Remote control and monitoring system of quick-switching blind plate valve based on floater control

Abstract

The invention relates to the technical field of industrial valve control, in particular to a floater control-based remote control and monitoring system for a quick-switching blind plate valve, which comprises a data acquisition unit, a cross-dimension data analysis unit and an instruction generation and feedback unit, and is characterized in that the data acquisition unit constructs a three-dimensional motion matrix through a floater trajectory tracking algorithm; an effective pressure fluctuation period is extracted in combination with a pressure-temperature coupling denoising algorithm, a cross-dimension data analysis unit extracts time-frequency fusion features by using a residual convolutional neural network, and dual verification of a fluid phase state is realized through a support vector machine and a dynamic time warping algorithm; and the instruction generation and feedback unit triggers a standard control strategy or a self-learning compensation mechanism according to the confidence coefficient, and corrects control deviation through a negative feedback optimization coefficient matrix, so that the problems of insufficient multi-source data fusion and poor complex working condition adaptability in the traditional technology are solved, accurate remote control and intelligent monitoring of the quick-switching blind plate valve are realized, and the working efficiency of the quick-switching blind plate valve is improved. And the leakage early warning success rate and the valve action reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial valve control technology, and more specifically, to a remote control and monitoring system for a quick-cut blind valve based on float control. Background Technology

[0002] Intelligent control of industrial valves is an important technology. In process industries such as petrochemicals and natural gas transportation, quick-cut blind valves are key shut-off devices in pipeline systems. Their remote control and monitoring technology, through real-time sensing of fluid status and precise execution of valve actions, is a core means to ensure the safe operation of pipelines and prevent leakage accidents.

[0003] With the development of industrial automation and intelligence, building a control system that integrates multi-source sensor data and intelligent algorithms is crucial for improving valve response speed and adapting to complex fluid conditions. Traditional single-parameter control modes are no longer sufficient to meet the precise control requirements under complex conditions such as gas-liquid two-phase flow and transient impact flow. However, traditional remote control technology for quick-cut blind valves suffers from core problems such as insufficient multi-source data fusion and poor adaptability to complex conditions. Existing solutions rely solely on single parameters such as pressure or displacement, failing to establish a cross-dimensional correlation model between float motion and fluid pressure. When gas-liquid two-phase flow occurs in the pipeline, the correlation delay between the sudden change in float angular acceleration and pressure fluctuation exceeds a preset threshold, resulting in valve lag and low success rate of leakage accident early warning. Rule-based control methods cannot effectively extract the time-frequency characteristics of fluid phase states. The accuracy of feature recognition in the 100-200Hz transient impact frequency band is low, increasing the probability of misjudging gas-liquid two-phase flow as liquid single-phase flow, leading to valve malfunction. In addition, traditional systems lack a self-learning compensation mechanism. When faced with the thermal expansion of float materials caused by pipeline temperature changes, they cannot adaptively correct displacement data deviations, increasing valve positioning errors. This lack of multi-source data correlation and insufficient intelligent adaptability ultimately results in low valve control accuracy under complex operating conditions, making it difficult to meet the high reliability control requirements of high-risk pipeline systems. To solve this technical problem, we provide a remote control and monitoring system for quick-cut blind valves based on float control. Summary of the Invention

[0004] The purpose of this invention is to provide a remote control and monitoring system for a quick-cut blind valve based on float control, so as to solve the problems mentioned in the background art.

[0005] Due to the insufficient fusion of multi-source data in traditional technologies, the correlation between float motion and fluid pressure is delayed, resulting in a low success rate of leak early warning. Therefore, this case study constructs a three-dimensional motion matrix through a data acquisition unit and combines it with a pressure-temperature coupling denoising algorithm to shorten the correlation delay and improve the success rate of leak early warning.

[0006] Traditional methods are poorly adaptable to complex working conditions, have low accuracy in identifying transient impact frequency band features, and result in frequent valve malfunctions. Therefore, this case study utilizes a residual convolutional neural network in a cross-dimensional data analysis unit with time-frequency fusion feature extraction to improve feature recognition accuracy and reduce valve malfunctions.

[0007] To achieve the above objectives, a remote control and monitoring system for a quick-cut blind valve based on float control is provided, comprising the following units: The data acquisition unit synchronously acquires velocity gradient and angular acceleration based on the float trajectory tracking algorithm and constructs a three-dimensional motion matrix. Then, it uses a pressure-temperature coupled denoising algorithm to extract the effective pressure fluctuation period. The cross-dimensional data analysis unit inputs the three-dimensional motion matrix and effective pressure fluctuation period into a pre-trained residual convolutional neural network. It extracts time-frequency fusion features through parallel time convolutional layers and frequency domain decomposition layers. Based on the time-frequency fusion features, it performs dual verification of fluid phase state. The first verification layer uses a support vector machine classifier to determine the basic phase state, and the second verification layer uses a dynamic time warping algorithm to match the historical working condition database and correct the basic phase state determination result. When the confidence level of the judgment result is greater than the preset threshold, the instruction generation and feedback unit triggers the preset control strategy library to generate standard action instructions. When the confidence level is less than or equal to the preset threshold, the self-learning compensation mechanism is activated to update the frequency domain decomposition layer weights of the residual convolutional neural network online, generate incremental control instructions and superimpose them on the standard instructions, synchronously record the deviation between the compensation instructions and the actual valve action, and construct a negative feedback optimization coefficient matrix.

