Dynamic optimization control method and system for multi-modal pump set

Abstract

The invention relates to the field of pump set control, in particular to a multi-modal pump set dynamic optimization control method and system, and the method comprises the steps: collecting the current waveform, pressure pulsation spectrum and flow time sequence data of each unit of a pump set through a distributed sensor array, and carrying out the time-frequency domain fusion processing to generate an operation state tensor; and inputting a multi-modal identification neural network, and outputting a dominant modal identifier and a modal confidence vector. And according to the modal confidence, activating a basic optimization strategy library or executing cross-modal strategy collaboration, and generating an optimization instruction. And dynamically selecting an instruction conversion path to obtain a frequency converter parameter set, loading the frequency converter parameter set to a programmable logic controller, and driving an execution mechanism to converge a moving track. The method can effectively cope with changes of multi-mode operation states and complex working conditions, and dynamic optimization of the pump set is achieved.

Description

Technical Field

[0001] The present invention relates to the field of pump group control, and specifically relates to a multi-modal pump group dynamic optimization control method and system. Background Art

[0002] In modern industrial and civil fields, pump group systems are widely used in various fluid transportation scenarios, and their operating efficiency and stability have an important impact on the production process and quality of life. Traditional pump group control methods mainly rely on single sensor data and fixed control strategies, and it is difficult to adapt to complex working conditions changes and multi-modal operating states. For example, the invention patent application with the patent number CN113464412A discloses a multi-pump parallel coordinated control method. This method preset the expected output water pressure value, selects one of the multiple intelligent control water pumps in the multi-pump parallel system as the host, and the remaining intelligent control water pumps as slaves, collects the actual output water pressure value of the multi-pump parallel system, and compares the actual output water pressure value with the expected output water pressure value, and adjusts the operating states of all intelligent control water pumps according to the comparison result to achieve the effect of voltage stabilization. However, this method only focuses on the stable control of water pressure, lacks comprehensive monitoring and dynamic optimization of the multi-modal operating states of the pump group, and is difficult to meet the high-efficiency operation requirements of the pump group system under complex working conditions. Summary of the Invention

[0003] The purpose of the present invention is to provide a multi-modal pump group dynamic optimization control method and system to solve the problem that the existing pump group control methods cannot effectively cope with multi-modal operating states and complex working conditions changes.

[0004] To achieve the above purpose, the following technical solutions are adopted.

[0005] A multi-modal pump group dynamic optimization control method includes the following steps, S1. Real-time collect the current waveform data, pressure pulsation spectrum and flow time series of each pump unit in the pump group system through a distributed sensor array, and perform time-frequency domain fusion processing on the three types of data to generate a pump group operating state tensor; S2. Input the pump group operating state tensor into a pre-constructed multi-modal recognition neural network, and output the current dominant mode identifier and mode confidence vector. The dominant mode identifier includes an efficient area operation mode, an overload warning mode, a low-frequency oscillation mode, a flow mismatch mode and a mixed conflict mode; S3. When the maximum value in the mode confidence vector exceeds a preset threshold, activate the basic optimization strategy library corresponding to the dominant mode identifier and output a basic optimization instruction; when it is identified as a mixed conflict mode or the maximum value in the mode confidence vector does not exceed the preset threshold, execute the cross-modal strategy coordination step: Extract the high-dimensional feature vector of the current operating state tensor; Calculate the compatibility weights of each basic optimization strategy; Generate a hybrid optimization instruction based on weight fusion; S4. Dynamically select an instruction conversion path according to the type of optimization instruction output in step S3 to obtain a frequency converter parameter set; S5. Load the frequency converter parameter set into the programmable logic controller of the target pump unit, and drive the actuator of the target pump unit to make the operation trajectory of the pump group converge within the target parameter envelope.

[0006] Optionally, it further includes S6. Collect the system response matrix after the instruction execution, calculate the policy effectiveness evaluation index, and trigger any of the following according to the index deviation degree: When the index deviation degree ≤ the preset value, finely adjust the convolution kernel parameters of the multi-modal recognition neural network; When the index deviation degree > the preset value, reconstruct the decision tree topology of the basic optimization strategy library.

[0007] Optionally, the specific steps of step S1 include S11. Synchronously collect the current waveform data, pressure pulsation spectrum and flow time series of each pump unit through a distributed sensor array; S12. Perform fundamental wave separation processing on the current waveform data and output the effective current component and harmonic distortion component; S13. Perform band energy integration on the pressure pulsation spectrum and output the main frequency band amplitude set and the secondary frequency band energy distribution ratio; S14. Perform sliding window statistical analysis on the flow time series and output the flow mean sequence and the fluctuation variance sequence; S15. Integrate the effective current component and harmonic distortion component output in step S12, the main frequency band amplitude set and the secondary frequency band energy distribution ratio output in step S13, and the flow mean sequence and the fluctuation variance sequence output in step S14 according to the following structure: The first physical dimension: the spatial position identifier of each pump unit; The second physical dimension: the physical characteristic category of three types of parameters, namely current / pressure / flow; The third physical dimension: the continuous sampling time window number; S16. Perform feature dimensionality reduction and compression on the integrated data structure to generate the operation state tensor of the pump group.

