An intelligent shutdown decision method based on loosening prediction of anode saturated reactor

Abstract

This invention discloses an intelligent shutdown decision-making method based on loosening prediction of anode saturated reactors. The method comprises the following steps: S1. For anode saturated reactors of the same model, historical information of different operating times is collected; S2. Based on this historical information of different operating times, characteristic information of the anode saturated reactors is calculated and converted into continuous time series data with time tags; S3. An LSTM network is constructed and trained using the continuous time series data with time tags, resulting in a trained LSTM network that serves as a looseness prediction model; and S4. Based on reinforcement learning, an intelligent shutdown decision is obtained. This method can dynamically determine the timing of shutdowns for maintenance before a fault occurs, balancing production continuity (avoiding frequent shutdowns) with equipment safety (avoiding major failures).

Description

Technical Field

[0001] The invention relates to an anode saturated reactor, and in particular to an intelligent shutdown decision method based on loosening prediction of the anode saturated reactor. Background Art

[0002] In the converter valves of HVDC transmission systems, the anode saturation reactor plays a critical role: as a nonlinear inductor, it effectively limits the high current rise rate that occurs during the initial turn-on phase of the thyristor, thereby protecting the thyristor. However, due to the magnetostrictive effect, the reactor can vibrate violently, which can cause the bolts connecting it to the valve tower to gradually loosen. It is worth noting that this type of fault is often difficult to detect in the early stages, but if the equipment continues to operate with the fault, it can eventually cause catastrophic damage. Summary of the Invention

[0003] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide an intelligent shutdown decision method based on loosening prediction of anode saturated reactor, which can dynamically decide the shutdown and maintenance timing before a fault occurs, and balance production continuity and equipment safety.

[0004] The object of the present invention is achieved through the following technical solution: an intelligent shutdown decision method based on loosening prediction of anode saturated reactor, comprising the following steps:

[0005] S1. For the same type of anode saturated reactor, collect historical information of different operating times;

[0006] S2. Calculate characteristic information of the anode saturated reactor based on historical information of different operating times and convert the characteristic information into continuous time series data with time tags;

[0007] S3. Build an LSTM network and train it using continuous time series data with time labels. The trained LSTM network is used as a prediction model for the degree of looseness.

[0008] S4. Based on reinforcement learning, intelligent shutdown decision is obtained.

[0009] The beneficial effects of this invention include the ability to dynamically determine the timing of downtime for maintenance before a fault occurs, balancing production continuity (avoiding frequent downtime) with equipment safety (avoiding major failures). The designed state space supports flexible deployment of the monitoring system at any time during equipment operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a flow chart of the method of the present invention;

[0011] Figure 2Schematic diagram of LSTM input and output;

[0012] Figure 3 This is a flowchart for reinforcement learning;

[0013] Figure 4 is the graph of the sigmoid function. DETAILED DESCRIPTION

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0015] like Figure 1 As shown, an intelligent shutdown decision method based on loosening prediction of anode saturated reactor includes the following steps:

[0016] S1. For the same type of anode saturated reactor, collect historical information of different operating times;

[0017] Initial data acquisition: A predictive model is constructed based on historical operating data to anticipate potential faults during online monitoring. For anode saturation reactors of the same model installed at converter stations nationwide, multiple sets of samples are collected for different shutdown intervals (e.g., operating durations of 5, 6, 8, 10, and 12 months). Soundprint data is collected immediately before shutdown using soundprint sensors installed on each reactor, and vibration acceleration data is collected in real time using a laser vibrometer. Furthermore, during shutdown inspections, a digital torque wrench is used to measure the loosening torque of the connecting bolts (i.e., the minimum torque at which the bolts begin to turn). The degree of bolt loosening is expressed as "(tightening torque - loosening torque) / tightening torque × 100%).

[0018] S2. Calculate characteristic information of the anode saturated reactor based on historical information of different operating times and convert the characteristic information into continuous time series data with time tags;

[0019] Sensor data: The voiceprint collection sensor obtains the voiceprint data of the day of the outage, and the laser vibrometer obtains the vibration acceleration data of the day of the outage;

[0020] Degree of looseness: As the increase of looseness leads to accelerated failure, a nonlinear fault propagation model is defined, assuming that the degree of looseness L ( t ) changes with time t (unit: day) following the exponential growth law, , the discrete data can be fitted with nonlinear least squares method to obtain unknowns k The degree of looseness is the output side of the LSTM.

