A water pump dynamic start-stop and pressure stabilization control method based on working condition identification

Abstract

This invention discloses a dynamic start-stop and pressure stabilization control method for water pumps based on operating condition identification, relating to the fields of mechatronics and automation control technology. The method includes the following steps: S1, outputting the predicted arrival time node of the reflected wave; S2, outputting operating condition attribute labels; S3, parsing the operating condition attribute labels; S4, inputting the dynamic fluctuation characteristics and graph topology space into an improved DCRNN model, introducing a bandgap filter based on acoustic impedance constraints and a resonant coupling diffusion mechanism to perform frequency domain filtering suppression and impedance matching coupling; S5, outputting a dynamic start-stop delay reference; S6, outputting the primary water hammer wave; S7, outputting the error compensation coefficient; S8, performing adaptive linkage correction on the dynamic start-stop delay reference and wave amplitude evolution reference. This invention overcomes the limitations of traditional methods, such as static threshold fixation, semantic simplicity, and neglect of spatiotemporal coupling topological constraints, providing an efficient solution for proactive defense against transient flow in fluid pipeline networks.

Description

Technical Field

[0001] This invention relates to the field of mechatronics and automation control technology, and in particular to a method for dynamic start-stop and pressure stabilization control of water pumps based on operating condition identification. Background Technology

[0002] With the continuous expansion of complex pipeline systems and the increasing complexity of their operating conditions, classical control methods face severe real-time computational challenges in accurately suppressing water hammer transient flows and achieving dynamic pressure stability. Existing conventional water hammer defense strategies, such as simple delayed start-stop and constant pressure compensation control, while utilizing basic valve regulation logic to reduce transient impacts, primarily rely on fixed empirical thresholds and static pressure deviations to generate control commands. Control methods based solely on static deviations ignore the complex physical topology implicit in the original pipeline system (such as pipe impedance distribution and nodal acoustic admittance) and deep spatiotemporal fluctuation coupling, leading to structural distortion of the high-dimensional flow field space when constructing the water hammer evolution model. This limits the accuracy of dynamic pump start-stop and precise pressure counterbalancing under complex operating conditions. Furthermore, classical water hammer suppression methods often struggle to fully utilize the spatiotemporal characteristics of transient flows in pipeline fluctuation data to constrain the delayed computation space, resulting in severe phase misalignment when dealing with the superposition and interference of multi-source reflected waves, increasing the risk of system oscillations and equipment damage.

[0003] Therefore, how to provide a method for dynamic start-up and pressure stabilization control of water pumps based on operating condition identification is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a dynamic start-stop and pressure stabilization control method for water pumps based on operating condition identification. It employs a bandgap filtering and resonant coupling diffusion mechanism based on acoustic impedance constraints, along with a closed-loop feedback optimization mechanism using residual feedforward adaptive correction. This method performs deep spatiotemporal dependency extraction on the dynamic fluctuation characteristics and graph topology space using an improved DCRNN model. High-dimensional physical evolution features, including the upper and lower boundary thresholds of the local resonant frequency bandgap and the resonant coupling space feature tensor, are extracted. Based on these features, standardized residuals are calculated to output error compensation coefficients. These error compensation coefficients are mapped to adaptive step-size scaling factors and linked correction feature vectors. Evolutionary step-size mapping is performed along the temporal step to generate the target derivation state transition matrix. The linked correction feature vector and the target derivation state transition matrix are injected as feedforward driving terms into the improved DCRNN model. The hidden states are updated through matrix multiplication, and the network parameters are iteratively corrected through rolling decoding until the dynamic deviation convergence condition is met. This mechanism effectively eliminates frequency domain noise interference and phase misalignment during the water hammer evolution state deduction process by establishing a closed-loop feedback path from "bandgap filter frequency domain zeroing" and "standardized residual calculation" to "state transition matrix feedforward mapping". This ensures that the generated dynamic start-stop delay benchmark can dynamically maintain the complex physical topological constraints of the transient flow in the pipeline network, achieving the technical effect of improving the accuracy of dynamic start-stop phase offsetting and pressure stability of water pumps while suppressing the superposition interference of multi-source reflected waves. This invention overcomes the limitations of traditional methods such as static threshold fixing, semantic simplicity, and neglect of spatiotemporal coupling topological constraints, providing an efficient solution for active defense against transient flow in fluid pipeline networks.

[0005] A method for dynamic start-up and pressure stabilization control of a water pump based on operating condition identification according to an embodiment of the present invention specifically includes: S1. Dynamically collect pipeline network parameters and pipe segment topology length, correct wave velocity through transient flow solution, extract wave propagation benchmark and steady-state constraint characteristics, and predict the arrival time node of reflected wave along the time axis. S2. Real-time acquisition of pipeline network operation status data, extraction of dynamic fluctuation features and operation parameter sequences through wavelet decomposition and sliding window, extraction of trajectory slope features of two-dimensional phase plane, and mapping output of operating condition attribute labels; S3. Parse the working condition attribute tags to determine the trigger status. If there is no trigger, maintain the status and update it in a closed loop periodically; otherwise, construct the topology space based on the pipeline topology. S4. The dynamic wave characteristics and graph topology space are input into the improved DCRNN model. A bandgap filter and resonant coupling diffusion mechanism based on acoustic impedance constraints are introduced to perform frequency domain filtering suppression and impedance matching coupling. After deep spatiotemporal dependency extraction and mapping, the primary water hammer evolution state, attribute identifier and wave amplitude evolution benchmark are output. S5. Map the initial water hammer evolution state and attribute identifier to the target phase time difference benchmark, combine the predicted arrival time node to perform inverse spatiotemporal calculation, and output the dynamic start-stop delay benchmark. S6. Based on the current system clock superimposed with dynamic start-stop delay reference, the absolute execution time is anchored, and the frequency converter is triggered in real time at the absolute execution time to perform the disconnection action and actively evolve and output the primary water hammer wave. S7. Real-time acquisition of pipeline pressure response under the action of primary water hammer wave, dynamic error calculation combined with primary water hammer evolution state and steady-state constraint characteristics, and output error compensation coefficient. S8. Based on the error compensation coefficient, perform adaptive linkage correction on the dynamic start-stop delay benchmark and amplitude evolution benchmark, and feed forward to the dynamic feature extraction and time series deduction stage to iteratively drive the next dynamic start-stop.

[0006] Optionally, S1 specifically includes: S11. Dynamically collect multi-point pressure fluctuation sequences and pipe segment topology lengths of the pipeline network, statistically analyze the spatiotemporal gradient distribution of transient flow based on the pressure fluctuation sequences, perform multi-dimensional feature cross-mapping calculations on the spatiotemporal gradient distribution and pipe segment topology lengths, and output an adaptively corrected wave velocity field. S12. Based on the adaptive correction of the wave velocity field, perform frequency domain transformation, extract the statistical distribution characteristics of wave energy in the frequency band as the wave propagation benchmark, and simultaneously extract the mean value of the amplitude envelope of the pressure wave sequence in the time window as the steady-state constraint feature. S13. Construct a dual judgment threshold based on the wave propagation benchmark and steady-state constraint characteristics, perform extreme value search and envelope truncation on the peaks and troughs of the pressure wave sequence along the time axis, and output the reflected wave characteristic sequence. S14. Perform spatiotemporal coupling calculations on the temporal intervals of adjacent peaks in the reflected wave characteristic sequence and the adaptively corrected wave velocity field. Perform peak-by-peak spatial tracing and propagation time accumulation extrapolation on the reflected wave characteristic sequence along the time axis, and output the predicted arrival time node of the reflected wave.

[0007] Optionally, S2 specifically includes: S21. Real-time acquisition of pipeline network operation status data, wavelet decomposition to extract multi-scale high-frequency fluctuation sequences based on operation status data, dividing the multi-scale high-frequency fluctuation sequences into sliding windows along the time axis, statistically analyzing the amplitude variance and energy distribution characteristics within each sliding window, and aggregating and outputting dynamic fluctuation characteristics. S22. Extract the instantaneous flow rate and pressure mean sequence based on the operating status data, synchronize the instantaneous flow rate and pressure mean sequence with the dynamic fluctuation characteristics in the time dimension, and fuse and output the operating parameter sequence. S23. Reconstruct the dynamic fluctuation characteristics and the operating parameter sequence in phase space, map the reconstruction result to a two-dimensional phase plane, extract the local tangent slope of the phase trajectory on the two-dimensional phase plane along the time axis, and output the trajectory slope characteristics based on the statistical distribution of the local tangent slope. S24. Calculate the spatial distribution density of slope scatter points in the trajectory slope feature. Based on the spatial distribution density, perform neighborhood connectivity search and adaptive density clustering segmentation on the slope scatter points, extract the boundary envelope of each cluster, map the boundary envelope to the original working condition type, and output the working condition attribute label.