[0008] Optionally, the specific steps of step S2 include S21. Input the operation state tensor of the pump group into the first convolutional layer of a pre-constructed multi-modal recognition neural network to extract spatial dimension features and output a primary feature map; S22. Input the primary feature map into a temporal convolutional layer to extract time-series correlation features and output a spatio-temporal fusion feature tensor; S23. Input the spatio-temporal fusion feature tensor into a modality classifier to calculate the matching probability values for each preset modality and generate a modality confidence vector; S24. Perform a maximum value screening operation on the modality confidence vector: If the maximum value ≥ the first threshold and the standard deviation of the probability distribution ≤ the second threshold, output the dominant modality identifier corresponding to a single modality; If the maximum value < the first threshold or the standard deviation of the probability distribution > the second threshold, output a mixed conflict modality identifier; Wherein the preset modalities include: efficient area operation modality, overload warning modality, low-frequency oscillation modality, and flow mismatch modality.

[0009] Optionally, the construction and invocation of the basic optimization strategy library include, Receiving the dominant modality identifier as a strategy selection signal; When the strategy selection signal is the efficient area operation modality: Obtain the current total efficiency benchmark value of the pump group, calculate the efficiency deviation coefficient of each pump unit, generate a start-stop priority queue for the pump units based on the efficiency deviation coefficient, and output a basic optimization instruction including the target start-stop pump numbers; When the strategy selection signal is the overload warning modality: Extract the real-time load rates of each pump unit, identify the overloaded pump units and the set of underloaded pump units, calculate the load transfer matrix, and output a basic optimization instruction including the load redistribution parameters; When the strategy selection signal is the low-frequency oscillation modality: Obtain the set of main frequency phase angles of the pressure pulsation, calculate the phase compensation amount and the damping injection amount, and output a basic optimization instruction including the phase shift instruction; When the strategy selection signal is the flow mismatch modality: Analyze the target flow - actual flow difference distribution, generate a flow correction gain coefficient vector, and output a basic optimization instruction including the gain coefficient.

[0010] Optionally, the specific steps of cross-modal strategy collaboration include: Extract the time-domain feature vector, frequency-domain feature vector, and time-series correlation feature vector from the pump group operation state tensor output in step S2, and combine them to generate a high-dimensional feature vector; Calculate the matching degree between the high-dimensional feature vector and the constraint conditions of each basic optimization strategy: Calculate the energy efficiency adaptation degree of the efficient area strategy, calculate the load balance degree of the overload strategy, calculate the phase compatibility degree of the oscillation strategy, calculate the flow response degree of the mismatch strategy, and output a set of strategy compatibility metrics; Perform normalization weighting based on the policy compatibility metric set: When there is a single metric value ≥ 0.8, set the policy weight to 0.7. When all metric values < 0.8, allocate the weight coefficients proportionally to generate a policy weight vector; Perform linear superposition on the basic optimization policy instructions according to the policy weight vector: Extract the start-stop priority queue of the high-efficiency area policy, extract the load transfer matrix of the overload policy, extract the phase compensation amount of the oscillation policy, extract the flow correction gain of the mismatch policy, and fuse them according to the weight coefficients to generate a hybrid optimization instruction.

[0011] Optionally, step S4 specifically includes S41. Determine the type of the optimization instruction output by step S3: If it is a basic optimization instruction, execute the linear conversion path: Extract the start-stop priority queue, load transfer matrix, phase compensation amount, and flow correction gain in the basic optimization instruction; Convert each parameter into a frequency modulation amount, phase offset amount, and power compensation amount through a linear mapping function; Combine to generate the first frequency converter parameter set; If it is a hybrid optimization instruction, execute the non-linear conversion path: Analyze the policy weight vector and the basic policy parameter set in the hybrid optimization instruction; Input the policy weight vector into the pre-trained neural network transducer; Generate a frequency modulation waveform, phase compensation curve, and power response surface through non-linear transformation; Integrate to generate the second frequency converter parameter set; S42. Perform physical dimension alignment processing on the converted frequency converter parameter set: Extract the pump group spatial position identifier, and reorganize the parameter set into a three-dimensional parameter matrix according to the spatial position identifier, where: The first dimension: the spatial index of the pump unit, The second dimension: the time window sequence, The third dimension: the parameter type channel; S43. Output the reorganized three-dimensional parameter matrix as the frequency converter parameter set for step S5 to load and execute.

[0012] Optionally, step S5 specifically includes: S51. Analyze the frequency converter parameter set output by step S4, and extract the parameter subset corresponding to the target pump unit spatial position identifier; S52. According to the parameter type channel identifier of the parameter subset: If it is a frequency modulation amount, generate a PWM waveform loading instruction; if it is a phase offset amount, generate a time delay compensation instruction; if it is a power compensation amount, generate an IGBT trigger pulse sequence; Combine them into an executable code block for the programmable logic controller; S53. Inject the executable code block into the programmable logic controller runtime environment of the target pump unit to drive the actuator to generate a mechanical adjustment action; S54. Collect the feedback signals of the actuator in real time: the actual change in motor speed, the valve opening adjustment value, and the instantaneous pressure response value, and construct an execution response triple; S55. Dynamically compare the execution response triple with the boundary thresholds of the target parameter envelope: If all parameters ∈ [lower threshold, upper threshold], it is determined to be in a converged state. If any parameter exceeds [lower threshold, upper threshold], perform trajectory correction: calculate the parameter deviation gradient value, generate a parameter compensation amount based on the gradient value, superimpose the parameter compensation amount on the parameter subset in step S51, and trigger a new round of execution loading; S56. When the converged state lasts for more than 3 control cycles, send a closed-loop confirmation signal to step S6.