[0021] Vibration acceleration peak value A: As the mechanical structure loosens, the system stiffness decreases, resulting in a significant increase in vibration.

[0022] Growth is slow in the early stages and accelerates in the later stages (non-linearly), consistent with exponential growth: a, b, c are fitting constants, and t is time.

[0023] Vibration acceleration high-frequency and low-frequency energy ratio B: , A i 100* i Hz corresponds to the amplitude of the acceleration waveform. As the mechanical structure loosens, the proportion of high-frequency energy increases, and this feature gradually increases over time, showing a nonlinear growth trend. a, b, c are fitting constants, and t is time.

[0024] Sound pressure level C: As the bolts become looser, the sound pressure level of the anode saturated reactor noise will increase, usually showing a smooth increase over time. a, b, c are fitting constants, and t is time.

[0025] Voiceprint spectrum complexity D: , , A i 100* i Hz corresponds to the amplitude of the sound wave, R i 100* i Hz corresponds to the energy density of the sound wave. As the bolt loosens more, the structure of the spectrum becomes more complex, so this feature will gradually increase over time. a, b, c are fitting constants, and t is time.

[0026] S3. Build an LSTM network and train it using continuous time series data with time labels to obtain the trained LSTM network as a prediction model for the degree of looseness; Figure 2 As shown:

[0027] LSTM input side: The input of LSTM is multimodal time series data, including the peak value and mean value of vibration acceleration (m / s 2 ), vibration acceleration high- and low-frequency energy ratio, sound pressure level (dB), and voiceprint spectrum complexity. Assume the time window is a sliding time window, with each time step corresponding to one day's data.

[0028] LSTM output side: looseness L

[0029] S4. Based on reinforcement learning, intelligent shutdown decision is obtained, such as Figure 3 As shown:

[0030] Building a reinforcement learning environment

[0031] 4.1 State Space Design:

[0032] The state space includes the current looseness and the rate of change of the looseness.

[0033] Status indication: s =( , );in L is the current predicted looseness, d L / d t is the rate of change of looseness; d L The difference between the current predicted looseness and the last predicted looseness, d t is one time step, i.e. one day;

[0034] 4.2 Action Space Design:

[0035] action a =0 (Continue): In this action, the model decides to keep the device running until the next decision moment. The interval between decision moments can be configured based on the device monitoring frequency, safety requirements, and maintenance management needs.

[0036] action a =1 (Stop for Maintenance): In this action, the model decides to stop the equipment and perform necessary maintenance.

[0037] 4.3 Reward Function Design:

[0038] Since a discontinuous or sudden reward function may cause the agent's final decision to be fixed at the threshold, the reward function is designed to be a continuous function. S express.

[0039] like Figure 4 As shown, the mathematical expression of the Sigmoid function is:

[0040] Reward function:

[0041]

[0042] First row: Change rate penalty (first item): If the rate of change of the equipment looseness is large, the operation will be penalized to prevent untimely shutdown during rapid changes; Balance reward (second item): When the equipment looseness is low and the rate of change is small, the equipment should be encouraged to continue operating and rewarded;

[0043] Second row: Equipment failure penalty: When the equipment looseness exceeds the threshold of 0.8, but the system does not shut down in time, the equipment fails and a -100 penalty is applied;

[0044] The third row: Timely shutdown reward (first item): When the degree of looseness is high, timely shutdown to prevent failure should be rewarded; Frequent shutdown penalty (second item): If the degree of looseness is low, the system will be penalized when it is shut down; Embody the engineering concept of "timely maintenance" (third and fourth items): Form an optimal shutdown interval, the third item: Penalize premature shutdown ( L <0.5), the fourth item: reward the best downtime ( L close to 0.6).

[0045] 4.4 Reinforcement Learning Algorithm Selection:

[0046] Since the state space s The model is trained using a deep Q-network (DQN) because the dimension is small (three-dimensional) and continuous. Traditional Q-learning algorithms use Q-tables to store state-action-value functions, which makes it difficult to handle continuous state spaces. DQN can handle continuous states through neural network approximation, so it was chosen as the reinforcement learning algorithm for this study.

[0047] To train the model:

[0048] 4.41 Initializing the Deep Q Network Structure

[0049] Build a neural network that takes a state vector as input and outputs Q values ​​corresponding to two actions. The network consists of two hidden layers, uses the ReLU activation function, and the output layer uses linear activation to represent the Q values. Initial weights are initialized with random numbers.