[0008] Optionally, S3 specifically includes: S31. Parse the state transition sequence of the working condition attribute label, identify the trigger state of the control path based on the state transition sequence, lock the current running parameters along the time axis to maintain the state in the absence of triggering, and perform closed-loop update based on the state maintenance duration. S32. When the triggering state is identified, the topology of the pipeline network is obtained and the node connection relationship and pipe segment physical coordinates are extracted. Graph structure encoding is performed on the node connection relationship and pipe segment physical coordinates to discretize and construct the graph topology space. S33. Extract the node monitoring values ​​from the running parameter sequence, align the node monitoring values ​​to the corresponding nodes in the graph topology space, and output the node dimension feature matrix; S34. Extract the static configuration parameters of the pipe segment from the operating parameter sequence, parse the pipe segment topology length and wave propagation benchmark, fuse the pipe segment topology length and wave propagation benchmark along the edge direction of the graph topology space, and adaptively generate the edge weight matrix based on the statistical distribution of the fusion calculation result to complete the weighting of the node and edge dimensions of the graph topology space.

[0009] Optionally, the improved DCRNN model includes a bandgap filtering and resonant coupling layer, a bidirectional temporal cyclic coding layer, a forward temporal decoding layer, and a multidimensional feature map output layer. The bandgap filtering and resonant coupling layer is used to introduce a bandgap filtering and resonant coupling diffusion mechanism based on acoustic impedance constraints. The specific execution process includes: Extract the pipe segment topology length and wave propagation reference from the edge weight matrix, calculate the acoustic admittance characteristics between nodes along the edge direction of the graph topology space, and map the acoustic admittance characteristics into a discrete acoustic impedance distribution matrix. Multidimensional statistical distribution calculations are performed on the dynamic fluctuation characteristics to extract the mean and variance distribution features of the fluctuation frequency at the node level. The mean and variance distribution features of the fluctuation frequency and the acoustic impedance distribution matrix are input into a two-layer mapping network with shared weights to perform nonlinear projection, and the upper and lower boundary thresholds of the local resonant frequency band gap are solved in real time and output as the node state changes. A frequency domain mask matrix is ​​constructed based on the upper and lower boundary thresholds of the local resonant frequency bandgap. The frequency domain mask matrix is ​​multiplied element-wise with the frequency domain transformation result of the dynamic wave characteristics. The frequency domain components falling into the bandgap are suppressed and cleared to zero, and the effective water hammer wave characteristics falling into the passband are extracted. The effective water hammer wave characteristics are used as the driving force and impedance matching product calculation is performed with the acoustic impedance distribution matrix in the feature space to output the resonant coupling space feature tensor. The bidirectional temporal cyclic coding layer is used to input the resonant coupled spatial feature tensor into the gated cyclic unit along the time axis to extract the forward temporal dependency, and input it into the gated cyclic unit along the time axis in the reverse direction to extract the backward temporal dependency. The forward and backward temporal dependencies are concatenated to output the global spatiotemporal hidden state tensor. The forward temporal decoding layer is used to take the global spatiotemporal hidden state tensor as the initial hidden state, and along the inference sequence of future time steps, it concatenates and nonlinearly maps the hidden state decoding tensor output by the previous time step with the input features of the corresponding time step, and outputs the hidden state decoding tensor of the current time step through cyclic rolling update. The multidimensional feature map output layer is used to input the hidden state decoding tensor of future time steps into three parallel fully connected mapping branches. The first branch outputs the primary water hammer evolution state through linear activation processing, the second branch outputs the attribute identifier through Softmax classification processing, and the third branch outputs the amplitude evolution benchmark through regression processing.

[0010] Optionally, S5 specifically includes: S51. Extract the joint probability distribution features of state amplitude and frequency based on the initial water hammer evolution state, extract the statistical proportion features of working condition type based on attribute identification, input the joint probability distribution features and statistical proportion features into a fully connected mapping network to perform nonlinear dimension reduction projection, and output the target phase time difference benchmark. S52. Extract the absolute coordinates of the time axis corresponding to the predicted arrival time node, perform a difference subtraction operation between the absolute coordinates of the time axis and the target phase time difference reference, and output the initial phase offset. S53. Extract the time series fluctuation variance of the initial phase offset sequence, calculate the nonlinear attenuation compensation factor based on the time series fluctuation variance, multiply the nonlinear attenuation compensation factor with the initial phase offset element by element, and output the dynamically corrected offset. S54. Obtain the pipe segment attribute parameters of the pipeline network topology and calculate the node-level limit offset threshold. Extract the node-level offset values ​​of the dynamically corrected offset sequence. Perform a node-by-node comparison between the node-level offset values ​​and the limit offset threshold. Identify abnormal offset nodes that exceed the limit offset threshold. Force reset the offset of abnormal offset nodes to the limit offset threshold. Retain the original values ​​of normal offset nodes that do not exceed the limit. Reconstruct and output the dynamic start-stop delay benchmark based on the reset results and the retained results.

[0011] Optionally, S6 specifically includes: S61. Obtain the system timestamp corresponding to the current system clock, add the system timestamp to the dynamic start / stop delay reference execution time axis, and output the absolute execution time. S62. Extract the real-time speed and frequency distribution characteristics of the inverter's current operating state sequence by anchoring the absolute execution time. Calculate the transient switching overshoot compensation value based on the real-time speed and frequency distribution characteristics. Subtract the transient switching overshoot compensation value from the execution time axis difference of the absolute execution time to output the corrected absolute execution time. S63. Input the correction absolute execution time into the hardware timer interrupt queue, extract the high-level pulse signal triggered when the correction absolute execution time arrives, write the high-level pulse signal as the underlying data stream into the inverter control register, and output the disconnect trigger instruction. S64. Respond to the disconnection trigger command to acquire the real-time operating electrical parameter sequence of the inverter output side, extract the current change timing feature and voltage change timing feature from the real-time operating electrical parameter sequence, perform feature-level tensor splicing on the current change timing feature and voltage change timing feature, and output the primary water hammer wave.

[0012] Optionally, S7 specifically includes: S71. Based on the multi-node pressure sensor array of the pipeline network, the real-time pressure sampling sequence under the action of primary water hammer wave is collected, and the real-time pressure sampling sequence is truncated by a sliding window to output the local pressure fluctuation subsequence. S72. Calculate the local mean and local variance of the local pressure fluctuation subsequence, and extract the transient pressure deviation distribution characteristics based on the local mean and local variance; S73. Extract the wave peak amplitude decay trajectory and wave front propagation delay sequence in the primary water hammer evolution state, extract the background pressure mean and background pressure variance of the real-time pressure sampling sequence during the period without water hammer disturbance, and use the background pressure mean and background pressure variance as adaptively generated steady-state constraint features. S74. Perform feature-level tensor concatenation on the transient pressure deviation distribution characteristics, wave peak amplitude attenuation trajectory and wave front propagation delay sequence to output the water hammer response joint tensor; perform standardized residual calculation on the water hammer response joint tensor and the background pressure mean and variance to output the dynamic deviation index sequence. S75. Extract the time-domain cumulative features of deviation and the local impact distribution features based on the dynamic deviation index sequence, and perform nonlinear state mapping to output the error compensation coefficient.

[0013] Optionally, S8 specifically includes: S81. Extract the proportional and differential components of the error compensation coefficient, perform element-wise multiplication of the proportional component with the dynamic start-stop delay reference, and perform element-wise multiplication of the differential component with the amplitude evolution reference to output the initial corrected delay reference and the initial corrected amplitude reference. S82. Extract the historical delay update sequence corresponding to the initial corrected delay benchmark and the historical amplitude update sequence corresponding to the initial corrected amplitude benchmark, and calculate the delay update variance of the historical delay update sequence and the amplitude update variance of the historical amplitude update sequence. S83. Calculate the multidimensional joint fluctuation distribution characteristics based on the delayed update variance and the amplitude update variance. Perform historical state truncation and local variance gradient calculation on the multidimensional joint fluctuation distribution characteristics through a time-series sliding window to extract the time-series convergence trend characteristics. Calculate the adaptive step size scaling factor based on the time-series convergence trend characteristics. S84. Perform element-wise multiplication of the adaptive step size scaling factor with the initial corrected delay reference and the initial corrected amplitude reference respectively, output the target corrected delay reference and the target corrected amplitude reference, and perform evolution step size mapping calculation along the time step to output the target deduced state transition matrix. S85. Concatenate the target correction delay benchmark and the target correction amplitude benchmark to output the linked correction feature vector. Perform tensor splicing bias on the input layer of the dynamic feature extraction stage based on the linked correction feature vector. At the same time, inject the linked correction feature vector and the target inference state transition matrix as feedforward driving terms into the improved DCRNN model. Perform feature fusion of the resonant coupling layer, hidden state matrix multiplication update of the bidirectional temporal cyclic coding layer, and rolling decoding of the forward temporal decoding layer in sequence to output the feedforward driving features of the next iteration cycle.