[0013] Optionally, the step S6 specifically includes: S61. Receive the execution response triple output by step S54, and aggregate it into a system response matrix according to the time window sequence, where the row dimension of the matrix corresponds to the pump unit space index, and the column dimension corresponds to the response parameter type; S62. Extract the following policy effectiveness evaluation elements from the system response matrix: the deviation of the actual energy efficiency improvement rate from the target value, the standard deviation of the load balance degree, the pressure oscillation attenuation rate, and the flow tracking response delay, and calculate the policy effectiveness evaluation index based on the weighted elements; S63. Calculate the index deviation degree between the policy effectiveness evaluation index and the target reference value, and the target reference value is set according to the historical optimal operation data of the pump group; S64. When the index deviation degree ≤ the preset value, perform model fine-tuning: extract the convolutional kernel parameter set of the multi-modal recognition neural network in step S2, calculate the parameter gradient based on the index deviation degree, and update the convolutional kernel parameters in the gradient direction to generate a fine-tuned model; S65. When the index deviation degree > the preset value, perform policy reconstruction: obtain the decision tree topology of the basic optimization policy library, parse the failure path nodes in the system response matrix, and reconstruct the decision tree branch structure based on the node failure frequency, and output the reconstructed decision tree topology; S66. Load the fine-tuned model or the reconstructed decision tree topology into the running environment for the next control cycle to call.

[0014] A multi-modal pump group dynamic optimization control system includes A multi-source sensing acquisition module for real-time collecting the current waveform data, pressure pulsation spectrum, and flow time series of each pump unit in the pump group system through a distributed sensor array; A time-frequency fusion module for performing time-frequency domain fusion processing on the current waveform data, pressure pulsation spectrum, and flow time series to generate a pump unit operation state tensor; A modal identification module for inputting the pump unit operation state tensor into a pre-constructed multi-modal identification neural network and outputting a current dominant mode identifier and a modal confidence vector; A strategy selection module for activating a basic optimization strategy library to output a basic optimization instruction when the maximum value in the modal confidence vector exceeds a preset threshold, and for executing cross-modal strategy collaboration to generate a hybrid optimization instruction when a hybrid conflict mode is identified or the maximum value in the modal confidence vector does not exceed the preset threshold; An instruction conversion module for dynamically selecting a conversion path according to the type of optimization instruction to generate a frequency converter parameter set; An execution control module for loading the frequency converter parameter set into the programmable logic controller of the target pump unit and driving the actuator to converge the pump unit operation trajectory within the target parameter envelope; An adaptive optimization module for collecting a system response matrix to calculate a strategy effectiveness evaluation index, and triggering fine-tuning of the convolution kernel parameters of the modal identification module or reconfiguration of the decision tree topology of the basic optimization strategy library according to the index deviation degree.

[0015] Compared with the prior art, the present invention has the following beneficial effects: In view of the problem that the pump unit control method in the prior art cannot effectively cope with multi-modal operation states and complex working condition changes, the present application proposes a multi-modal pump unit dynamic optimization control method. The method uses a distributed sensor array to collect in real time the current waveform data, pressure pulsation spectrum, and flow time series of each pump unit in the pump unit system, and performs time-frequency domain fusion processing on these data to generate a pump unit operation state tensor. Then, a pre-constructed multi-modal identification neural network is used to identify the current dominant mode, and corresponding optimization strategies are selected according to the modal confidence vector. Finally, the optimization instruction is converted into a frequency converter parameter set and loaded into the programmable logic controller of the target pump unit to achieve dynamic optimization of the pump unit operation trajectory.

[0016] First, through the fusion of multimodal data and modal recognition of neural networks, the current operating mode of the pump set can be accurately identified, thereby providing targeted optimization strategies for different operating states, effectively solving the problem in the prior art that it is unable to adapt to multimodal operating states. Second, this method can dynamically select the conversion path of optimization instructions, generate a set of inverter parameters suitable for different working conditions, and further improve the operating efficiency and stability of the pump set. In addition, by collecting the system response matrix after the execution of the instructions and calculating the policy effectiveness evaluation index, the effect of the optimization strategy can be monitored in real time, and model fine-tuning or policy reconstruction can be triggered according to the index deviation degree to ensure the long-term stable operation of the system. Other technical effects also include: improving the accuracy and reliability of data through detailed data collection and processing steps; enhancing the accuracy of modal recognition through specific modal recognition steps; achieving refined optimization for different modes by constructing and invoking the basic optimization strategy library; improving the adaptability and flexibility of the optimization strategy through cross-modal strategy coordination steps; enhancing the dynamic response ability of the system by dynamically selecting the instruction conversion path; improving the control accuracy and response speed of the actuator through specific execution control steps; and achieving the adaptive optimization and long-term stable operation of the system through the adaptive optimization module. Description of the Drawings

[0017] Figure 1 is a schematic flow chart of the steps of an embodiment of a multimodal pump set dynamic optimization control method according to the present invention.

[0018] Figure 2 is a schematic module structure diagram of an embodiment of a multimodal pump set dynamic optimization control system according to the present invention. Detailed Embodiments

[0019] The present invention will be described in detail below with reference to the drawings and in conjunction with the embodiments. It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0020] The following detailed descriptions are all exemplary descriptions, aiming to provide further detailed descriptions of the present invention. Unless otherwise specified, all technical terms adopted by the present invention have the same meaning as commonly understood by those of ordinary skill in the art to which the present application belongs. The terms used in the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present invention.