[0008] As a further improvement to this technical solution, the method for constructing a three-dimensional motion matrix by the data acquisition unit is as follows: Raw float displacement data is captured by an axially arranged Hall sensor array. Moving average filtering is applied to 20 consecutive sampling points of the raw float displacement data, and the displacement change rate between adjacent sampling points is calculated as the velocity gradient reference value. Three-dimensional angular velocity data is acquired by a MEMS gyroscope embedded in the top of the float, and the high-frequency noise components of the three-dimensional angular velocity data are filtered out. The angular acceleration is obtained by performing time differentiation on the processed three-dimensional angular velocity data. The velocity gradient and angular acceleration are time-stamped, and a three-dimensional vector sequence is generated using a Kalman data fusion algorithm. The thermal expansion coefficient of the float material is introduced to compensate for temperature drift in the displacement data, ultimately forming a three-dimensional motion matrix.

[0009] As a further improvement to this technical solution, the pressure-temperature coupled denoising algorithm performs the following: The original pressure waveform is acquired by a pipeline pressure sensor and normalized. Pipeline wall temperature data is acquired simultaneously. The thermal deformation of the current temperature relative to the calibration temperature is calculated based on the linear expansion coefficient of the pipe material. The zero-point offset of the pressure sensor is corrected proportionally. Discrete wavelet transform is applied to the compensated pressure signal to extract the reconstructed signal in the 0.1 to 10 Hz frequency band as the effective pressure fluctuation. The local peak detection algorithm is used to identify the time interval between adjacent peaks, generate a pressure fluctuation period time series, and extract the effective pressure fluctuation period.

[0010] As a further improvement to this technical solution, the residual convolutional neural network processing logic includes: The three-dimensional motion matrix is ​​input into the dilated causal convolutional layer according to the time step. The long-period dependence of the displacement velocity gradient is captured by the dilation coefficient of the convolutional kernel which is multiplied layer by layer, and the displacement feature vector after time dimension compression is output. The pressure fluctuation cycle time series is input into the adjustable Q-factor wavelet transform layer, and the frequency domain mode is separated by a multi-scale filter bank to output the frequency domain feature vector. The displacement feature vector and the frequency domain feature vector are concatenated to form a hybrid feature. The contribution weights of the two types of features are calculated through an attention weight allocation mechanism, and the weighted fusion is used to generate a 128-dimensional time-frequency fusion feature vector.

[0011] As a further improvement to this technical solution, the fluid phase dual verification method includes: The 128-dimensional time-frequency fusion feature vector is input into a pre-trained support vector machine classifier to calculate the probability distribution values ​​of three phase states: liquid single-phase flow, gas-liquid two-phase flow, and transient impingement flow. The phase state corresponding to the maximum probability is taken as the basic judgment result. The dynamic time warping algorithm is used to perform morphological similarity matching between the current pressure fluctuation cycle sequence and the template sequence labeled with phase in the historical operating condition database, and the phase label of the historical template with a similarity of more than 85% is selected as the reference result; When the basic judgment result is consistent with the historical reference result, the phase label is output directly. When they are inconsistent, the historical reference result is used first and the confidence downgrade flag is triggered.

[0012] As a further improvement to this technical solution, the method for determining confidence level and generating threshold is as follows: A sliding window statistical analysis is performed on 10 consecutive phase state determination results. The proportion of the same phase state label within the window is calculated as the real-time confidence level. Based on the standard deviation data of valve execution error in the historical control cycle, the confidence level threshold is adjusted inversely according to the error fluctuation amplitude. When the real-time confidence level is higher than the dynamically adjusted confidence level threshold, the standard control command generation process is initiated; otherwise, the self-learning compensation mechanism is activated.

[0013] As a further improvement to this technical solution, the self-learning compensation mechanism executes the following steps: For the original sensor data segment when the compensation mechanism is triggered, an expanded dataset is generated using an overlapping sliding window resampling technique, and Gaussian white noise is injected to simulate the working condition disturbance. The gradient of the loss function of the frequency domain decomposition layer of the residual convolutional neural network is calculated based on the expanded dataset. The gradient descent algorithm with momentum factor is used to update the filter bank weight parameters. The control strength gain coefficient is calculated according to the weight update amplitude. The amplitude of the original standard action command is amplified according to the gain coefficient to generate an incremental compensation command.

[0014] As a further improvement to this technical solution, the method for superimposing and controlling the execution of the instructions is as follows: Based on the historical action data of the valve actuator, a nonlinear transfer function model including dead time and response hysteresis is established. After the incremental compensation command is input into the PID compensator to eliminate high-frequency oscillation components, it is algebraically superimposed with the standard action command in the time domain. The amplitude of the superimposed command is limited to not exceed the maximum stroke range of the actuator.

[0015] As a further improvement to this technical solution, the negative feedback optimization coefficient matrix is ​​constructed as follows: The deviation data between the actual displacement and the commanded displacement of the valve are collected in real time. The integral value of the absolute error in the time domain and the energy distribution ratio in the frequency domain are calculated respectively. The evolution law of the historical deviation sequence is learned through the long short-term memory network to predict the error decay trend of the next three control cycles. The integral value of the time domain error, the energy ratio in the frequency domain, and the predicted error decay rate are combined into a three-dimensional vector, which is multiplied by the weight coefficient to generate an optimization matrix. This matrix is ​​fed back to the data acquisition unit in real time.