[0014] The beneficial effects of this invention are: An improved DCRNN model based on acoustic impedance-constrained bandgap filtering and resonant coupling diffusion mechanism was introduced to establish a physical rule-driven deep spatiotemporal dependency extraction and error feedforward closed-loop verification system. The bandgap filtering and resonant coupling layer calculates the upper and lower boundary thresholds of the local resonant frequency bandgap based on acoustic admittance features to construct a frequency domain mask matrix. Element-wise multiplication suppression and zeroing are performed on the frequency domain transformation results of dynamic wave characteristics to extract effective water hammer wave features and perform impedance matching product calculation to output the resonant coupling spatial feature tensor. The bidirectional temporal cyclic coding layer inputs the resonant coupling spatial feature tensor into a gated cyclic unit along the time axis in both forward and reverse directions to extract forward and backward temporal dependencies and performs a concatenation operation to output the global spatiotemporal hidden state tensor. The adaptive step-size scaling factor is calculated by combining the error compensation coefficients output from the standardized residual solution. Evolutionary step-size mapping is performed along the temporal step to calculate the target inference state transition matrix. This target inference state transition matrix is ​​then injected as a feedforward driving term into the bidirectional temporal cyclic coding layer to perform hidden state matrix multiplication update and rolling decoding. This system achieves multi-dimensional feature enhancement from the physical acoustic filtering space to the deep temporal deduction space through acoustic impedance frequency domain bandgap hard constraint, forward and reverse time-dependent feature extraction, and residual feedforward state transition matrix verification. This ensures that the output feedforward driving features have extremely high capture accuracy for the water hammer wave phase evolution boundary and dynamic start-stop delay offset. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a water pump dynamic start-stop and pressure stabilization control method based on operating condition identification proposed in this invention; Figure 2 This is a flowchart illustrating the working principle of an improved DCRNN model for a water pump dynamic start-stop and pressure stabilization control method based on operating condition identification proposed in this invention. Detailed Implementation

[0016] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figure 1 and Figure 2 A method for dynamic start-up and pressure stabilization control of water pumps based on operating condition identification, specifically including: S1. Dynamically collect pipeline network parameters and pipe segment topology length, correct wave velocity through transient flow solution, extract wave propagation benchmark and steady-state constraint characteristics, and predict the arrival time node of reflected wave along the time axis. S2. Real-time acquisition of pipeline network operation status data, extraction of dynamic fluctuation features and operation parameter sequences through wavelet decomposition and sliding window, extraction of trajectory slope features of two-dimensional phase plane, and mapping output of operating condition attribute labels; S3. Parse the working condition attribute tags to determine the trigger status. If there is no trigger, maintain the status and update it in a closed loop periodically; otherwise, construct the topology space based on the pipeline topology. S4. The dynamic wave characteristics and graph topology space are input into the improved DCRNN model. A bandgap filter and resonant coupling diffusion mechanism based on acoustic impedance constraints are introduced to perform frequency domain filtering suppression and impedance matching coupling. After deep spatiotemporal dependency extraction and mapping, the primary water hammer evolution state, attribute identifier and wave amplitude evolution benchmark are output. S5. Map the initial water hammer evolution state and attribute identifier to the target phase time difference benchmark, combine the predicted arrival time node to perform inverse spatiotemporal calculation, and output the dynamic start-stop delay benchmark. S6. Based on the current system clock superimposed with dynamic start-stop delay reference, the absolute execution time is anchored, and the frequency converter is triggered in real time at the absolute execution time to perform the disconnection action and actively evolve and output the primary water hammer wave. S7. Real-time acquisition of pipeline pressure response under the action of primary water hammer wave, dynamic error calculation combined with primary water hammer evolution state and steady-state constraint characteristics, and output error compensation coefficient. S8. Based on the error compensation coefficient, perform adaptive linkage correction on the dynamic start-stop delay benchmark and amplitude evolution benchmark, and feed forward to the dynamic feature extraction and time series deduction stage to iteratively drive the next dynamic start-stop.

[0018] In this embodiment, S1 specifically includes: S11. Dynamically acquire multi-point pressure fluctuation sequences and pipe segment topology lengths of the pipeline network, calculate the difference between data points at adjacent time steps in the pressure fluctuation sequence and divide it by the sampling time step to obtain the time-pressure change rate, divide the time-pressure change rate by the pipe segment topology length to obtain the transient flow spatiotemporal gradient distribution; construct a weight matrix containing values ​​of 0.6 and 0.4, perform matrix multiplication with the eigenvector formed by concatenating the transient flow spatiotemporal gradient distribution and the pipe segment topology length, input the calculation result into the hyperbolic tangent activation function, and output the adaptively corrected wave velocity field.

[0019] S12. Perform a Fast Fourier Transform on the pressure fluctuation sequence, calculate the square of the amplitude of the frequency domain transformation result to obtain the fluctuation energy distribution, sum the fluctuation energy distribution values ​​in the 0 Hz to 50 Hz frequency band and divide by the total number of frequency points to obtain the fluctuation propagation reference; at the same time, set the time window length to 10 seconds, slide and extract data segments in the pressure fluctuation sequence, calculate the mean of the amplitude envelope range of the current time window and the previous time window data segment with the maximum value minus the minimum value, and output the steady-state constraint characteristics.

[0020] S13. Multiply the wave propagation benchmark by 1.2 as the first judgment threshold, multiply the steady-state constraint feature by 0.8 as the second judgment threshold, traverse the pressure wave sequence along the time axis, mark the data points greater than the first judgment threshold as wave peaks, mark the data points less than the second judgment threshold as wave troughs, cut off abnormal wave points where the time interval between two consecutive wave peaks is less than 0.1 seconds, combine the retained wave peaks and troughs in time order, and output the reflected wave feature sequence.

[0021] S14. Extract the spatial coordinates of the pipeline network corresponding to the last wave peak in the reflected wave feature sequence as the source point, calculate the remaining straight-line distance from the source point to the preset monitoring target node of the pipeline network, divide the remaining straight-line distance by the wave velocity value of the node corresponding to the source point in the adaptive correction wave velocity field to obtain the remaining propagation time; extract the absolute coordinates of the timestamp corresponding to the last wave peak, add the absolute coordinates of the timestamp to the remaining propagation time to perform propagation time accumulation and deduction, and output the predicted arrival time node of the reflected wave.

[0022] In this embodiment, S2 specifically includes: S21. Real-time acquisition of pipeline network operation status data is input into the db4 wavelet basis function to perform 3-level discrete wavelet decomposition, extracting the high-frequency detail coefficients from the 1st to the 3rd levels as a multi-scale high-frequency fluctuation sequence; a sliding window with a time length of 50 sampling points is set and truncates with a step size of 1, the sum of squares of the data points within the window is calculated as the energy distribution feature, and the sum of the squares of the data points within the window minus the mean is calculated and divided by the total number as the amplitude variance, and the two are spliced ​​together in time order to output the dynamic fluctuation feature.

[0023] S22. Extract the instantaneous flow rate and pressure mean value sequence based on the operating status data. Multiply the instantaneous flow rate value by 0.6 and add the pressure mean value by 0.4 to obtain the fusion weight sequence. Extract the start timestamp of each window in the dynamic fluctuation feature. Downsample the fusion weight sequence according to the timestamp to generate the mean value and generate the fusion weight feature vector. Concatenate it with the dynamic fluctuation feature in the feature dimension to output the operating parameter sequence.

[0024] S23. Set the time delay to 3 and the embedding dimension to 2. Use the current time step value of the running parameter sequence as the first dimension and the value of moving 3 time steps as the second dimension to construct two-dimensional phase points and map them to the two-dimensional phase plane. Calculate the difference between the ordinates of adjacent two-dimensional phase points and divide it by the difference between the abscissas to obtain the local tangent slope. Use the first dimension value of the phase point as the abscissa and the local tangent slope as the ordinate to construct two-dimensional scattered points and output the trajectory slope feature.

[0025] S24. Set the neighborhood search radius to 0.5, traverse the trajectory slope features to calculate the Euclidean distance between slope scatter points, mark the scatter points with a distance less than 0.5 as connected and count the number as spatial distribution density; divide the connected regions with a spatial distribution density greater than 5 into clusters, extract the maximum and minimum values ​​of the vertical coordinate slope values ​​within the cluster to construct the slope threshold boundary interval, calculate the average slope of the connected scatter points within the interval, compare it with the preset mapping dictionary containing three categories: steady state, pump start and pump stop, and output the interval category into which the average value falls as the label for the working condition attribute label.

[0026] In this embodiment, S3 specifically includes: S31. Read the working condition attribute label sequence, calculate the difference between the label values ​​of adjacent time steps to construct a state transition sequence, and set the difference greater than 0 as the triggered state and equal to 0 as the non-triggered state; in the non-triggered state, lock the running parameter sequence of the current time step as the first row, and copy it down 10 rows to construct an 11-row state maintenance matrix; calculate the variance of the data in each column of the state maintenance matrix, set the variance threshold to 0.05, when the variance is less than 0.05, directly output the state maintenance matrix, when the variance is greater than or equal to 0.05, remove the first row and append the running parameter sequence of the latest time step until the variance is less than 0.05, complete the periodic closed-loop update and output the state maintenance matrix.