[0021] Embodiment 1

[0022] As Figure 1As shown in the figure, this embodiment provides a multi-modal pump group dynamic optimization control method. This method uses a distributed sensor array to collect the current waveform data, pressure pulsation spectrum, and flow time series of each pump unit in the pump group system in real time, and performs time-frequency domain fusion processing on these data to generate the pump group operation state tensor. Then, a pre-constructed multi-modal recognition neural network is used to identify the current dominant mode, and the corresponding optimization strategy is selected according to the modal confidence vector. Finally, the optimization instruction is converted into a frequency converter parameter set and loaded into the programmable logic controller of the target pump unit to achieve the dynamic optimization of the pump group operation trajectory.

[0023] Step S1: Data acquisition and processing In the pump group system, the current waveform data, pressure pulsation spectrum, and flow time series of each pump unit are collected in real time through a distributed sensor array. Specifically, the current waveform data is collected by a high-precision current sensor, the pressure pulsation spectrum is collected by a high-sensitivity pressure sensor, and the flow time series is collected by a high-precision flow sensor. These sensors are distributed on each pump unit of the pump group system and can collect key data during the operation of the pump group in real time and accurately.

[0024] The collected current waveform data is first subjected to fundamental wave separation processing to separate the effective value component and harmonic distortion component of the current. The specific steps are as follows: Use the fast Fourier transform (FFT) to perform frequency domain analysis on the current waveform data to extract the fundamental wave and each harmonic.

[0025] Calculate the effective value of the fundamental wave as the effective value component of the current.

[0026] Calculate the total effective value of each harmonic as the harmonic distortion component.

[0027] The pressure pulsation spectrum performs band energy integration and outputs the main frequency band amplitude set and the secondary frequency band energy distribution ratio. The specific steps are as follows: Divide the pressure pulsation spectrum into multiple frequency bands, such as 0-10Hz, 10-20Hz, 20-30Hz, etc.

[0028] Perform energy integration on each frequency band to calculate the total energy of the frequency band.

[0029] Extract the amplitude set of the main frequency band, that is, the amplitudes of the several frequency bands with the largest energy.

[0030] Calculate the energy distribution ratio of the secondary frequency band, that is, the ratio of the energy of each secondary frequency band to the total energy.

[0031] The flow time series is subjected to sliding window statistical analysis to output the flow mean sequence and the fluctuation variance sequence. The specific steps are as follows: Segment the traffic time series using a sliding window. The window size can be set according to actual needs, such as 1 second, 5 seconds, etc.

[0032] Calculate the mean of the traffic data within each window to form a traffic mean sequence.

[0033] Calculate the variance of the traffic data within each window to form a traffic fluctuation variance sequence.

[0034] Integrate the processed data according to the following structure: The first physical dimension: the spatial position identifiers of each pump unit, such as pump unit 1, pump unit 2, etc.

[0035] The second physical dimension: the physical characteristic categories of three types of parameters, current / pressure / flow.

[0036] The third physical dimension: the continuous sampling time window numbers, such as window 1, window 2, etc.

[0037] Further perform feature dimensionality reduction and compression on the integrated data structure to generate a pump group operation state tensor. The specific steps are as follows: Use principal component analysis (PCA) to reduce the dimensionality of the integrated data and extract the main features.

[0038] Construct the dimensionality-reduced feature data into a multi-dimensional tensor, where each dimension corresponds to different physical meanings, such as spatial position, parameter category, time window, etc.

[0039] Step S2, Modal Recognition and Optimization Strategy Selection Input the generated pump group operation state tensor into a pre-constructed multi-modal recognition neural network. This neural network extracts spatial dimension features through the first convolutional layer and outputs a primary feature map; then extracts time series correlation features through the time dimension convolutional layer and outputs a spatio-temporal fusion feature tensor. The specific steps are as follows: The first convolutional layer uses multiple convolutional kernels to perform convolutional operations on the operation state tensor to extract features in the spatial dimension.

[0040] The time dimension convolutional layer performs convolutional operations on the primary feature map to extract the correlation features of the time series and generates a spatio-temporal fusion feature tensor.

[0041] The spatio-temporal fusion feature tensor is further input into a modal classifier to calculate the matching probability values of each preset mode and generate a modal confidence vector. The specific steps are as follows: The modal classifier can use classification algorithms such as support vector machine (SVM), decision tree, etc.

[0042] Classify the spatio-temporal fusion feature tensor and calculate the matching probability values of each preset mode.

[0043] Construct the matching probability values into a modal confidence vector, e.g., [0.8, 0.1, 0.05, 0.05], indicating that the probability of the high-efficiency region operation mode is 0.8, and the probabilities of other modes are 0.1, 0.05, and 0.05 respectively.

[0044] In step S3, the maximum value in the modal confidence vector is used to determine the current dominant mode. If the maximum value exceeds the preset threshold and the standard deviation of the probability distribution is within a certain range, the dominant mode identifier corresponding to the single mode is output; otherwise, the mixed conflict mode identifier is output. The specific steps are as follows: Set the preset threshold, e.g., 0.7, and the standard deviation threshold of the probability distribution, e.g., 0.1.

[0045] If the maximum value ≥ 0.7 and the standard deviation of the probability distribution ≤ 0.1, output the dominant mode identifier corresponding to the single mode.

[0046] Otherwise, output the mixed conflict mode identifier.

[0047] The preset modes include the high-efficiency region operation mode, overload warning mode, low-frequency oscillation mode, flow mismatch mode, etc. When a single mode is identified, activate the basic optimization strategy library corresponding to the dominant mode identifier and output the basic optimization instruction. The specific steps are as follows: High-efficiency region operation mode: Obtain the current total efficiency reference value of the pump group, calculate the efficiency deviation coefficient of each pump unit, generate a start-stop priority queue for the pump units, and output the basic optimization instruction including the target start-stop pump numbers.