[0016] As a further improvement to this technical solution, the negative feedback optimization coefficient matrix performs the following fault diagnosis method in closed-loop applications: When the integral value of the absolute error in the time domain in the optimization coefficient matrix exceeds the set safety threshold for three consecutive control cycles, and the frequency domain energy ratio exhibits high-frequency oscillation characteristics, the deep fault diagnosis mode is activated. The deep fault diagnosis mode synchronously retrieves the original three-dimensional motion matrix, time-frequency fusion features, and instruction overlay records, and predicts the fault source through correlation analysis.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: In the remote control and monitoring system of a fast-cut blind valve based on float control, the float trajectory tracking algorithm and pressure-temperature coupling denoising algorithm of the data acquisition unit are used to simultaneously acquire the three-dimensional motion matrix of the float and the effective pressure fluctuation period, solving the problem of insufficient multi-source data fusion, shortening the correlation delay between float motion and fluid pressure, and improving the success rate of leakage early warning. The cross-dimensional data analysis unit uses a residual convolutional neural network to extract time-frequency fusion features, combined with a fluid phase dual verification mechanism, to accurately identify complex working conditions such as gas-liquid two-phase flow, reduce the phase misjudgment rate, and reduce valve malfunction. The command generation and feedback unit determines the trigger standard or self-learning compensation command through confidence level determination, and combines negative feedback to optimize the coefficient matrix, adaptively correcting displacement deviations caused by temperature drift, etc., improving valve positioning accuracy and meeting the high reliability control requirements of high-risk pipelines. Attached Figure Description

[0018] Figure 1 This is an overall block diagram of the present invention.

[0019] The meanings of the labels in the diagram are as follows: 1. Data acquisition unit; 2. Cross-dimensional data analysis unit; 3. Instruction generation and feedback unit. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides a remote control and monitoring system for a quick-cut blind valve based on float control. Please refer to [link to relevant documentation]. Figure 1 As shown, it includes the following units: Data acquisition unit 1 synchronously acquires velocity gradient and angular acceleration according to float trajectory tracking algorithm and constructs three-dimensional motion matrix, and then uses pressure-temperature coupling denoising algorithm to extract effective pressure fluctuation period; To achieve precise capture of the float's motion state, data acquisition unit 1 constructs a three-dimensional motion matrix through multi-sensor fusion and dynamic compensation, as detailed below: Since the axial displacement of the float is the core parameter of valve control, the original displacement data of the float is first captured by an axially arranged Hall sensor array. The output voltage is linearly related to the float position. The original displacement data of the float is processed by moving average filtering for 20 consecutive sampling points to eliminate high-frequency electromagnetic interference. After filtering, the displacement change rate between adjacent sampling points is calculated as the velocity gradient reference value. In order to obtain the rotation state of the float, three-dimensional angular velocity data is collected by a MEMS gyroscope embedded in the top of the float. The high-frequency noise component of the three-dimensional angular velocity data is filtered out. A fourth-order Butterworth low-pass filter is used to filter out the high-frequency noise component and eliminate the sudden change in angular velocity caused by the impact of the pipeline fluid. The time derivative is performed on the processed three-dimensional angular velocity data to obtain the angular acceleration, improve the measurement accuracy of the angular acceleration, and meet the requirements for capturing the rotation state of the float under the gas-liquid two-phase flow condition. Due to the slight deviation in the sampling clocks of the Hall sensor and the gyroscope, a hardware synchronization triggering mechanism is used to align the velocity gradient and angular acceleration with timestamps. The velocity gradient and angular acceleration are combined into a six-dimensional state vector to construct a Kalman data fusion model. A three-dimensional vector sequence is generated using the Kalman data fusion algorithm to reduce the root mean square error of the fused data. The float temperature is collected in real time, and the thermal expansion coefficient of the float material is queried and introduced. The temperature drift is calculated based on the thermal expansion coefficient of the float material, and temperature drift compensation is performed on the displacement data. Finally, the compensated velocity gradient and angular acceleration are arranged in time sequence to form a three-dimensional motion matrix. Each row of the matrix corresponds to the six-dimensional motion state at a certain moment, improving the correction rate of displacement error caused by temperature drift. This provides a high-precision motion state data foundation for subsequent fluid phase analysis and control command generation.

[0022] To accurately extract the effective pressure fluctuation period of the pipeline and eliminate the interference of temperature drift on the pressure signal, the pressure-temperature coupled denoising algorithm achieves signal purification through multi-step preprocessing and frequency domain analysis. The specific implementation method is as follows: Because pipeline pressure fluctuations and temperature changes are coupled, the original pressure waveform is first acquired using a pipeline pressure sensor and then normalized. Simultaneously, pipeline wall temperature data is obtained. The thermal deformation relative to the calibration temperature is calculated based on the linear expansion coefficient of the pipe material. This deformation causes additional stress on the sensor diaphragm, leading to zero-point offset. The zero-point offset of the pressure sensor is then corrected using this value. The corrected pressure signal is normalized to the [-1, 1] interval for subsequent frequency domain analysis. Discrete wavelet transform decomposition is performed on the compensated pressure signal to obtain approximation coefficients (low frequency) and detail coefficients (high frequency). The reconstructed signal in the 0.1 to 10 Hz frequency band is extracted as the effective pressure fluctuation. The 0.1–10 Hz frequency band is extracted through frequency band division: Layer 5 The approximation coefficients correspond to 0-3.125Hz, the fourth layer of detail coefficients corresponds to 3.125-6.25Hz, and the third layer of detail coefficients corresponds to 6.25-12.5Hz. Only the first two layers of detail coefficients and the fifth layer of approximation coefficients are retained for reconstruction. Pipe vibration and slow drift noise are filtered out. Adaptive threshold local peak detection is used on the reconstructed pressure signal. The adaptive threshold local peak detection algorithm identifies the time interval between adjacent peaks. First, the signal mean and standard deviation are calculated. The peak threshold is set to the mean + 1.5 standard deviations to avoid interference from small fluctuations. The signal is traversed to find peak points, and the time interval between adjacent peaks is calculated to generate a pressure fluctuation cycle time series. The effective pressure fluctuation cycle is extracted. To eliminate outliers, a filtering principle of 3 times the standard deviation is used. If the time interval between adjacent peaks exceeds the mean time interval... A time interval of 3 times the standard deviation is considered an invalid period and corrected by linear interpolation.