[0027] S32. When the trigger state is identified, extract the total number of nodes in the pipeline network to construct an all-zero adjacency matrix of the same dimension, traverse the node pairs with physical connections, set the corresponding element in the adjacency matrix to the value 1, and output the node connection relationship; at the same time, extract the spatial horizontal and vertical coordinates of each node and concatenate them row by row to output the physical coordinates of the pipe segment, and concatenate the node connection relationship and the physical coordinates of the pipe segment in the feature dimension to output the graph topology space.

[0028] S33. Extract the node monitoring values ​​corresponding to the trigger state timestamps in the running parameter sequence. Input the value of each data point in the node monitoring value into the S-type activation function to perform nonlinear mapping calculation and output the mapping value. The calculation process of the S-type activation function is to calculate the value 1 and add the sum of the results calculated by raising the power of the natural constant e to the opposite number of the data point value. Divide the value 1 by the sum to obtain the mapping value. Align and fill all the mapping values ​​according to the node arrangement order in the graph topology space to construct a node dimension feature matrix with the number of rows equal to the total number of nodes and the number of columns equal to 1.

[0029] S34. Extract the static configuration parameters of the pipe segments from the operating parameter sequence. According to the position of the node pair corresponding to the value 1 in the node connection relationship in the graph topology space, analyze the pipe segment topology length and wave propagation reference of the corresponding pipe segment. Divide the pipe segment topology length by the wave propagation reference to obtain the single-sided time impedance. Construct a zero-edge weight matrix of the same dimension and fill the position corresponding to the value 1 with the single-sided time impedance. Calculate the mean of all single-sided time impedances. Take the absolute value of the difference between each single-sided time impedance and the mean and input it into the hyperbolic tangent activation function to calculate the edge weight correction value. The calculation process of the hyperbolic tangent activation function is to take the difference between twice the single-sided time impedance and the power of the opposite of twice the single-sided time impedance, divide it by the sum of the corresponding powers, and use the edge weight correction value to replace the non-zero elements in the edge weight matrix. Output the graph topology space with the node and edge dimensions weighted.

[0030] In this embodiment, the improved DCRNN model includes a bandgap filtering and resonant coupling layer, a bidirectional temporal cyclic coding layer, a forward temporal decoding layer, and a multidimensional feature mapping output layer: The bandgap filtering and resonant coupling layer is used to introduce a bandgap filtering and resonant coupling diffusion mechanism based on acoustic impedance constraints. The specific execution process includes: Extract the pipe segment topology length and wave propagation reference from the edge weight matrix. Divide the pipe segment topology length by the wave propagation reference to obtain the basic admittance of the connection. Add the value 0.1 to the basic admittance of the connection to obtain the acoustic admittance feature between nodes. Fill the acoustic admittance feature into a zero matrix of the same dimension according to the connection position in the graph topology space, and output the acoustic impedance distribution matrix.

[0031] The mean and variance of the dynamic fluctuation features are calculated column-wise to obtain the mean and variance distribution features of the fluctuation frequency. These two are concatenated into the first input vector, and the acoustic impedance distribution matrix is ​​flattened into the second input vector. The two are concatenated along the feature dimension to obtain the fused input vector. The fused input vector is input into a two-layer mapping network with 5 nodes and 8 hidden layer dimensions. The product of the fused input vector and the first layer weight matrix is ​​calculated, plus the first bias vector. The first layer hidden state is calculated by inputting the modified linear unit activation function. The calculation process of the modified linear unit activation function is to determine whether the input value is greater than 0. If it is greater than 0, the input value is output; if it is less than or equal to 0, the value 0 is output. The product of the first layer hidden state and the second layer weight matrix is ​​calculated, plus the second bias vector. The output mapping value is calculated by inputting the sigmoid activation function. The calculation process of the sigmoid activation function is to calculate the sum of the result of the calculation with the natural constant e as the base and the opposite number of the input value as the exponent. The value 1 is divided by the sum to obtain the mapping value. The mapping value is added to the value of 10 to output the upper boundary threshold of the local resonant frequency bandgap, and added to the value of 5 to output the lower boundary threshold of the local resonant frequency bandgap.

[0032] A zero-matrix of the same dimension as the dynamic wave feature is constructed as the frequency domain mask matrix. The frequency values ​​of each frequency component in the frequency domain transformation result of the dynamic wave feature are traversed. When the frequency value is greater than the lower boundary threshold of the local resonant frequency bandgap and less than the upper boundary threshold of the local resonant frequency bandgap, the element at the corresponding position of the frequency domain mask matrix is ​​set to 0; otherwise, it is set to 1. The frequency domain mask matrix and the frequency domain transformation result of the dynamic wave feature are multiplied element by element. The inverse frequency domain transformation is performed on the multiplication result to output the effective water hammer wave feature. The effective water hammer wave feature is used as the driving vector. The product of the driving vector and the corresponding node position element in the acoustic impedance distribution matrix is ​​calculated. All product results are combined to output the resonant coupling spatial feature tensor.

[0033] The bidirectional temporal recurrent coding layer is used to sequentially input the resonant coupled spatial feature tensor into the forward gated recurrent unit step by step from the first step to the last step. At each time step, the concatenated vector of the input data and the hidden state of the previous time step is multiplied by the reset gate weight matrix and the reset gate bias. The result is then input into the sigmoid activation function to calculate the reset gate state. The calculation process of the sigmoid activation function is to calculate the sum of the calculated value 1 and the result of a power raised to the power of the natural constant e with the opposite of the input value as the exponent. The value 1 is then divided by this sum to obtain the reset gate state. The reset gate state is then multiplied by the previous time step. The result of the hidden state is concatenated with the current input data and multiplied by the update gate weight matrix plus the update gate bias. The result is then input into the sigmoid activation function to calculate the update gate state. The result of the calculation is 1 minus the update gate state multiplied by the hidden state of the previous time step plus the update gate state multiplied by the candidate hidden state. The result of the calculation is the forward temporal dependency of the current time step. The resonant coupled spatial feature tensor is input into the reverse gated recurrent unit in sequence from the last step to the first step in time step. The above calculation process is repeated to output the backward temporal dependency. The forward temporal dependency and the backward temporal dependency are concatenated in the feature dimension to output the global spatiotemporal hidden state tensor.

[0034] The forward temporal decoding layer sets the inference sequence length for future time steps to 10, using the last hidden state in the global spatiotemporal hidden state tensor as the initial decoding hidden state, and setting the initial decoding input to the value 0. At each time step of the future inference sequence, the hidden state decoding tensor output from the previous time step is concatenated with the input features of the corresponding time step along the feature dimension. The concatenation result is multiplied by the decoding layer weight matrix and the decoding layer bias is added. The hyperbolic tangent activation function is then used to calculate the hidden state decoding tensor of the current time step. The calculation process of the hyperbolic tangent activation function is to take the difference between twice the input value of the natural constant e and the power of the opposite of twice the input value, divide by the sum of the corresponding powers, and use the hidden state decoding tensor of the current time step as the input of the next time step to perform a loop rolling update, outputting the hidden state decoding tensors of all future time steps.

[0035] The multidimensional feature map output layer is used to construct three parallel fully connected mapping branches. The output dimension of each fully connected mapping branch is set to 1. The first branch calculates the hidden state decoding tensor of the future time step, multiplies it by the first branch weight matrix and adds the first branch bias, and directly outputs the value as the primary water hammer evolution state through a linear activation function. The second branch calculates the hidden state decoding tensor of the future time step, multiplies it by the second branch weight matrix and adds the second branch bias, and inputs the result into the Softmax classification function to process and output the attribute label. The calculation process of the Softmax classification function is to take the result of the power calculation with the natural constant e as the base and the input value as the exponent, and divide the power calculation result by the sum of the power calculation results of all categories to obtain the attribute label. The third branch calculates the hidden state decoding tensor of the future time step, multiplies it by the third branch weight matrix and adds the third branch bias, and directly outputs the value as the amplitude evolution benchmark through a linear activation function.

[0036] The improved DCRNN model proposed in this step is similar to the traditional DCRNN in terms of its temporal inference mechanism. Both are based on the theory of graph structure deep spatiotemporal feature extraction and state space mapping. That is, by projecting the dynamic fluctuation features of pipeline nodes onto a high-dimensional graph topology space for spatial edge feature aggregation, the weight update mechanism of gated recurrent units is used to capture the dynamic evolution law of temporal state, and both use a fully connected mapping layer to map the final hidden state decoding tensor into a scalarized water hammer evolution benchmark and working condition attribute label.