[0048] Calculate the efficiency deviation coefficient of each pump unit, e.g., efficiency deviation coefficient = current efficiency / reference efficiency.

[0049] Generate a start-stop priority queue according to the efficiency deviation coefficient, give priority to starting the pump units with high efficiency, and deactivate the pump units with low efficiency.

[0050] Overload warning mode: Extract the real-time load rates of each pump unit, identify the set of overloaded pump units and underloaded pump units, calculate the load transfer matrix, and output the basic optimization instruction including the load redistribution parameters.

[0051] Calculate the real-time load rate of each pump unit, e.g., load rate = current load / rated load.

[0052] Identify the overloaded pump units and underloaded pump units, calculate the load transfer matrix, and transfer the load of the overloaded pump units to the underloaded pump units.

[0053] Low-frequency oscillation mode: Obtain the set of main frequency phase angles of the pressure pulsation, calculate the phase compensation amount and damping injection amount, and output the basic optimization instruction including the phase shift instruction.

[0054] Calculate the set of main frequency phase angles of the pressure pulsation, for example, extract the main frequency phase angle through Fourier transform.

[0055] Calculate the phase compensation amount and damping injection amount, generate a phase offset command, adjust the operating phase of the pump unit, and reduce oscillation.

[0056] Flow mismatch mode: Analyze the distribution of the difference between the target flow rate and the actual flow rate, generate a flow correction gain coefficient vector, and output a basic optimization instruction including the gain coefficient.

[0057] Calculate the distribution of the difference between the target flow rate and the actual flow rate, for example, difference = target flow rate - actual flow rate.

[0058] Generate a flow correction gain coefficient vector, adjust the flow rate of the pump unit, and make the actual flow rate close to the target flow rate.

[0059] When it is identified as the hybrid conflict mode or the maximum value in the mode confidence vector does not exceed the preset threshold, execute the cross-modal strategy coordination steps. The specific steps are as follows: Extract the time-domain feature vector, frequency-domain feature vector, and time-series correlation feature vector from the pump group operating state tensor, and combine them to generate a high-dimensional feature vector.

[0060] Calculate the matching degree between the high-dimensional feature vector and the constraint conditions of each basic optimization strategy, and output a set of strategy compatibility metrics.

[0061] Calculate the energy efficiency adaptation degree of the high-efficiency zone strategy, for example, energy efficiency adaptation degree = current energy efficiency / target energy efficiency.

[0062] Calculate the load balancing degree of the overload strategy, for example, load balancing degree = current load balancing degree / target load balancing degree.

[0063] Calculate the phase compatibility degree of the oscillation strategy, for example, phase compatibility degree = current phase compatibility degree / target phase compatibility degree.

[0064] Calculate the flow response degree of the mismatch strategy, for example, flow response degree = current flow response degree / target flow response degree.

[0065] Perform normalized weighting based on the set of strategy compatibility metrics to generate a strategy weight vector.

[0066] When there is a single metric value ≥ 0.8, set the strategy weight to 0.7.

[0067] When all metric values < 0.8, allocate the weight coefficients proportionally to generate a strategy weight vector.

[0068] Linearly superimpose the basic optimization strategy instructions according to the strategy weight vector to generate a hybrid optimization instruction.

[0069] The start-stop priority queue of the extraction efficient area strategy.

[0070] The load transfer volume matrix of the extraction overload strategy.

[0071] The phase compensation amount of the extraction oscillation strategy.

[0072] The flow correction gain of the extraction mismatch strategy.

[0073] Fuse according to the weight coefficient to generate a hybrid optimization instruction.

[0074] Instruction conversion and execution control Step S4: Dynamically select the instruction conversion path according to the type of optimization instruction. The specific steps are as follows: Basic optimization instruction: Extract the start-stop priority queue, load transfer volume matrix, phase compensation amount, and flow correction gain in the basic optimization instruction.

[0075] Convert each parameter into a frequency modulation amount, phase offset amount, and power compensation amount through a linear mapping function.

[0076] Combine to generate the first frequency converter parameter set.

[0077] Hybrid optimization instruction: Analyze the strategy weight vector and basic strategy parameter set in the hybrid optimization instruction.

[0078] Input the strategy weight vector into the pre-trained neural network transducer.

[0079] Generate a frequency modulation waveform, phase compensation curve, and power response surface through a non-linear transformation.

[0080] Integrate to generate the second frequency converter parameter set.

[0081] Perform physical dimension alignment processing on the converted frequency converter parameter set, extract the pump group spatial position identifier, and reorganize the parameter set into a three-dimensional parameter matrix according to the spatial position identifier. The specific steps are as follows: Extract the pump group spatial position identifier, such as pump unit 1, pump unit 2, etc.

[0082] Reorganize the parameter set into a three-dimensional parameter matrix according to the spatial position identifier, where: The first dimension: the spatial index of the pump unit.

[0083] The second dimension: the time window sequence.

[0084] The third dimension: the parameter type channel.

[0085] Output the reorganized three-dimensional parameter matrix as the frequency converter parameter set for loading and execution.

[0086] Step S5: Load the frequency converter parameter set into the programmable logic controller of the target pump unit, and drive the actuator to generate mechanical adjustment actions. The specific steps are as follows: Analyze the frequency converter parameter set and extract the parameter subset corresponding to the spatial position identifier of the target pump unit.