[0050] 4.42 Generating Experience through Interaction between Agent and Environment

[0051] Run multiple training cycles in the simulation environment. In each cycle, the agent ε -The greedy strategy selects actions and records state transitions (current state, action taken, reward obtained, next state) to form experience data.

[0052] 4.43 Experience Replay and Target Network Update

[0053] The experience data generated during the interaction is stored in a replay buffer. Every set number of steps, experience is randomly sampled from the replay pool to train the main Q network. The mean squared error loss function is used to optimize the Q-value approximation process, and the target network weights are periodically synchronized.

[0054] 4.44 Training Convergence and Policy Extraction

[0055] Repeat the above process until the average reward change of the model is less than 0.1 for 100 consecutive rounds. Finally, extract the trained model as the intelligent shutdown decision controller to determine whether to perform equipment maintenance operations online.

[0056] The present invention has been described and illustrated in detail so that those skilled in the art can understand and apply the present invention. Those skilled in the art will appreciate that various modifications and variations based on the above description may be made. Such modifications and variations, provided they do not depart from the spirit and scope of the present invention, are intended to be within the scope of the appended claims.

Claims

1. An intelligent shutdown decision-making method based on loosening prediction of anode saturated reactor, characterized by: The following steps are involved: S1. For the same type of anode saturated reactor, collect historical information of different operating times; The running time is in days, with each day being a time step. The historical information at different running times, i.e., the historical information at different time steps, includes: The soundprint data collected by the soundprint collection sensor, the vibration acceleration data collected by the laser vibrometer, and the loosening torque of the reactor connection bolts collected using a digital torque wrench; The voiceprint data includes the effective value of the sound pressure, the root mean square value of the sound pressure in a time step, and when i=1,2,…30, 100* i The amplitude of the sound wave corresponding to Hz; The vibration acceleration data includes: each vibration acceleration peak value within a time step, and when i=1,2,…30, 100* i The amplitude of the vibration acceleration waveform corresponding to Hz; S2. Calculate characteristic information of the anode saturated reactor based on historical information of different operating times and convert the characteristic information into continuous time series data with time tags; The step S2 comprises: S201. Calculate characteristic information of the anode saturated reactor, including acceleration peak mean, high-frequency and low-frequency energy ratio of vibration acceleration, sound pressure level, soundprint spectrum complexity, and looseness; Based on the vibration acceleration data collected by the laser vibrometer, calculate the peak value A of the vibration acceleration and the high-frequency and low-frequency energy ratio B of the vibration acceleration: Among them, the vibration acceleration peak mean A is obtained by averaging the vibration acceleration peak values; The calculation method of the high-frequency and low-frequency energy ratio B of vibration acceleration is as follows: ； A i 100* i Hz corresponds to the amplitude of the acceleration waveform; According to the voiceprint data collected by the voiceprint collection sensor, the sound pressure level C and the voiceprint spectrum complexity D are calculated: The sound pressure level C is calculated as follows: C= 20*lg(P / P 0 ) Where P is the effective value of the measured sound pressure, P 0 is the reference sound pressure, which is 20 μPa; The voiceprint spectrum complexity D is calculated as follows: , , A i 100* i Hz corresponds to the amplitude of the sound wave, R i 100* i Hz corresponds to the energy density of the sound wave; Calculate the degree of looseness based on the loosening torque of the reactor connection bolts collected by the digital torque wrench L : L = (tightening torque - loosening torque) / tightening torque * 100% When the connecting bolts of the reactor are in a fully tightened state, the torque of the connecting bolts of the reactor is collected in advance using a digital torque wrench, and the obtained torque is the tightening torque; S202. Fit the relationship between the peak-mean value of acceleration, the high-frequency and low-frequency energy ratio of vibration acceleration, the sound pressure level, the complexity of the soundprint spectrum, the degree of looseness, and the running time t: A1. Define the average value of the peak value of vibration acceleration when the running time is t as : At different running times t, the corresponding acceleration peak mean is obtained according to step S201 and recorded as , through multiple and the corresponding running time t, the fitting constant a 1, b1, c1, and thus obtain the relationship between the peak value and the average value of vibration acceleration and the running time; A2. Define the high-frequency and low-frequency energy ratio of acceleration when the running time is t. : At different running times t, the corresponding acceleration high-frequency and low-frequency energy ratios are obtained according to step S201 and are recorded as , through multiple and the corresponding running time t, the fitting constant a 2, b2, c2, thus obtaining the relationship between the high-frequency and low-frequency energy ratio of acceleration and the running time; A3. Define the sound pressure level when the running time is t. : At different running times t, the corresponding sound pressure levels are obtained according to step S201 and are recorded as , through multiple and the corresponding running time t, the fitting constant a 3, b3, c3, and thus the relationship between the sound pressure level and the running time is obtained; A4. Define the voiceprint spectrum complexity when the running time is t : At different running times t, the corresponding voiceprint spectrum complexity is obtained according to step S201 and is recorded as , through multiple and the corresponding running time t, the fitting constant a 4, b4, c4, and thus obtain the relationship between the voiceprint spectrum complexity and the running time; A5. Define the degree of looseness when the running time is t. : At different running times t, the corresponding looseness degree is obtained according to step S201 and is recorded as , through multiple and the corresponding running time t, fitting k , thus obtaining the relationship between the degree of looseness and the running time; S203. Based on the transformation relationships obtained by fitting A1-A5, obtain continuous time series data with time tags, including the peak-mean value of acceleration, the high-frequency and low-frequency energy ratio of vibration acceleration, the sound pressure level, the complexity of the voiceprint spectrum, and the degree of looseness. In the continuous time series data, each day is a time step. S3. Build an LSTM network and train it using continuous time series data with time labels. The trained LSTM network is used as a prediction model for the degree of looseness. S4. Based on reinforcement learning, intelligent shutdown decision is obtained.