[0037] The difference lies in that this invention breaks through the limitation of traditional DCRNN's purely data-driven graph convolutional aggregation ignoring the acoustic physical boundaries of the fluid network. It adds a bandgap filtering and resonant coupling layer before the bottom-level graph message passing, introducing a bandgap filtering and resonant coupling diffusion mechanism based on acoustic impedance constraints. This mechanism abandons the traditional approach of directly performing graph convolutional diffusion on the original frequency domain fluctuation features. Instead, it calculates the acoustic admittance features along the graph edge direction, mapping them to an acoustic impedance distribution matrix. It then combines the node frequency distribution to solve the upper and lower boundary thresholds of the local resonant frequency bandgap in real time to construct a frequency domain mask matrix. After suppressing and zeroing the frequency components falling within the bandgap, it performs impedance matching multiplication with the acoustic impedance distribution matrix to calculate and output the resonant coupling spatial feature tensor, rather than directly inputting it into the bidirectional temporal cyclic coding layer.

[0038] The beneficial effects of the improvements are that this invention, through acoustic impedance bandgap hard constraint and impedance matching diffusion, forcibly embeds the acoustic passband filtering and physical impedance coupling mechanism into the forward propagation of the graph network. This breaks the limitation of traditional DCRNN in the easy confusion of invalid oscillation modes and real water hammer wave characteristics under complex multi-source reflection wave superposition, which leads to phase inference divergence. It realizes the accurate conversion from undifferentiated graph message diffusion to acoustic and physical strong constraint space aggregation. This design significantly enhances the model's ability to isolate and filter environmental noise and parasitic frequency domain interference. It can extract pure tensors in the feature space that strictly follows the real physical acoustic characteristics of the pipeline network. Combined with bidirectional time-series dependency extraction and rolling decoding, it effectively improves the absolute accuracy and robustness of the water hammer wave phase evolution boundary and dynamic start-stop delay benchmark inference.

[0039] In this embodiment, S5 specifically includes: S51. Based on the primary water hammer evolution state and attribute identifier, calculate the covariance value and the proportion of each category respectively, concatenate the two into a dimension-reduced input vector, input it into a fully connected mapping network to calculate its product with the weight matrix plus the bias vector, input the result into a sigmoid activation function to calculate the output mapping value. The calculation process of the sigmoid activation function is to calculate the sum of the calculated value 1 plus the sum of the results calculated with the natural constant e as the base and the opposite number of the input value as the exponent, divide the value 1 by the sum to obtain the mapping value, and multiply the mapping value by the value 5 to output the target phase time difference benchmark.

[0040] S52. Extract the absolute coordinate values ​​of the time axis of the predicted arrival time node, calculate the difference between the absolute coordinate values ​​of the time axis and the target phase time difference reference, and output the initial phase offset.

[0041] S53. Calculate the variance of the initial phase offset sequence to obtain the time series fluctuation variance. Calculate the value 1 divided by the sum of the time series fluctuation variance and the value 0.01 to obtain the nonlinear attenuation compensation factor. Multiply each value in the initial phase offset sequence by the nonlinear attenuation compensation factor to output the dynamically corrected offset sequence.

[0042] S54. Read the pipe wall thickness value of the pipeline topology and divide it by 1000 to obtain the node-level limit offset threshold. Traverse the dynamic correction offset sequence to extract the node-level offset value. When the difference between the absolute value of the node-level offset value and the limit offset threshold is greater than 0, replace the node-level offset value with the limit offset threshold. Otherwise, retain the original value. Reassemble the replaced and retained values ​​according to the node order and output the dynamic start-stop delay benchmark.

[0043] The dynamic start-stop delay benchmark calculation process proposed in this step is similar to the traditional water hammer passive delay control mechanism in that it is based on the theory of wave state feature extraction and time axis mapping. That is, by projecting the evolution characteristics of water hammer waves and the timing characteristics of reflected waves onto the time dimension to calculate the phase difference, the difference operation mechanism is used to capture the delay law of control action, and numerical mapping is used to convert the characteristic deviation into a scalar execution delay benchmark.

[0044] The difference lies in that this invention breaks through the limitations of traditional fixed empirical delay or simple linear interpolation that ignores the physical limits of the pipeline topology and the attenuation of sequence fluctuations. It adds nonlinear attenuation compensation and limit offset truncation steps, and projects the primary water hammer evolution state and operating conditions into a target phase time difference benchmark by dimensionality reduction. It calculates the nonlinear attenuation compensation factor based on the variance of time series fluctuations to perform dynamic offset correction, and replaces the traditional unconstrained output with a node-level limit offset threshold. It performs node-by-node comparison between the dynamic correction offset and the pipe segment attribute parameters, and forces the reset and truncation of abnormal offset nodes that exceed the threshold. Finally, it reconstructs and outputs a dynamic start-stop delay benchmark, rather than a single linear difference calculation.

[0045] The beneficial effects of the improvements are that, through attenuation compensation and limit truncation, the physical displacement constraints of pipeline nodes are rigidly embedded into the delay benchmark calculation process, breaking the limitations of traditional methods that are prone to waveform misalignment or mechanical overruns due to unreasonable delay settings under complex working conditions. This achieves a precise conversion from fixed empirical delay to adaptive delay with strong topological physical constraints. This design significantly enhances the dynamic tracking and defense capabilities against transient flow timing fluctuations and variances. It can accurately truncate abnormal offset risks in the node-level safety boundary space. Combined with nonlinear attenuation fusion correction, it effectively improves the accuracy of inverter cut-in timing and the absolute safety of pipeline equipment operation scheduling.

[0046] In this embodiment, S6 specifically includes: S61. Read the system timestamp value corresponding to the current system clock, extract the delay value of each node in the dynamic start-stop delay benchmark, calculate the sum of the system timestamp value and the delay values ​​of all nodes in the dynamic start-stop delay benchmark, and output the sum as the absolute execution time.

[0047] S62. Extract the inverter's current operating state sequence within the 10 sampling periods before the absolute execution time. Calculate the difference between the values ​​of two adjacent data points in the state sequence and take the absolute value. Arrange all absolute values ​​in chronological order to construct a real-time speed-frequency distribution feature. Calculate the sum of all values ​​in the real-time speed-frequency distribution feature and divide by the total number of values ​​to obtain the average frequency fluctuation. Multiply the average frequency fluctuation by 0.5 to obtain the transient switching overshoot compensation value. Calculate the difference between the time axis value of the absolute execution time and the transient switching overshoot compensation value, and use the difference as the output to correct the absolute execution time.

[0048] S63. Input the time axis value corresponding to the corrected absolute execution time into the hardware timer interrupt queue. When the system clock value reaches the time axis value of the corrected absolute execution time, trigger the interrupt response and extract the high-level pulse signal generated by the interrupt response. Write the value 1 of the high-level pulse signal into the storage bit with the offset address of value 0 in the inverter control register and output the cut-in trigger instruction.

[0049] S64. In response to the disconnection trigger command, the real-time operating electrical parameter sequence containing current and voltage data on the output side of the frequency converter is acquired. The difference between two adjacent data points in the current data is calculated to obtain the current difference sequence, and the difference between two adjacent data points in the voltage data is calculated to obtain the voltage difference sequence. The current difference sequence is used as the first row and the voltage difference sequence is used as the second row to construct a time-series mutation matrix. The time-series mutation matrix is ​​directly used as the feature tensor to output the primary water hammer wave.

[0050] In this embodiment, S7 specifically includes: S71. Read the value of each data point in the real-time pressure sampling sequence under the action of primary water hammer wave collected by the multi-node pressure sensor array of the pipeline network. Set the length of the sliding window to 20. Slide the sliding window from the first data point to the end on the real-time pressure sampling sequence with a step size of 1. Extract 20 data points in each sliding window. Arrange all the extracted results in chronological order and output the local pressure fluctuation subsequence.

[0051] S72. For each subsequence in the local pressure fluctuation subsequence, calculate the sum of all data point values ​​in the subsequence and divide it by 20 to obtain the local mean. Calculate the sum of the squares of each data point value minus the local mean and divide it by 20 to obtain the local variance. Calculate the difference between each data point value and the local mean, divide the difference by the square root of the sum of the local variance and the value 0.01, arrange all the calculation results in the original time order, and output the transient pressure deviation distribution characteristics.

[0052] S73. Read the values ​​of each state data point in the initial water hammer evolution state, extract the maximum value among all state data points and connect them in time order to obtain the wave peak amplitude attenuation trajectory, extract the time point coordinates corresponding to all maximum values ​​and subtract them to obtain the wavefront propagation delay sequence; read the values ​​of the first 100 data points in the real-time pressure sampling sequence, calculate the sum of the values ​​of the first 100 data points and divide it by 100 to obtain the background pressure mean, calculate the sum of the squares of the first 100 data points minus the background pressure mean and divide it by 100 to obtain the background pressure variance, and combine the background pressure mean and background pressure variance to output the steady-state constraint characteristics.