[0087] According to the parameter type channel identifier of the parameter subset: If it is a frequency modulation quantity, generate a PWM waveform loading instruction.

[0088] If it is a phase offset quantity, generate a time delay compensation instruction.

[0089] If it is a power compensation quantity, generate an IGBT trigger pulse sequence.

[0090] Combine them into an executable code block that can be recognized by the programmable logic controller.

[0091] Inject the executable code block into the runtime environment of the programmable logic controller of the target pump unit to drive the actuator to generate mechanical adjustment actions.

[0092] Collect the feedback signals of the actuator in real time, including the actual change in motor speed, the valve opening adjustment value, and the instantaneous pressure response value, and construct an execution response triple. The specific steps are as follows: Collect the actual change in motor speed, for example, through a speed sensor.

[0093] Collect the valve opening adjustment value, for example, through a position sensor.

[0094] Collect the instantaneous pressure response value, for example, through a pressure sensor.

[0095] Construct an execution response triple, for example, (speed change amount, valve opening adjustment value, pressure response value).

[0096] Dynamically compare the execution response triple with the boundary threshold of the target parameter envelope. The specific steps are as follows: If all parameters are within the threshold range, for example, [threshold lower limit, threshold upper limit], it is determined to be in a converged state.

[0097] If any parameter exceeds the threshold range, perform trajectory correction: Calculate the parameter deviation gradient value, for example, through differential calculation.

[0098] Generate a parameter compensation amount based on the gradient value, for example, through a proportional integral derivative (PID) controller.

[0099] Superimpose the parameter compensation amount on the parameter subset to trigger a new round of execution loading.

[0100] When the convergence state lasts for more than 3 control cycles, send a closed-loop confirmation signal to the subsequent steps.

[0101] Step S6, Adaptive Optimization Collect the system response matrix after the instruction execution, calculate the policy effectiveness evaluation index, and trigger model fine-tuning or policy reconstruction according to the index deviation degree. The specific steps are as follows: Aggregate and generate the system response matrix according to the time window sequence, where the row dimension of the matrix corresponds to the spatial index of the pump unit, and the column dimension corresponds to the type of response parameter.

[0102] Extract the policy effectiveness evaluation elements from the system response matrix, including: The deviation of the actual energy efficiency improvement rate from the target value.

[0103] The standard deviation of the load balance degree.

[0104] The pressure oscillation decay rate.

[0105] The flow tracking response delay.

[0106] Calculate the policy effectiveness evaluation index based on element weighting, such as weighted sum.

[0107] Calculate the index deviation degree between the policy effectiveness evaluation index and the target reference value, and the target reference value is set according to the historical optimal operation data of the pump group. The specific steps are as follows: Set the target reference value, such as the energy efficiency improvement rate reference value, the load balance degree reference value, etc.

[0108] Calculate the deviation degree between the policy effectiveness evaluation index and the target reference value, for example, deviation degree = current index - reference value.

[0109] When the index deviation degree is less than or equal to the preset value, perform model fine-tuning. The specific steps are as follows: Extract the set of convolutional kernel parameters of the multi-modal recognition neural network.

[0110] Calculate the parameter gradient based on the index deviation degree, for example, through the backpropagation algorithm.

[0111] Update the convolutional kernel parameters in the gradient direction to generate the fine-tuned model.

[0112] When the index deviation degree is greater than the preset value, perform policy reconstruction. The specific steps are as follows: Obtain the decision tree topology of the basic optimization policy library.

[0113] Parse the failure path nodes in the system response matrix.

[0114] Reconstruct the decision tree branch structure based on the node failure frequency, and output the reconstructed decision tree topology.

[0115] Load the fine-tuned model or reconstructed decision tree topology into the operating environment for invocation in the next control cycle.

[0116] Through the above steps, this embodiment provides a multi-modal pump unit dynamic optimization control method, which can effectively cope with the multi-modal operating states and complex working condition changes of the pump unit system, and improve the operating efficiency and stability of the pump unit. This method can not only monitor and optimize the operating state of the pump unit in real time, but also achieve the long-term stable operation of the system through the adaptive optimization module, and has broad application prospects and significant economic benefits.

[0117] Embodiment 2 As Figure 2 shown, this embodiment provides a multi-modal pump unit dynamic optimization control system. This system is based on the multi-modal pump unit dynamic optimization control method described in the above Embodiment 1, and realizes the efficient and stable operation of the pump unit system through the coordinated work of multiple modules. The system includes a multi-source sensing and acquisition module, a time-frequency fusion module, a modal identification module, a strategy selection module, an instruction conversion module, an execution control module, and an adaptive optimization module.

[0118] The multi-source sensing and acquisition module is responsible for collecting the current waveform data, pressure pulsation spectrum, and flow time series of each pump unit in the pump unit system in real time through a distributed sensor array. Specifically, this module includes: Current sensor: A high-precision current sensor for collecting the current waveform data of each pump unit in real time.

[0119] Pressure sensor: A high-sensitivity pressure sensor for collecting the pressure pulsation spectrum of each pump unit in real time.

[0120] Flow sensor: A high-precision flow sensor for collecting the flow time series of each pump unit in real time.

[0121] These sensors are distributed on each pump unit of the pump unit system, and can collect the key data during the operation of the pump unit in real time and accurately, and transmit the collected data to the time-frequency fusion module.

[0122] The time-frequency fusion module receives the data transmitted by the multi-source sensing and acquisition module, performs time-frequency domain fusion processing on the current waveform data, pressure pulsation spectrum, and flow time series, and generates a pump unit operation state tensor. The specific processing steps include: Fundamental wave separation processing: Perform fundamental wave separation processing on the current waveform data to extract the effective current component and harmonic distortion component.