2. The intelligent shutdown decision method based on loosening prediction of anode saturated reactor according to claim 1, characterized in that: In step S1, when collecting historical information of different operating times for anode saturated reactors of the same model, the information needs to be collected during a fixed period of time before the reactor is shut down, where the fixed period of time is one day.

3. The intelligent shutdown decision method based on loosening prediction of anode saturated reactor according to claim 1, characterized in that: When the LSTM network is trained in step S3, at each running time t, the data ( , , , ) as the input feature of the LSTM network, and the corresponding degree of looseness As the expected output, the LSTM network is trained to obtain a mature LSTM network as a prediction model for the degree of looseness.

4. The intelligent shutdown decision method based on loosening prediction of anode saturated reactor according to claim 1, characterized in that: In step S4, when the equipment operation time is unknown, starting from any time point, the acceleration peak mean A, the vibration acceleration high-low frequency energy ratio B, the sound pressure level C, and the voiceprint spectrum complexity D are collected according to step S201, and sent to a well-trained LSTM network to predict the corresponding looseness degree. After the prediction starts, the looseness degree prediction is performed every time step. The second prediction is used as the decision starting point, and an intelligent shutdown decision is obtained based on reinforcement learning.

5. The intelligent shutdown decision method based on loosening prediction of anode saturated reactor according to claim 4, characterized in that: Starting from the second prediction as the decision starting point, based on the reinforcement learning method, the intelligent shutdown decision is obtained including: S401. State Space Design: Status indication: s =( , d / d t ); where L is the currently predicted looseness, dL / dt is the rate of change of the looseness; dL is the difference between the currently predicted looseness and the last predicted looseness, and dt is a time step, i.e., one day; S402. Action Space Design: action a =0: Under this action, the model decides to keep the device running until the next decision point; action a =1: In this action, the model decides to suspend the equipment for maintenance; S403. Reward function design: Since the discontinuous and sudden reward function causes the agent's final decision to be fixed at the threshold, the reward function is designed to be a continuous function, and the Sigmoid function is introduced and used. S express: The mathematical expression of the Sigmoid function is: The reward function is designed as: S404. Model training: The model training process uses the deep Q network, namely the DQN network: S4041. Initialize the deep Q network structure: Construct a neural network that takes a state vector as input and outputs two Q values ​​corresponding to actions. The network consists of two hidden layers, uses the ReLU activation function, and the output layer uses linear activation to represent the Q value. The initial weights are initialized with random numbers. S4042. Generating Experience through Interaction between Agents and Environments Run multiple training cycles in the simulation environment. In each cycle, the agent ε - Greedy strategy selects actions and records state transitions, including current state, action execution, reward, and next state, to form experience data; S4043. Experience Replay and Target Network Update: The experience data generated during the interaction is stored in the replay buffer. Every set number of steps, experience is randomly sampled from the replay pool to train the main Q network. The mean square error loss function is used to optimize the Q value approximation process and periodically synchronize the target network weights. S4044. Training Convergence and Strategy Extraction: The above process is repeated until the average reward change of the model is less than 0.1 for 100 consecutive rounds. Finally, the trained model is extracted as the intelligent shutdown decision controller to determine whether to perform equipment maintenance operations online.

Images