[0053] S74. The transient pressure deviation distribution characteristics are used as the first row, the peak amplitude attenuation trajectory is used as the second row, and the wavefront propagation delay sequence is used as the third row. The feature-level tensors are spliced ​​together to output the water hammer response joint tensor. Each value in the first row of the water hammer response joint tensor is extracted, the difference between the value and the mean background pressure is calculated, and the difference is divided by the square root of the sum of the variance of the background pressure and the value 0.01. All division results are arranged in their original positions to output the dynamic deviation index sequence.

[0054] S75. Calculate the cumulative sum of all values ​​in the dynamic deviation index sequence to obtain the deviation time-domain cumulative feature. Extract values ​​greater than 1.5 from the dynamic deviation index sequence and count their number, dividing by the total sequence length to obtain the local impact distribution feature. Concatenate the deviation time-domain cumulative feature and the local impact distribution feature along the feature dimension to obtain the state mapping input vector. Construct a fully connected mapping network with an output dimension of 1. Calculate the product of the state mapping input vector and the network weight matrix, plus the bias vector. Input the result into the hyperbolic tangent activation function to calculate the current state mapping value. The calculation process of the hyperbolic tangent activation function is to take the difference between twice the input value and the power of the opposite of twice the input value, divide by the sum of the corresponding powers, and multiply the current state mapping value by the value 0.1 to output the error compensation coefficient.

[0055] In this embodiment, S8 specifically includes: S81. Read the first two values ​​in the error compensation coefficient as the proportional component and the differential component, respectively. Extract the delay value of each node in the dynamic start-stop delay reference. Calculate the proportional component and multiply it by the delay value of each node to obtain each element in the initial corrected delay reference. Extract each amplitude value in the amplitude evolution reference. Calculate the differential component and multiply it by the amplitude value to obtain each element in the initial corrected amplitude reference. Arrange all elements in their original order and output the initial corrected delay reference and the initial corrected amplitude reference.

[0056] S82. Read the data of the first 10 historical periods corresponding to the initial corrected delay benchmark to construct a historical delay update sequence. Calculate the sum of all values ​​in the historical delay update sequence and divide by 10 to obtain the delay mean. Calculate the sum of the squares of each value minus the delay mean and divide by 10 to obtain the delay update variance. Read the data of the first 10 historical periods corresponding to the initial corrected amplitude benchmark to construct a historical amplitude update sequence. Calculate the amplitude update variance using the same summation and mean square difference steps.

[0057] S83. Concatenate the delayed update variance and the amplitude update variance along the feature dimension to obtain the multidimensional joint fluctuation distribution feature. Set the length of the time series sliding window to a value of 5. Slide the multidimensional joint fluctuation distribution feature with a step size of 1 to extract a subsequence of length 5. Calculate the difference between two adjacent values ​​in each subsequence as the gradient value. Calculate the sum of the squares of all gradient values ​​and divide by the total number of gradient values ​​to obtain the local variance gradient. Calculate the negative of the local variance gradient and add a value of 1 to obtain the time series convergence trend feature. Calculate the value of 1 and divide by the sum of the time series convergence trend feature and a value of 0.05 to obtain the adaptive step size scaling factor.

[0058] S84. Extract each element from the initial corrected delay reference and multiply it by the adaptive step size scaling factor. Arrange the multiplication results in the original order and output the target corrected delay reference. Extract each element from the initial corrected amplitude reference and multiply it by the adaptive step size scaling factor. Arrange the multiplication results in the original order and output the target corrected amplitude reference. Construct a diagonal matrix with 50 rows and 50 columns. Set the elements on the main diagonal to be the adaptive step size scaling factor and the elements on the off-diagonal to be 0. Output this matrix as the target derivation state transition matrix.

[0059] S85. Using the target correction delay benchmark as the first row and the target correction amplitude benchmark as the second row, concatenate them to output a linked correction feature vector. Extract each tensor value from the input layer of the dynamic feature extraction stage. Calculate the sum of each tensor value and the corresponding value in the linked correction feature vector. Replace the original tensor values ​​with all the summed results to complete the tensor concatenation bias. Input the linked correction feature vector as the first input and the target derivation state transition matrix as the second input into the improved DCRNN model. Calculate the product of the transpose matrices of the first and second inputs in the resonant coupling layer and input it into the corrected linear unit activation function. The corrected linear unit activation function determines whether the input value is greater than 0. If it is greater than 0, output the input value; if it is less than or equal to 0, output the value 0, thus obtaining the fused features. The matrix is ​​calculated as follows: In the bidirectional temporal recurrent coding layer, the product of the fused feature matrix and the hidden state weight matrix is ​​added to the product of the hidden state matrix and the recurrent weight matrix of the previous time step. The result is input into the hyperbolic tangent activation function to calculate the hidden state matrix of the current time step. The calculation process of the hyperbolic tangent activation function is to take the difference between twice the input value and twice the opposite power of the input value, and divide by the sum of the corresponding powers. In the forward temporal decoding layer, the product of the hidden state matrix and the decoding weight matrix of the current time step is added to the decoding bias vector. The result is input into the sigmoid activation function to calculate the feedforward driving feature of the next iteration cycle. The calculation process of the sigmoid activation function is to calculate the value 1 and add the sum of the results calculated with the natural constant e as the base and the opposite power of the input value as the exponent. The value 1 is divided by the sum to obtain the mapped value output.

[0060] The adaptive linkage correction and feedforward drive process proposed in this step is similar to the traditional water hammer control error feedback adjustment mechanism in that it is based on the error compensation and state space update theory. That is, the error allocation calculation is performed by projecting the actual pressure response deviation characteristics onto the control parameter space, and the dynamic correction law of the control reference is captured by the proportional and derivative weight adjustment mechanism. Both adopt a closed loop to map the error calculation results into the updated control output.

[0061] The difference lies in that this invention breaks through the limitation of traditional error feedback, which only corrects the output of the peripheral control while ignoring the cognitive bias within the deep time series model. It adds steps for extracting time series convergence trend features and mapping the state transition matrix. Based on the historical update variance, it calculates the adaptive step size scaling factor and performs nonlinear step size mapping on the correction benchmark. Instead of the traditional single-point error superposition, it uses the target inferred state transition matrix and the linked correction feature vector as feedforward driving terms and directly injects them into the resonant coupling layer, bidirectional time series recurrent encoding layer and forward time series decoding layer of the improved DCRNN model to perform hidden state matrix multiplication and update, outputting the feedforward driving features for the next iteration cycle, rather than directly compensating for the single underlying control quantity.

[0062] The beneficial effects of the improvements are that, through state transition matrix mapping and deep network feedforward injection, the historical convergence trend constraints are forcibly implanted into the forward propagation of the deep graph network. This breaks the limitations of traditional methods, which are prone to prediction model mismatch and phase inference divergence due to error accumulation in the dynamic evolution of complex pipeline networks. It achieves accurate conversion from feedback correction of peripheral actuators to feedforward correction of deep spatiotemporal inference models. This design significantly enhances the defense capability against sudden changes in variance of temporal fluctuations and model drift. It can accurately calibrate the evolution trajectory of water hammer waves in high-dimensional hidden state space. Combined with adaptive step size scaling linkage, it effectively improves the global optimization convergence of dynamic start-stop delay benchmark and the absolute stability of iterative defense under complex pipeline network conditions.

[0063] Example 1: To verify the feasibility of this invention in water hammer prevention and pressure control of fluid pipeline networks, the method of this invention was applied to the dynamic scheduling and water hammer suppression system of the main water supply network system of a large municipal water group (hereinafter referred to as "Company W"). In traditional water supply network operation systems, control strategies based on fixed empirical delay start-stop or simple PID constant pressure compensation are usually adopted. These methods not only make it difficult to accurately calculate the phase evolution trajectory of water hammer waves under complex operating conditions and the superposition of multiple reflected waves, but also cannot accurately match the physical acoustic impedance characteristics of the pipeline network, which can easily lead to misjudgment of the inverter's disconnection timing or severe water hammer overshoot. To solve the above problems, Company W decided to adopt the pump dynamic start-stop and pressure stabilization control method based on operating condition identification proposed in this invention.

[0064] During implementation, Company W first utilized high-frequency pressure sensor arrays deployed at key nodes of the pipeline network and the inlet and outlet of pumping stations to acquire multi-dimensional pressure fluctuation sequences. Through transient flow spatiotemporal gradient calculation and cross-mapping of pipe segment topology lengths, an aligned input data stream containing adaptively corrected wave velocity fields and steady-state constraint features was constructed. Simultaneously, Company W's hydraulic experts performed precise two-dimensional phase plane trajectory slope extraction and adaptive density clustering segmentation labeling on the collected multi-source operating condition data, serving as the benchmark for model training and operating condition trigger state discrimination.