[0123] Band energy integration: Perform band energy integration on the pressure pulsation spectrum, and output the main frequency band amplitude set and the secondary frequency band energy distribution ratio.

[0124] Sliding window statistical analysis: Perform sliding window statistical analysis on the traffic time series, and output the traffic mean sequence and the fluctuation variance sequence.

[0125] Data integration: Integrate the processed data according to the following structure: The first physical dimension: The spatial position identifier of each pump unit.

[0126] The second physical dimension: The physical characteristic categories of three types of parameters, namely current / pressure / flow.

[0127] The third physical dimension: The continuous sampling time window number.

[0128] Feature dimensionality reduction and compression: Perform feature dimensionality reduction and compression on the integrated data structure to generate the pump group operation state tensor.

[0129] The generated pump group operation state tensor will be transmitted to the modal recognition module.

[0130] The modal recognition module receives the pump group operation state tensor generated by the time-frequency fusion module, identifies the current dominant mode through a pre-constructed multi-modal recognition neural network, and outputs a modal confidence vector. The specific steps include: Spatial dimension feature extraction: Input the pump group operation state tensor into the first convolutional layer of the multi-modal recognition neural network to extract spatial dimension features and output a primary feature map.

[0131] Time series correlation feature extraction: Input the primary feature map into the time dimension convolutional layer to extract time series correlation features and output a spatio-temporal fusion feature tensor.

[0132] Modal classification: Input the spatio-temporal fusion feature tensor into the modal classifier, calculate the matching probability values of each preset mode, and generate a modal confidence vector.

[0133] Modal judgment: Based on the maximum value and the standard deviation of the probability distribution in the modal confidence vector, judge the current dominant mode. If the maximum value exceeds the preset threshold and the standard deviation of the probability distribution is within a certain range, output the dominant mode identifier corresponding to the single mode; otherwise, output the mixed conflict mode identifier.

[0134] The modal recognition module transmits the modal confidence vector and the dominant mode identifier to the strategy selection module.

[0135] The strategy selection module selects the corresponding optimization strategy according to the modal confidence vector and the dominant mode identifier output by the modal recognition module, and outputs an optimization instruction. The specific steps include: Activation of the basic optimization strategy library: When the maximum value in the modal confidence vector exceeds the preset threshold, activate the basic optimization strategy library corresponding to the dominant mode identifier and output a basic optimization instruction.

[0136] Cross-modal strategy collaboration: When it is identified as a mixed conflict mode or the maximum value in the modal confidence vector does not exceed the preset threshold, execute the cross-modal strategy collaboration step to generate a mixed optimization instruction.

[0137] The specific implementation methods of the basic optimization strategy library and cross-modal strategy collaboration are the same as the steps described in Embodiment 1. The strategy selection module transmits the optimization instruction to the instruction conversion module.

[0138] The instruction conversion module dynamically selects the instruction conversion path according to the type of optimization instruction output by the strategy selection module and generates a frequency converter parameter set. The specific steps include: Basic optimization instruction conversion: Extract the start-stop priority queue, load transfer matrix, phase compensation amount, and flow correction gain in the basic optimization instruction.

[0139] Convert each parameter into a frequency modulation amount, phase offset amount, and power compensation amount through a linear mapping function.

[0140] Combine to generate the first frequency converter parameter set.

[0141] Mixed optimization instruction conversion: Analyze the strategy weight vector and the basic strategy parameter set in the mixed optimization instruction.

[0142] Input the strategy weight vector into the pre-trained neural network transducer.

[0143] Generate a frequency modulation waveform, phase compensation curve, and power response surface through a non-linear transformation.

[0144] Integrate to generate the second frequency converter parameter set.

[0145] The instruction conversion module transmits the frequency converter parameter set to the execution control module.

[0146] The execution control module receives the frequency converter parameter set generated by the instruction conversion module, loads it into the programmable logic controller of the target pump unit, and drives the actuator to generate a mechanical adjustment action. The specific steps include: Parameter set analysis: Analyze the frequency converter parameter set and extract the parameter subset corresponding to the spatial position identifier of the target pump unit.

[0147] Instruction generation: If it is a frequency modulation amount, generate a PWM waveform loading instruction.

[0148] If it is a phase offset amount, generate a time delay compensation instruction.

[0149] If it is a power compensation amount, generate an IGBT trigger pulse sequence.

[0150] Code block assembly: Assemble the generated instructions into programmable logic controller executable code blocks.

[0151] Code block injection: Inject the executable code block into the programmable logic controller runtime environment of the target pump unit to drive the actuator to produce mechanical adjustment action.

[0152] Feedback signal acquisition: Real-time collection of actuator feedback signals, including actual motor speed change, valve opening adjustment value, and pressure transient response value, to construct an execution response triplet.

[0153] Dynamic comparison: The execution response triplet is dynamically compared with the boundary threshold of the target parameter envelope. If all parameters are within the threshold range, it is determined to be in a convergence state. If any parameter exceeds the threshold range, trajectory correction is performed, the parameter deviation gradient value is calculated, and the parameter compensation amount is generated. The parameter compensation amount is superimposed on the parameter subset, triggering a new round of execution loading.

[0154] Closed-loop confirmation: When the convergence state lasts for more than 3 control cycles, a closed-loop confirmation signal is sent to the adaptive optimization module.

[0155] The execution control module transmits the system response matrix to the adaptive optimization module.