[0023] The cross-dimensional data analysis unit 2 inputs the three-dimensional motion matrix and the effective pressure fluctuation period into the pre-trained residual convolutional neural network. It extracts time-frequency fusion features through parallel time convolutional layers and frequency domain decomposition layers. Based on the time-frequency fusion features, it performs dual verification of fluid phase state. The first verification layer uses a support vector machine classifier to determine the basic phase state, and the second verification layer uses a dynamic time warping algorithm to match the historical working condition database and correct the basic phase state determination result. To achieve cross-dimensional feature fusion of the float's motion state and pressure fluctuations, a residual convolutional neural network constructs a time-frequency fusion feature vector through dilated causal convolution, adjustable Q-factor wavelet transform, and attention mechanism. The specific implementation method is as follows: The processing logic of residual convolutional neural networks includes: Since the long-period dependence of the float's motion is crucial for fluid phase identification, the three-dimensional motion matrix is ​​input into an expanded causal convolutional layer according to the time step. This layer adopts a three-layer convolutional structure, with the expansion coefficient of each convolutional kernel increasing by a power of 2. By progressively increasing the expansion coefficient of the convolutional kernel, the long-period dependence of displacement velocity gradient is captured, ensuring causality in the time dimension. Taking an input matrix with a time step of 128 as an example, the first convolutional layer captures the correlation between velocity gradient and angular acceleration in adjacent time steps through a receptive field of 3×1 (d=1). The second layer captures the motion trend within 8 time steps through a receptive field of 7×1 (d=2). The third layer captures the long-period dependence of 16 time steps through a receptive field of 15×1 (d=4). After batch normalization and ReLU activation, the output of each layer skips downsampling through residual connections, and finally outputs a displacement feature vector with a dimension of 16×6, where d represents the layer number and the output is a displacement feature vector compressed in the time dimension. To separate the multi-scale frequency domain patterns of pressure fluctuations, the time series of pressure fluctuation cycles is input into an adjustable Q-factor wavelet transform layer. This layer contains 8 bandpass filters. The Q-factor (quality factor) can be dynamically adjusted through backpropagation, with an initial value of Q=5 and an adjustment range of 2-10. Each filter group corresponds to a different center frequency. Through multi-scale filtering, 8 frequency domain patterns are separated, and the energy proportion of each pattern constitutes an 8-dimensional frequency domain feature vector. Compared with wavelet transform with a fixed Q value, this improves the accuracy of frequency domain feature extraction. The frequency domain patterns are separated by a multi-scale filter bank, and the frequency domain feature vector is output. The displacement feature vector and the frequency domain feature vector are concatenated to form a hybrid feature. The contribution weights of the two types of features are calculated using an attention weight allocation mechanism, and then weighted and fused to generate a 128-dimensional time-frequency fusion feature vector. The 16×6 displacement feature vector (flattened to 96 dimensions) is concatenated with the 8-dimensional frequency domain feature vector to form a 104-dimensional hybrid feature, which is then input into the attention weight allocation module. This module contains two fully connected layers: the first layer maps the 104-dimensional feature to 32 dimensions using the LeakyReLU activation function; the second layer outputs a 2-dimensional weight vector [w1, w2], which, after softmax normalization, satisfies w1 + w2 = 1. The weight calculation is based on feature correlation: the higher the correlation between the velocity gradient abrupt change in the displacement feature and the high-frequency energy in the frequency domain feature, the larger the weight of w1, and vice versa. An intermediate feature is generated through weighted fusion consensus. The fused feature is then reduced to 128 dimensions by a 1×1 convolution, where the first 96 dimensions correspond to the displacement feature weight allocation result, and the last 32 dimensions correspond to the frequency domain feature mapping result. This fusion method enhances the complementarity of time-frequency features.

[0024] To achieve accurate identification of fluid phase states and address the misjudgment problem of single algorithms under complex working conditions, a dual verification method for fluid phase states utilizes a dual mechanism of support vector machine classification and dynamic time warping algorithm matching to form a complementary verification method. The dual verification method for fluid phase states includes: A 128-dimensional time-frequency fusion feature vector is input into a pre-trained support vector machine (SVM) classifier. This classifier uses a radial basis function kernel function, with kernel parameters determined through grid search. The penalty parameter C=100. The training dataset contains 10,000 labeled samples: 4,000 for liquid single-phase flow, 3,500 for gas-liquid two-phase flow, and 2,500 for transient impingement flow. Each sample contains a time-frequency feature vector and label for the corresponding phase. For the input feature vector... Support vector machines compute decision functions. Generate scores for three types of phase states, among which For radial basis functions and kernel functions, , For support vector weights, As a bias term, the score is converted into a probability distribution value using the Softmax function. , For each of the three phase states—liquid single-phase flow, gas-liquid two-phase flow, and transient impingement flow—the probability distribution values ​​of the three phase states are calculated, and the phase state corresponding to the maximum probability value is taken as the basic judgment result. A dynamic time warping algorithm is used to match the current pressure fluctuation cycle sequence with the template sequences labeled with phases in the historical operating condition database based on morphological similarity. The phase label of the historical template with a similarity of more than 85% is selected as the reference result. The historical operating condition database contains 2,000 sets of pressure fluctuation cycle sequence templates labeled with phases, stored according to phase: 800 sets for liquid single-phase flow, 700 sets for gas-liquid two-phase flow, and 500 sets for transient impact flow. The length of each template sequence is standardized to 50 cycle points. For the current pressure fluctuation cycle sequence, detrending processing is first performed: the mean of the sequence is calculated and subtracted to eliminate the influence of slow changes in pipeline pressure. Then, the sequence length is unified to 50 points through linear interpolation, consistent with the dimension of the template sequence. When the basic judgment result is consistent with the historical reference result, the phase label is directly output. When they are inconsistent, the historical reference result is used first and the confidence downgrade flag is triggered.