[0065] Company W constructs a graph topology space, aligns node monitoring values ​​to the corresponding node's output node-dimensional feature matrix, and combines the pipe segment topology length and wave propagation benchmark along the edge direction to generate an edge weight matrix. Next, the dynamic wave features and the graph topology space are input into an improved DCRNN model. Bandgap filtering and a resonant coupling layer are used to calculate the acoustic admittance features between nodes and map them to an acoustic impedance distribution matrix. The upper and lower boundary thresholds of the local resonant frequency bandgap are solved in real time to construct a frequency domain mask matrix. Subsequently, the frequency domain transformation results of the dynamic wave features are subjected to element-wise multiplication to suppress and zero out the in-bandgap components, extracting effective water hammer wave features and performing impedance matching multiplication to calculate the output resonant coupling space feature tensor.

[0066] In the core calculation and control phase, this invention extracts the global spatiotemporal hidden state tensor through a bidirectional time-series cyclic encoding layer and a forward time-series decoding layer, and outputs the primary water hammer evolution state, attribute identifier, and amplitude evolution benchmark through rolling decoding. Combining the predicted arrival time of the reflected wave, inverse spatiotemporal calculation is performed to output a dynamic start-stop delay benchmark, obtaining the current system clock superimposed with the dynamic start-stop delay benchmark to anchor the absolute execution time. Finally, at the absolute execution time, the system triggers a high-level pulse signal to write to the inverter control register to execute a disconnect / reconnection action, actively evolving and outputting the primary water hammer wave. Combined with the error compensation coefficient output from the standardized residual calculation, a feedforward closed-loop update is performed, achieving a closed-loop transition from condition awareness to proactive defense.

[0067] During implementation, the technical team at Company W discovered that, compared to traditional methods of manually setting delays and constant pressure, the method of this invention significantly improves the accuracy of transient flow phase offsetting and the real-time stability of pressure in pipeline networks. Traditional methods cannot quantify the superimposed phase shift of multi-source reflected waves and have poor filtering effects on frequency domain oscillation modes under complex topologies. In contrast, the method of this invention effectively achieves active and precise evolution of water hammer waves and dynamic and stable pressure control through acoustic impedance bandgap filtering, graph topology spatiotemporal extrapolation, and adaptive residual feedforward correction.

[0068] To further verify the actual performance of the method of the present invention, Company W conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1. Performance Comparison of Water Hammer Defense and Start-up / Stop Control Methods for Company W Water Supply Network

[0069] As shown in Table 1, the performance of the water hammer prevention and control system for water supply networks has been comprehensively improved after applying the method of this invention. The accuracy of operating condition identification has increased from 81.2% of the traditional method to 96.5%, and the start-stop delay calculation error has decreased from 85.0 milliseconds to 6.5 milliseconds, significantly improving the accuracy of phase offset and providing a reliable basis for subsequent active evolution. The maximum water hammer pressure peak deviation has decreased from 18.5% to 2.1%, effectively avoiding the risk of extreme damage to the pipeline network. The pressure stabilization convergence time has been significantly shortened from 45 seconds to 8.5 seconds, significantly enhancing the system's timeliness. In addition, the invalid oscillation mode filtering rate has increased from 65.0% to 98.2%, the number of frequency converter overshoot triggers has decreased from 12.0 times / 100 times to 0.5 times / 100 times, and the pipeline burst accident rate has dropped to 0, significantly reducing equipment wear and maintenance costs. The satisfaction rate of operation and maintenance scheduling has also significantly improved, from 84.0% to 98.0%.

[0070] Through the method of this invention, Company W has successfully achieved active and precise evolution and dynamic and stable pressure control of multi-source water hammer waves in complex pipe networks. This effectively eliminates phase misalignment and overshoot damage caused by the superposition of reflected waves, ensures the safe operation of the main water supply network, significantly improves the intelligence and digitalization level of water dispatching, significantly reduces the workload of dispatching personnel, enhances the stability and robustness of the pump station control system, and provides strong technical support for the construction of modern smart water affairs.

[0071] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic start-stop and pressure stabilization control of a water pump based on operating condition identification, characterized in that, Includes the following steps: S1. Dynamically collect pipeline network parameters and pipe segment topology length, correct wave velocity through transient flow solution, extract wave propagation benchmark and steady-state constraint characteristics, and predict the arrival time node of reflected wave along the time axis. S2. Real-time acquisition of pipeline network operation status data, extraction of dynamic fluctuation features and operation parameter sequences through wavelet decomposition and sliding window, extraction of trajectory slope features of two-dimensional phase plane, and mapping output of operating condition attribute labels; S3. Parse the working condition attribute tags to determine the trigger status. If there is no trigger, maintain the status and update it in a closed loop periodically; otherwise, construct the topology space based on the pipeline topology. S4. The dynamic wave characteristics and graph topology space are input into the improved DCRNN model. A bandgap filter and resonant coupling diffusion mechanism based on acoustic impedance constraints are introduced to perform frequency domain filtering suppression and impedance matching coupling. After deep spatiotemporal dependency extraction and mapping, the primary water hammer evolution state, attribute identifier and wave amplitude evolution benchmark are output. S5. Map the initial water hammer evolution state and attribute identifier to the target phase time difference benchmark, combine the predicted arrival time node to perform inverse spatiotemporal calculation, and output the dynamic start-stop delay benchmark. S6. Based on the current system clock superimposed with dynamic start-stop delay reference, the absolute execution time is anchored, and the frequency converter is triggered in real time at the absolute execution time to perform the disconnection action and actively evolve and output the primary water hammer wave. S7. Real-time acquisition of pipeline pressure response under the action of primary water hammer wave, dynamic error calculation combined with primary water hammer evolution state and steady-state constraint characteristics, and output error compensation coefficient. S8. Based on the error compensation coefficient, perform adaptive linkage correction on the dynamic start-stop delay benchmark and amplitude evolution benchmark, and feed forward to the dynamic feature extraction and time series deduction stage to iteratively drive the next dynamic start-stop.

2. The method for dynamic start-stop and pressure stabilization control of a water pump based on operating condition identification according to claim 1, characterized in that, S1 specifically includes: S11. Dynamically collect multi-point pressure fluctuation sequences and pipe segment topology lengths of the pipeline network, statistically analyze the spatiotemporal gradient distribution of transient flow based on the pressure fluctuation sequences, perform multi-dimensional feature cross-mapping calculations on the spatiotemporal gradient distribution and pipe segment topology lengths, and output an adaptively corrected wave velocity field. S12. Based on the adaptive correction of the wave velocity field, perform frequency domain transformation, extract the statistical distribution characteristics of wave energy in the frequency band as the wave propagation benchmark, and simultaneously extract the mean value of the amplitude envelope of the pressure wave sequence in the time window as the steady-state constraint feature. S13. Construct a dual judgment threshold based on the wave propagation benchmark and steady-state constraint characteristics, perform extreme value search and envelope truncation on the peaks and troughs of the pressure wave sequence along the time axis, and output the reflected wave characteristic sequence. S14. Perform spatiotemporal coupling calculations on the temporal intervals of adjacent peaks in the reflected wave characteristic sequence and the adaptively corrected wave velocity field. Perform peak-by-peak spatial tracing and propagation time accumulation extrapolation on the reflected wave characteristic sequence along the time axis, and output the predicted arrival time node of the reflected wave.

3. The method for dynamic start-stop and pressure stabilization control of a water pump based on operating condition identification according to claim 1, characterized in that, S2 specifically includes: S21. Real-time acquisition of pipeline network operation status data, wavelet decomposition to extract multi-scale high-frequency fluctuation sequences based on operation status data, dividing the multi-scale high-frequency fluctuation sequences into sliding windows along the time axis, statistically analyzing the amplitude variance and energy distribution characteristics within each sliding window, and aggregating and outputting dynamic fluctuation characteristics. S22. Extract the instantaneous flow rate and pressure mean sequence based on the operating status data, synchronize the instantaneous flow rate and pressure mean sequence with the dynamic fluctuation characteristics in the time dimension, and fuse and output the operating parameter sequence. S23. Reconstruct the dynamic fluctuation characteristics and the operating parameter sequence in phase space, map the reconstruction result to a two-dimensional phase plane, extract the local tangent slope of the phase trajectory on the two-dimensional phase plane along the time axis, and output the trajectory slope characteristics based on the statistical distribution of the local tangent slope. S24. Calculate the spatial distribution density of slope scatter points in the trajectory slope feature. Based on the spatial distribution density, perform neighborhood connectivity search and adaptive density clustering segmentation on the slope scatter points, extract the boundary envelope of each cluster, map the boundary envelope to the original working condition type, and output the working condition attribute label.

4. The method for dynamic start-stop and pressure stabilization control of a water pump based on operating condition identification according to claim 1, characterized in that, S3 specifically includes: S31. Parse the state transition sequence of the working condition attribute label, identify the trigger state of the control path based on the state transition sequence, lock the current running parameters along the time axis to maintain the state in the absence of triggering, and perform closed-loop update based on the state maintenance duration. S32. When the triggering state is identified, the topology of the pipeline network is obtained and the node connection relationship and pipe segment physical coordinates are extracted. Graph structure encoding is performed on the node connection relationship and pipe segment physical coordinates to discretize and construct the graph topology space. S33. Extract the node monitoring values ​​from the running parameter sequence, align the node monitoring values ​​to the corresponding nodes in the graph topology space, and output the node dimension feature matrix; S34. Extract the static configuration parameters of the pipe segment from the operating parameter sequence, parse the pipe segment topology length and wave propagation benchmark, fuse the pipe segment topology length and wave propagation benchmark along the edge direction of the graph topology space, and adaptively generate the edge weight matrix based on the statistical distribution of the fusion calculation result to complete the weighting of the node and edge dimensions of the graph topology space.

