claims

1. A method for predicting the maintenance time of blast furnace TRT equipment based on LSTM model, characterized in that The steps include: Step 1: Build a sliding window adaptive algorithm based on the vibration signal change characteristics. This algorithm dynamically adjusts the input data sliding window size of the LSTM model according to the fluctuation amplitude and change rate of the vibration signal. When the vibration signal fluctuates, the sliding window is automatically reduced to allow the LSTM model to capture subtle changes in the vibration signal in a timely manner. When the vibration signal tends to be stable, the sliding window is automatically expanded to increase the coverage of the training data. Step 2: Adopt adaptive sliding window to obtain TRT equipment training data and train the LSTM model. The LSTM model captures the complex time dependency in the vibration signal through the gating mechanism, including long-term trends, periodic fluctuations and short-term anomalies. Through the training data adaptively adjusted by the sliding window, the LSTM model continuously learns the vibration mode of the equipment under different working conditions. The training process adopts batch gradient descent method and is optimized in combination with the self-learning mechanism. When new training data enters, the LSTM model gradually updates parameters to adapt to the latest equipment operation status to ensure that the prediction results are always accurate. Step 3: Apply the trained LSTM model to the real-time collected vibration signal, predict the vibration trend of the TRT equipment in the next n days at the current point, and output the predicted value of the vibration signal in the future period of time; Step 4: Based on the prediction results of the vibration signal and the preset shutdown maintenance threshold, determine the time when the TRT equipment reaches the threshold. For the first time point when the threshold is reached, the LSTM model calculates the time difference between this point and the current time point to obtain the remaining days before the TRT equipment needs maintenance.

2. The method for predicting the maintenance time of blast furnace TRT equipment based on LSTM model according to claim 1 is characterized in that: During the operation of the TRT equipment, the LSTM model is updated regularly and the latest vibration signal data is used to fine-tune the LSTM model. Through incremental training, the LSTM model can adapt to changes in the working conditions of the TRT equipment and reduce prediction deviations caused by changes in data distribution.

3. The method for predicting the maintenance time of blast furnace TRT equipment based on LSTM model according to claim 1 is characterized in that: The actual maintenance data of TRT equipment is compared with the prediction results of the LSTM model, a feedback mechanism is established, the error between the actual TRT equipment failure time and the LSTM model prediction time is analyzed, and the adaptive algorithm of the sliding window, the structure and hyperparameters of the LSTM model are dynamically adjusted to ensure that the LSTM model maintains stability and accuracy in long-term predictions.

4. The method for predicting the maintenance time of blast furnace TRT equipment based on LSTM model according to claim 1 is characterized in that: The vibration signal is collected from the intelligent operation and maintenance platform to obtain time series vibration data, with one data point per minute. After accumulating historical data for a period of time, it is used as a data set for initial training and verification of the LSTM model. The vibration data is preprocessed, and the vibration data during the shutdown of the TRT equipment is identified and removed by setting a speed threshold to avoid interference to the LSTM model caused by abnormal fluctuations in the vibration signal during the shutdown period. The outliers of the vibration data are detected and removed using a box plot to prevent extreme data from affecting the training of the LSTM model. The preprocessed vibration data is used to calculate an average value every day as a data point, and the aggregated vibration data is used for subsequent fitting and prediction.

5. The method for predicting the maintenance time of blast furnace TRT equipment based on LSTM model according to claim 1 is characterized in that: In the sliding window adaptive algorithm, the sliding window size is represented by the number of aggregated points, the time series length of the vibration signal is N aggregated points, the initial sliding window size is set to W0 aggregated points, and the sliding window size is dynamically adjusted according to the real-time changes of the vibration signal, including: Fluctuation amplitude detection, within the current sliding window, calculates the standard deviation of the vibration signal as a measure of the fluctuation amplitude. The formula is as follows: s t =std(X t-W+1 ,…,X t ) (1) Among them, σ t is the standard deviation of the vibration signal in the sliding window at time t, X i is the vibration signal value at the i-th moment; Change rate detection uses the difference in standard deviations between adjacent sliding windows to calculate the change rate. The formula is as follows: Among them, R t is the rate of change in the sliding window at time t, σ t-1 is the standard deviation in the previous sliding window, and W is the size of the current sliding window; Adjust the sliding window size according to the fluctuation amplitude and change rate. Set the sliding window reduction condition. When the fluctuation amplitude exceeds the preset threshold A or the change rate R t When the preset threshold C is exceeded, the vibration signal is considered to be in a state of violent fluctuation. At this time, the size of the sliding window is W t Reduce it by the following formula: W t =max(W min ,W t-1 -△W) (3) Among them, W min is the minimum sliding window size, △W is the sliding window reduction step size; Set the sliding window expansion condition. When the fluctuation amplitude is lower than the preset threshold B or the change rate R t If the value is lower than the preset threshold D, the vibration signal is considered to be in a stable state. At this time, the size of the sliding window is W t Expand according to the following formula: IN t =min(W max ,IN t-1 +△W) (4) Among them, W max is the maximum sliding window size, △W is the window expansion step size; After each adjustment of the sliding window size, the sliding window moves forward one time point to obtain the data segment in the next sliding window, continue to perform fluctuation amplitude and change rate detection, and adjust the sliding window.

6. The method for predicting the maintenance time of blast furnace TRT equipment based on LSTM model according to claim 5 is characterized in that: The preset threshold A is 2 times the standard deviation of the vibration signal, the preset threshold C=0.1, and the minimum sliding window size W min =10, the sliding window reduction step size and expansion step size △W=5, the preset threshold B is 1 times the standard deviation of the vibration signal, the preset threshold D=0.05, the maximum sliding window size W max =100.

7. The method for predicting the maintenance time of blast furnace TRT equipment based on LSTM model according to claim 1 is characterized in that: The LSTM model training process inputs the training set data into the LSTM model in batches for iterative training. At the end of each training, the validation set is used to evaluate the LSTM model performance, the validation loss and evaluation indicators are recorded, and the training process is dynamically adjusted according to the early stopping strategy and the learning rate adjustment strategy. During the training process, the LSTM model weights with the smallest validation loss are saved as the final model parameters, and the LSTM model structure and weights are saved as files for subsequent loading and use.

8. The method for predicting the maintenance time of blast furnace TRT equipment based on LSTM model according to claim 7 is characterized in that: The early stopping strategy is to monitor the validation set loss, and when the validation loss does not decrease after 10 consecutive trainings, stop training to prevent overfitting; the learning rate adjustment strategy is to reduce the learning rate by 50% when the validation set loss does not decrease after 5 consecutive trainings.

Images