[0025] Confidence level determination and threshold generation methods: A sliding window statistical analysis is performed on 10 consecutive phase state determination results. The proportion of the same phase state label within the window is calculated as the real-time confidence level. Based on the standard deviation data of valve execution error in the historical control cycle, the confidence level threshold is adjusted inversely according to the error fluctuation amplitude. When the real-time confidence level is higher than the dynamically adjusted confidence level threshold, the standard control command generation process is initiated; otherwise, the self-learning compensation mechanism is activated.

[0026] To accurately determine the reliability of phase state determination results and adaptively adjust the control strategy, the confidence level determination and threshold generation method uses sliding window statistics and dynamic threshold adjustment to achieve intelligent evaluation of the system's operating state. The specific implementation method is as follows: A sliding window statistical analysis is performed on 10 consecutive phase determination results (from the output of fluid phase dual verification). The window shifts to the right with each new determination result, retaining the latest 10 labels each time. For example, if the current phase label sequence in the window is [gas-liquid two-phase flow, gas-liquid two-phase flow, transient impingement flow, gas-liquid two-phase flow, gas-liquid two-phase flow, liquid single-phase flow, gas-liquid two-phase flow, gas-liquid two-phase flow, gas-liquid two-phase flow, gas-liquid two-phase flow], and gas-liquid two-phase flow appears 8 times, then the real-time confidence level is calculated as 8 / 10 × 100% = 80%. To eliminate the impact of accidental misjudgments on the confidence level and to avoid drastic fluctuations in confidence level caused by a single misjudgment, median filtering is used to denoise the phase labels in the window. Ten labels are counted according to phase category. Only phase labels that appear 3 or more times are retained for confidence calculation. If all phase labels in the window appear less than 3 times, the confidence is initialized to 50%. For example, if the labels in the window are [gas-liquid two-phase flow, transient impingement flow, gas-liquid two-phase flow, transient impingement flow, liquid single-phase flow, gas-liquid two-phase flow, transient impingement flow, liquid single-phase flow, gas-liquid two-phase flow, liquid single-phase flow], and each phase appears 3-4 times, the confidence is initialized to 50%. More judgment results are needed to clarify the phase trend. Valve execution error data (deviation between actual displacement and commanded displacement) within the historical control cycle is stored in a circular buffer. The standard deviation of the error is calculated in real time. The confidence threshold is adjusted using an inverse proportional formula. The basic threshold is set at 70%. To adjust the coefficients, this formula ensures that the error fluctuation amplitude is inversely proportional to the threshold exponentially. This guarantees that the system lowers the decision threshold and promptly activates the compensation mechanism when the execution error is large. The threshold calculation module updates the historical error standard deviation every second and adjusts it according to the latest... The threshold is refreshed, and to prevent frequent threshold changes, a 0.5% adjustment dead zone is set. If the difference between the current threshold and the newly calculated threshold is less than 0.5%, the original threshold is maintained to avoid threshold jitter caused by noise. When the real-time confidence level is greater than or equal to the dynamically adjusted confidence threshold, the standard control command generation process is initiated. Standard action commands corresponding to the determined phase state are retrieved from the preset control strategy library. The strategy library contains 20 predefined strategies, each indexed in multiple dimensions based on phase state, pressure level, and temperature range to ensure the specificity of the commands. When the real-time confidence level is less than the dynamic threshold, the self-learning compensation mechanism is activated. First, the current sensor data segment, i.e., the touch... The data before and after the launch time (100ms) are expanded, and 50 new samples are generated using overlapping sliding window resampling technology. Gaussian white noise (with a standard deviation of 5% of the original data) is injected to simulate the operating condition disturbance. Then, the loss function gradient of the frequency domain decomposition layer of the residual convolutional neural network is calculated based on the expanded dataset. The filter bank weight parameters are updated using a stochastic gradient descent algorithm with a driving factor. Finally, the control strength gain coefficient is calculated based on the weight update amplitude. The standard command amplitude is amplified according to the gain coefficient to generate an incremental compensation command. This effectively solves the problems of coarse confidence assessment and lagging threshold adjustment in traditional technologies, and provides a reliable decision basis for the precise control of the fast-cut blind valve.

[0027] Implementation steps of the self-learning compensation mechanism: For the original sensor data segment when the compensation mechanism is triggered, an expanded dataset is generated using an overlapping sliding window resampling technique, and Gaussian white noise is injected to simulate the working condition disturbance. The gradient of the loss function of the frequency domain decomposition layer of the residual convolutional neural network is calculated based on the expanded dataset. The gradient descent algorithm with momentum factor is used to update the filter bank weight parameters. The control strength gain coefficient is calculated according to the weight update amplitude. The amplitude of the original standard action command is amplified according to the gain coefficient to generate an incremental compensation command.

[0028] To address the problem of model prediction bias under complex operating conditions, a self-learning compensation mechanism achieves adaptive updates of the control strategy through a closed-loop process of data augmentation, model optimization, and command correction. The specific implementation method is as follows: When the confidence level triggers the compensation mechanism, the original sensor data segments are extracted 50ms before and after the trigger time. The overlapping sliding window technique is used to resample the data segments. The window length is set to 60 points (60ms) and the overlap rate is 50% (i.e., 30 points are slid each time). A new sample set is generated to ensure the integrity of the data features. Gaussian white noise is injected into each new sample to simulate the actual working condition disturbance. The noise standard deviation is set to 15% of the original data standard deviation. The expanded dataset is input into the frequency domain decomposition layer (adjustable Q-factor wavelet transform layer) of the residual convolutional neural network. The cross-entropy between the current phase state determination result and the actual phase state label is used as the loss function. The gradient of the loss function of the frequency domain decomposition layer is calculated through the backpropagation algorithm. The center frequency and Q-factor parameters of the filter bank are optimized. The filter bank parameters are updated using the stochastic gradient descent algorithm with a momentum factor (0.9). The update amplitude of the weight parameters of the frequency domain decomposition layer is calculated. The weight update amplitude is mapped to the control strength gain coefficient. The amplitude of the original standard action command is amplified according to the gain coefficient to generate an incremental compensation command, which provides self-evolution capability for the intelligent control of the fast-cut blind valve.