5. The method for dynamic start-stop and pressure stabilization control of a water pump based on operating condition identification according to claim 1, characterized in that, The improved DCRNN model includes a bandgap filtering and resonant coupling layer, a bidirectional temporal recurrent coding layer, a forward temporal decoding layer, and a multidimensional feature map output layer. The bandgap filtering and resonant coupling layer is used to introduce a bandgap filtering and resonant coupling diffusion mechanism based on acoustic impedance constraints. The specific execution process includes: Extract the pipe segment topology length and wave propagation reference from the edge weight matrix, calculate the acoustic admittance characteristics between nodes along the edge direction of the graph topology space, and map the acoustic admittance characteristics into a discrete acoustic impedance distribution matrix. Multidimensional statistical distribution calculations are performed on the dynamic fluctuation characteristics to extract the mean and variance distribution features of the fluctuation frequency at the node level. The mean and variance distribution features of the fluctuation frequency and the acoustic impedance distribution matrix are input into a two-layer mapping network with shared weights to perform nonlinear projection, and the upper and lower boundary thresholds of the local resonant frequency band gap are solved in real time and output as the node state changes. A frequency domain mask matrix is ​​constructed based on the upper and lower boundary thresholds of the local resonant frequency bandgap. The frequency domain mask matrix is ​​multiplied element-wise with the frequency domain transformation result of the dynamic wave characteristics. The frequency domain components falling into the bandgap are suppressed and cleared to zero, and the effective water hammer wave characteristics falling into the passband are extracted. The effective water hammer wave characteristics are used as the driving force and impedance matching product calculation is performed with the acoustic impedance distribution matrix in the feature space to output the resonant coupling space feature tensor. The bidirectional temporal cyclic coding layer is used to input the resonant coupled spatial feature tensor into the gated cyclic unit along the time axis to extract the forward temporal dependency, and input it into the gated cyclic unit along the time axis in the reverse direction to extract the backward temporal dependency. The forward and backward temporal dependencies are concatenated to output the global spatiotemporal hidden state tensor. The forward temporal decoding layer is used to take the global spatiotemporal hidden state tensor as the initial hidden state, and along the inference sequence of future time steps, it concatenates and nonlinearly maps the hidden state decoding tensor output by the previous time step with the input features of the corresponding time step, and outputs the hidden state decoding tensor of the current time step through cyclic rolling update. The multidimensional feature map output layer is used to input the hidden state decoding tensor of future time steps into three parallel fully connected mapping branches. The first branch outputs the primary water hammer evolution state through linear activation processing, the second branch outputs the attribute identifier through Softmax classification processing, and the third branch outputs the amplitude evolution benchmark through regression processing.

6. The method for dynamic start-stop and pressure stabilization control of a water pump based on operating condition identification according to claim 1, characterized in that, S5 specifically includes: S51. Extract the joint probability distribution features of state amplitude and frequency based on the initial water hammer evolution state, extract the statistical proportion features of working condition type based on attribute identification, input the joint probability distribution features and statistical proportion features into a fully connected mapping network to perform nonlinear dimension reduction projection, and output the target phase time difference benchmark. S52. Extract the absolute coordinates of the time axis corresponding to the predicted arrival time node, perform a difference subtraction operation between the absolute coordinates of the time axis and the target phase time difference reference, and output the initial phase offset. S53. Extract the time series fluctuation variance of the initial phase offset sequence, calculate the nonlinear attenuation compensation factor based on the time series fluctuation variance, multiply the nonlinear attenuation compensation factor with the initial phase offset element by element, and output the dynamically corrected offset. S54. Obtain the pipe segment attribute parameters of the pipeline network topology and calculate the node-level limit offset threshold. Extract the node-level offset values ​​of the dynamically corrected offset sequence. Perform a node-by-node comparison between the node-level offset values ​​and the limit offset threshold. Identify abnormal offset nodes that exceed the limit offset threshold. Force reset the offset of abnormal offset nodes to the limit offset threshold. Retain the original values ​​of normal offset nodes that do not exceed the limit. Reconstruct and output the dynamic start-stop delay benchmark based on the reset results and the retained results.

7. The method for dynamic start-stop and pressure stabilization control of a water pump based on operating condition identification according to claim 1, characterized in that, S6 specifically includes: S61. Obtain the system timestamp corresponding to the current system clock, add the system timestamp to the dynamic start / stop delay reference execution time axis, and output the absolute execution time. S62. Extract the real-time speed and frequency distribution characteristics of the inverter's current operating state sequence by anchoring the absolute execution time. Calculate the transient switching overshoot compensation value based on the real-time speed and frequency distribution characteristics. Subtract the transient switching overshoot compensation value from the execution time axis difference of the absolute execution time to output the corrected absolute execution time. S63. Input the correction absolute execution time into the hardware timer interrupt queue, extract the high-level pulse signal triggered when the correction absolute execution time arrives, write the high-level pulse signal as the underlying data stream into the inverter control register, and output the disconnect trigger instruction. S64. Respond to the disconnection trigger command to acquire the real-time operating electrical parameter sequence of the inverter output side, extract the current change timing feature and voltage change timing feature from the real-time operating electrical parameter sequence, perform feature-level tensor splicing on the current change timing feature and voltage change timing feature, and output the primary water hammer wave.

8. The method for dynamic start-stop and pressure stabilization control of a water pump based on operating condition identification according to claim 1, characterized in that, Specifically, S7 includes: S71. Based on the multi-node pressure sensor array of the pipeline network, the real-time pressure sampling sequence under the action of primary water hammer wave is collected, and the real-time pressure sampling sequence is truncated by a sliding window to output the local pressure fluctuation subsequence. S72. Calculate the local mean and local variance of the local pressure fluctuation subsequence, and extract the transient pressure deviation distribution characteristics based on the local mean and local variance; S73. Extract the wave peak amplitude decay trajectory and wave front propagation delay sequence in the primary water hammer evolution state, extract the background pressure mean and background pressure variance of the real-time pressure sampling sequence during the period without water hammer disturbance, and use the background pressure mean and background pressure variance as adaptively generated steady-state constraint features. S74. Perform feature-level tensor concatenation on the transient pressure deviation distribution characteristics, wave peak amplitude attenuation trajectory and wave front propagation delay sequence to output the water hammer response joint tensor; perform standardized residual calculation on the water hammer response joint tensor and the background pressure mean and variance to output the dynamic deviation index sequence. S75. Extract the time-domain cumulative features of deviation and the local impact distribution features based on the dynamic deviation index sequence, and perform nonlinear state mapping to output the error compensation coefficient.

9. The method for dynamic start-stop and pressure stabilization control of a water pump based on operating condition identification according to claim 1, characterized in that, S8 specifically includes: S81. Extract the proportional and differential components of the error compensation coefficient, perform element-wise multiplication of the proportional component with the dynamic start-stop delay reference, and perform element-wise multiplication of the differential component with the amplitude evolution reference to output the initial corrected delay reference and the initial corrected amplitude reference. S82. Extract the historical delay update sequence corresponding to the initial corrected delay benchmark and the historical amplitude update sequence corresponding to the initial corrected amplitude benchmark, and calculate the delay update variance of the historical delay update sequence and the amplitude update variance of the historical amplitude update sequence. S83. Calculate the multidimensional joint fluctuation distribution characteristics based on the delayed update variance and the amplitude update variance. Perform historical state truncation and local variance gradient calculation on the multidimensional joint fluctuation distribution characteristics through a time-series sliding window to extract the time-series convergence trend characteristics. Calculate the adaptive step size scaling factor based on the time-series convergence trend characteristics. S84. Perform element-wise multiplication of the adaptive step size scaling factor with the initial corrected delay reference and the initial corrected amplitude reference respectively, output the target corrected delay reference and the target corrected amplitude reference, and perform evolution step size mapping calculation along the time step to output the target deduced state transition matrix. S85. Concatenate the target correction delay benchmark and the target correction amplitude benchmark to output the linked correction feature vector. Perform tensor splicing bias on the input layer of the dynamic feature extraction stage based on the linked correction feature vector. At the same time, inject the linked correction feature vector and the target inference state transition matrix as feedforward driving terms into the improved DCRNN model. Perform feature fusion of the resonant coupling layer, hidden state matrix multiplication update of the bidirectional temporal cyclic coding layer, and rolling decoding of the forward temporal decoding layer in sequence to output the feedforward driving features of the next iteration cycle.

Images