[0029] When the confidence level of the judgment result is greater than the preset threshold, the instruction generation and feedback unit 3 triggers the preset control strategy library to generate standard action instructions. When the confidence level is less than or equal to the preset threshold, the self-learning compensation mechanism is started, the frequency domain decomposition layer weights of the residual convolutional neural network are updated online, incremental control instructions are generated and superimposed on the standard instructions, the deviation between the compensation instructions and the actual valve action is recorded synchronously, and a negative feedback optimization coefficient matrix is ​​constructed.

[0030] Methods for superimposing and controlling the execution of instructions: Based on the historical action data of the valve actuator, a nonlinear transfer function model including dead time and response hysteresis is established. After the incremental compensation command is input into the PID compensator to eliminate high-frequency oscillation components, it is algebraically superimposed with the standard action command in the time domain. The amplitude of the superimposed command is limited to not exceed the maximum stroke range of the actuator.

[0031] To ensure the accuracy and safety of command execution, the command superposition and execution control method achieves the organic integration of standard commands and compensation commands by establishing a nonlinear model, PID filtering, and dynamic limiting. The specific implementation method is as follows: Historical action data of the valve actuator under different command amplitudes (including command sequences and actual displacement feedback, with a sample size of ≥200 sets) were collected. Dead time and hysteresis parameters were identified through step response testing. Dead time is defined as the time window during which the actuator does not respond after the command is issued, and is determined by statistically analyzing the average delay at 50% of the command amplitude. Response hysteresis is calculated by the displacement difference between the forward and reverse strokes. A nonlinear transfer function model including dead time and hysteresis is constructed. The linear transfer function and the nonlinear element are connected in series to form a complete model. The incremental compensation command is input into the PID compensator. The filtered incremental compensation command and the standard action command are algebraically superimposed in the time domain. During superposition, the timing consistency of the commands is considered: the timestamp alignment ensures that the two commands are superimposed within the same sampling period (1ms) to avoid sudden shocks. The maximum stroke range of the actuator is defined, and bidirectional amplitude limiting is implemented on the superimposed command. To avoid command abrupt changes caused by amplitude limiting, an S-shaped smoothing function is used to process the amplitude limiting boundary to extend the service life of the actuator.

[0032] Construction of negative feedback optimization coefficient matrix: The deviation data between the actual displacement and the commanded displacement of the valve are collected in real time. The integral value of the absolute error in the time domain and the energy distribution ratio in the frequency domain are calculated respectively. The evolution law of the historical deviation sequence is learned through the long short-term memory network to predict the error decay trend of the next three control cycles. The integral value of the time domain error, the energy ratio in the frequency domain, and the predicted error decay rate are combined into a three-dimensional vector, which is multiplied by the weight coefficient to generate an optimization matrix. This matrix is ​​fed back to the data acquisition unit 1 in real time.

[0033] To achieve dynamic optimization of valve control deviation and continuous improvement of system performance, a closed-loop control mechanism is constructed using a negative feedback optimization coefficient matrix through multi-dimensional error analysis and time-series prediction. The specific implementation method is as follows: Since the multi-dimensional characteristics of valve control deviation are the foundation for optimization, high-precision displacement sensors are used to collect real-time deviation data between the actual and commanded valve displacements. For the collected deviation sequence, the time-domain absolute error integral value is calculated, which is the sum of the absolute values ​​of the deviation within a single control cycle. This value reflects the overall scale of the error. Simultaneously, frequency domain analysis is performed on the deviation sequence, using Fast Fourier Transform to calculate the energy distribution ratios of low, medium, and high frequencies, thus characterizing the frequency characteristics of the error and laying a data foundation for precise optimization. Given the significant value of historical deviation sequence evolution for future error prediction, the time-domain error integral values ​​and frequency-domain energy distribution ratios of the past 20 control cycles are used as input to construct a Long Short-Term Memory (LSTM) network model. This network contains two layers of memory units, each with 128 neurons. By learning the time dependency of the deviation sequence, the error decay rate for the next three control cycles is predicted, i.e., the error in the next cycle versus the error in the current cycle. The ratio enables the system to perceive error change trends in advance, providing forward-looking support for optimization decisions. To transform multi-dimensional error characteristics into executable optimization instructions, a three-dimensional feature vector is formed by the normalized value of the time-domain error integral in the current period, the proportion of high-frequency components in the frequency-domain energy ratio, and the first-step prediction error attenuation rate. The feature vector is linearly transformed by a trained weight matrix to generate a 6-dimensional optimization matrix. For example, the three-dimensional vector is [0.5, 0.3, 0.8]. After weight mapping, a matrix containing elements such as the sensor sampling frequency adjustment coefficient and the Kalman filter parameter correction value is obtained. This matrix is ​​fed back to the data acquisition unit 1 in real time through serial communication to dynamically adjust parameters such as the sampling frequency of the Hall sensor and the float thermal expansion compensation coefficient. This closed-loop feedback mechanism improves the convergence speed of control error, realizes the self-optimization and continuous improvement of system performance, and effectively solves the problems of fixed parameters and poor adaptability in traditional control schemes.

[0034] The negative feedback optimization coefficient matrix performs the following fault diagnosis method in closed-loop applications: When the integral value of the absolute error in the time domain in the optimization coefficient matrix exceeds the set safety threshold for three consecutive control cycles, and the frequency domain energy ratio exhibits high-frequency oscillation characteristics, the deep fault diagnosis mode is activated. The deep fault diagnosis mode synchronously retrieves the original three-dimensional motion matrix, time-frequency fusion features, and instruction overlay records, and predicts the fault source through correlation analysis.

[0035] To promptly detect and locate potential faults in valve control systems, the negative feedback optimization coefficient matrix performs fault diagnosis through multi-dimensional feature correlation in closed-loop applications. The specific implementation method is as follows: Since single-dimensional error characteristics may lead to misjudgments, the system simultaneously monitors error characteristics in both the time and frequency domains to improve the accuracy of fault warnings. First, it calculates the time-domain absolute error integral value in the optimization coefficient matrix in real time, i.e., accumulating the absolute value of displacement deviation within each control cycle. When this value exceeds the set safety threshold for three consecutive cycles, it indicates a persistently large control deviation in the system. Second, it performs frequency domain analysis on the deviation sequence, calculating the proportion of high-frequency energy to total energy. If this proportion exceeds 30% for three consecutive cycles and exhibits periodic fluctuations, it is determined that high-frequency oscillation characteristics exist. This dual monitoring mechanism reduces the false alarm rate and ensures the reliability of the warning. To comprehensively analyze the root cause of the fault, after the deep diagnostic mode is activated, three types of key data are retrieved simultaneously: the original three-dimensional motion matrix of the past 10 control cycles is obtained through data acquisition unit 1, which contains six-dimensional motion data such as the float's velocity gradient and angular acceleration; the corresponding time-frequency fusion feature vector sequence is retrieved from the cross-dimensional data analysis unit 2. This feature is generated by a residual convolutional neural network, reflecting the comprehensive characteristics of the fluid phase and the time and frequency domains. The system simultaneously retrieves the instruction superposition records from instruction generation and feedback unit 3, including the time-domain superposition sequence of standard instructions and incremental compensation instructions, to ensure timing consistency. After data retrieval, cyclic redundancy check is used for integrity verification to avoid diagnostic bias due to data errors and improve the credibility of subsequent analysis. To locate the root cause of the fault from complex data, a dynamic correlation analysis method is used to cross-validate multi-source data. First, a time-feature correlation matrix is ​​constructed, and the angular acceleration mutation points in the three-dimensional motion matrix, the high-frequency energy peaks in the time-frequency features, and the abnormal pulses in the instruction sequence are spatiotemporally matched. The time difference and correlation coefficient between each feature are calculated. For example, it is found that the moment when the float angular acceleration suddenly increases in a certain period is synchronized with the overshoot pulse in the instruction superposition sequence, and the high-frequency energy of the time-frequency features suddenly increases, indicating that there may be fluid impact or mechanical jamming. Then, a random forest classifier trained on a historical fault database is used as input, and the above-mentioned correlation features are input to output the probability distribution of each fault source. This correlation analysis method combines data-driven and mechanism analysis, providing an intelligent guarantee for the safe operation of pipelines.

[0036] In this invention, the data acquisition unit 1 constructs a three-dimensional motion matrix using a float trajectory tracking algorithm and extracts the effective pressure fluctuation period using a pressure-temperature coupled denoising algorithm. The cross-dimensional data analysis unit 2 extracts time-frequency fusion features using a residual convolutional neural network and achieves dual verification of fluid phase state through a support vector machine and a dynamic time warping algorithm. The command generation and feedback unit 3 triggers a standard control strategy or a self-learning compensation mechanism based on confidence level and corrects control deviations through negative feedback optimization of the coefficient matrix. This solves the problems of insufficient multi-source data fusion and poor adaptability to complex working conditions in traditional technologies, realizing precise remote control and intelligent monitoring of the quick-cut blind valve, and improving the success rate of leakage early warning and the reliability of valve operation.

[0037] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A remote control and monitoring system for a quick-cut blind valve based on float control, characterized in that, Includes the following units: The data acquisition unit (1) synchronously acquires the velocity gradient and angular acceleration according to the float trajectory tracking algorithm and constructs a three-dimensional motion matrix, and then uses the pressure-temperature coupling denoising algorithm to extract the effective pressure fluctuation period; The cross-dimensional data analysis unit (2) inputs the three-dimensional motion matrix and effective pressure fluctuation period into the pre-trained residual convolutional neural network. It extracts time-frequency fusion features through parallel time convolutional layers and frequency domain decomposition layers. Based on the time-frequency fusion features, it performs dual verification of fluid phase state. The first verification layer judges the basic phase state through a support vector machine classifier. The second verification layer uses a dynamic time warping algorithm to match the historical working condition database and correct the basic phase state judgment result. The instruction generation and feedback unit (3) triggers the generation of standard action instructions from the preset control strategy library when the confidence level of the judgment result is greater than the preset threshold. When the confidence level is less than or equal to the preset threshold, it starts the self-learning compensation mechanism, updates the frequency domain decomposition layer weights of the residual convolutional neural network online, generates incremental control instructions and superimposes them on the standard instructions, synchronously records the deviation between the compensation instructions and the actual valve action, and constructs a negative feedback optimization coefficient matrix.

2. The remote control and monitoring system for a quick-cut blind valve based on float control according to claim 1, characterized in that: The method for the data acquisition unit (1) to construct the three-dimensional motion matrix is ​​as follows: Raw float displacement data is captured by an axially arranged Hall sensor array. Moving average filtering is applied to 20 consecutive sampling points of the raw float displacement data, and the displacement change rate between adjacent sampling points is calculated as the velocity gradient reference value. Three-dimensional angular velocity data is acquired by a MEMS gyroscope embedded in the top of the float, and the high-frequency noise components of the three-dimensional angular velocity data are filtered out. The angular acceleration is obtained by performing time differentiation on the processed three-dimensional angular velocity data. The velocity gradient and angular acceleration are time-stamped, and a three-dimensional vector sequence is generated using a Kalman data fusion algorithm. The thermal expansion coefficient of the float material is introduced to compensate for temperature drift in the displacement data, ultimately forming a three-dimensional motion matrix.

3. The remote control and monitoring system for a quick-cut blind valve based on float control according to claim 2, characterized in that: The pressure-temperature coupled denoising algorithm is executed as follows: The original pressure waveform is acquired by a pipeline pressure sensor and normalized. Pipeline wall temperature data is acquired simultaneously. The thermal deformation of the current temperature relative to the calibration temperature is calculated based on the linear expansion coefficient of the pipe material. The zero-point offset of the pressure sensor is corrected proportionally. Discrete wavelet transform is applied to the compensated pressure signal to extract the reconstructed signal in the 0.1 to 10 Hz frequency band as the effective pressure fluctuation. The local peak detection algorithm is used to identify the time interval between adjacent peaks, generate a pressure fluctuation period time series, and extract the effective pressure fluctuation period.

4. The remote control and monitoring system for a quick-cut blind valve based on float control according to claim 3, characterized in that: The residual convolutional neural network processing logic includes: The three-dimensional motion matrix is ​​input into the dilated causal convolutional layer according to the time step. The long-period dependence of the displacement velocity gradient is captured by the dilation coefficient of the convolutional kernel which is multiplied layer by layer, and the displacement feature vector after time dimension compression is output. The pressure fluctuation cycle time series is input into the adjustable Q-factor wavelet transform layer, and the frequency domain mode is separated by a multi-scale filter bank to output the frequency domain feature vector. The displacement feature vector and the frequency domain feature vector are concatenated to form a hybrid feature. The contribution weights of the two types of features are calculated through an attention weight allocation mechanism, and the weighted fusion is used to generate a 128-dimensional time-frequency fusion feature vector.

5. The remote control and monitoring system for a quick-cut blind valve based on float control according to claim 4, characterized in that: The fluid phase dual verification method includes: The 128-dimensional time-frequency fusion feature vector is input into a pre-trained support vector machine classifier to calculate the probability distribution values ​​of three phase states: liquid single-phase flow, gas-liquid two-phase flow, and transient impingement flow. The phase state corresponding to the maximum probability is taken as the basic judgment result. The dynamic time warping algorithm is used to perform morphological similarity matching between the current pressure fluctuation cycle sequence and the template sequence labeled with phase in the historical operating condition database, and the phase label of the historical template with a similarity of more than 85% is selected as the reference result; When the basic judgment result is consistent with the historical reference result, the phase label is output directly. When they are inconsistent, the historical reference result is used first and the confidence downgrade flag is triggered.

6. The remote control and monitoring system for a quick-cut blind valve based on float control according to claim 5, characterized in that: The confidence level determination and threshold generation method are as follows: A sliding window statistical analysis is performed on 10 consecutive phase state determination results. The proportion of the same phase state label within the window is calculated as the real-time confidence level. Based on the standard deviation data of valve execution error in the historical control cycle, the confidence level threshold is adjusted inversely according to the error fluctuation amplitude. When the real-time confidence level is higher than the dynamically adjusted confidence level threshold, the standard control command generation process is initiated; otherwise, the self-learning compensation mechanism is activated.

7. The remote control and monitoring system for a quick-cut blind valve based on float control according to claim 6, characterized in that: The self-learning compensation mechanism is implemented in the following steps: For the original sensor data segment when the compensation mechanism is triggered, an expanded dataset is generated using an overlapping sliding window resampling technique, and Gaussian white noise is injected to simulate the working condition disturbance. The gradient of the loss function of the frequency domain decomposition layer of the residual convolutional neural network is calculated based on the expanded dataset. The gradient descent algorithm with momentum factor is used to update the filter bank weight parameters. The control strength gain coefficient is calculated according to the weight update amplitude. The amplitude of the original standard action command is amplified according to the gain coefficient to generate an incremental compensation command.

8. The remote control and monitoring system for a quick-cut blind valve based on float control according to claim 7, characterized in that: The method for superimposing and controlling the execution of the instructions: Based on the historical action data of the valve actuator, a nonlinear transfer function model including dead time and response hysteresis is established. After the incremental compensation command is input into the PID compensator to eliminate high-frequency oscillation components, it is algebraically superimposed with the standard action command in the time domain. The amplitude of the superimposed command is limited to not exceed the maximum stroke range of the actuator.

9. The remote control and monitoring system for a quick-cut blind valve based on float control according to claim 8, characterized in that: Construction of the negative feedback optimization coefficient matrix: The deviation data between the actual displacement and the commanded displacement of the valve are collected in real time. The integral value of the absolute error in the time domain and the energy distribution ratio in the frequency domain are calculated respectively. The evolution law of the historical deviation sequence is learned through the long short-term memory network, and the error decay trend of the next three control cycles is predicted. The integral value of the time domain error, the energy ratio in the frequency domain, and the predicted error decay rate are combined into a three-dimensional vector, which is multiplied by the weight coefficient to generate an optimization matrix. This matrix is ​​fed back to the data acquisition unit in real time (1).

10. The remote control and monitoring system for a quick-cut blind valve based on float control according to claim 9, characterized in that: The negative feedback optimization coefficient matrix performs the following fault diagnosis method in closed-loop applications: When the integral value of the absolute error in the time domain in the optimization coefficient matrix exceeds the set safety threshold for three consecutive control cycles, and the frequency domain energy ratio exhibits high-frequency oscillation characteristics, the deep fault diagnosis mode is activated. The deep fault diagnosis mode synchronously retrieves the original three-dimensional motion matrix, time-frequency fusion features, and instruction overlay records, and predicts the fault source through correlation analysis.

Images