[0156] The adaptive optimization module receives the system response matrix transmitted by the execution control module, calculates the strategy effectiveness evaluation index, and triggers model fine-tuning or strategy reconstruction based on the index deviation. The specific steps include: System response matrix aggregation: Aggregate the system response matrix by time window sequence, where the row dimension of the matrix corresponds to the pump unit spatial index and the column dimension corresponds to the response parameter type.

[0157] Extraction of strategic effectiveness evaluation factors: Extract the deviation of the actual energy efficiency improvement rate from the target value.

[0158] Extract the standard deviation of load balancing.

[0159] Extract the pressure oscillation decay rate.

[0160] Extract traffic tracking response latency.

[0161] Calculation of strategy effectiveness evaluation index: Calculate the strategy effectiveness evaluation index based on the weighted extraction factors.

[0162] Index deviation calculation: Calculate the index deviation between the strategy effectiveness evaluation index and the target benchmark value. The target benchmark value is set based on the historical optimal operating data of the pump group.

[0163] Model fine-tuning or strategy reconstruction: When the indicator deviation is less than or equal to the preset value, perform model fine-tuning: Extract the set of convolutional kernel parameters of the multi-modal recognition neural network.

[0164] Calculate the parameter gradient based on the metric deviation.

[0165] Update the convolutional kernel parameters in the gradient direction to generate a fine-tuned model.

[0166] When the metric deviation is greater than the preset value, perform policy reconstruction: Obtain the decision tree topology of the basic optimization policy library.

[0167] Analyze the failure path nodes in the system response matrix.

[0168] Reconstruct the decision tree branch structure based on the node failure frequency and output the reconstructed decision tree topology.

[0169] Model or topology loading: Load the fine-tuned model or the reconstructed decision tree topology into the operating environment for use in the next control cycle.

[0170] Through the collaborative work of the above modules, this embodiment provides a multi-modal pump group dynamic optimization control system, which can effectively cope with the multi-modal operating states and complex working condition changes of the pump group system, and improve the operating efficiency and stability of the pump group. This system can not only monitor and optimize the operating state of the pump group in real time, but also achieve the long-term stable operation of the system through the adaptive optimization module, and has broad application prospects and significant economic benefits.

[0171] As is known by common technical knowledge, the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above-disclosed embodiments are illustrative in all respects and not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are encompassed by the present invention.

Claims

1. A multi-modal pump unit dynamic optimization control method, characterized in that Including the following steps, S1. Real-time collect the current waveform data, pressure pulsation spectrum and flow time series of each pump unit in the pump group system through a distributed sensor array, and perform time-frequency domain fusion processing on the three types of data to generate a pump group operation state tensor; S2. Input the pump group operation state tensor into a pre-constructed multi-modal recognition neural network, and output the current dominant mode identifier and mode confidence vector. The dominant mode identifier includes high-efficiency area operation mode, overload warning mode, low-frequency oscillation mode, flow mismatch mode and mixed conflict mode; S3. When the maximum value in the mode confidence vector exceeds the preset threshold, activate the basic optimization strategy library corresponding to the dominant mode identifier and output a basic optimization instruction; when it is identified as a mixed conflict mode or the maximum value in the mode confidence vector does not exceed the preset threshold, execute the cross-modal strategy coordination steps: Extract the high-dimensional feature vector of the current operation state tensor; Calculate the compatibility weights of each basic optimization strategy; Generate a mixed optimization instruction based on weight fusion; S4. According to the type of optimization instruction output in step S3, dynamically select the instruction conversion path to obtain the frequency converter parameter set; S5. Load the frequency converter parameter set into the programmable logic controller of the target pump unit, and drive the actuator of the target pump unit to make the operation trajectory of the pump group converge within the target parameter envelope.

2. The multimodal pump unit dynamic optimization control method according to claim 1, wherein, It also includes, S6. Collect the system response matrix after the instruction is executed, calculate the strategy effectiveness evaluation index, and trigger any of the following according to the index deviation: When the index deviation ≤ the preset value, fine-tune the convolution kernel parameters of the multi-modal recognition neural network; When the index deviation > the preset value, reconstruct the decision tree topology of the basic optimization strategy library.

3. A multimodal pump unit dynamic optimization control method according to claim 1, characterized in that, The specific steps of step S1 Include, S11. Synchronously collect the current waveform data, pressure pulsation spectrum and flow time series of each pump unit through a distributed sensor array; S12. Perform fundamental wave separation processing on the current waveform data and output the effective current component and harmonic distortion component; S13. Perform frequency band energy integration on the pressure pulsation spectrum and output the main frequency band amplitude set and the secondary frequency band energy distribution ratio; S14. Perform sliding window statistical analysis on the flow time series and output the flow mean sequence and fluctuation variance sequence; S15. Integrate the effective current component and harmonic distortion component output in step S12, the main frequency band amplitude set and the secondary frequency band energy distribution ratio output in step S13, and the flow mean sequence and fluctuation variance sequence output in step S14 according to the following structure: The first physical dimension: the spatial position identifier of each pump unit; The second physical dimension: the physical feature category of the three types of parameters of current / pressure / flow; The third physical dimension: the continuous sampling time window number; S16. Perform feature dimensionality reduction and compression on the integrated data structure to generate the pump group operation state tensor.

4. A multimodal pump unit dynamic optimization control method according to claim 1, characterized in that, The specific steps of step S2 Include, S21. Input the pump group operation state tensor into the first convolutional layer of a pre-constructed multi-modal recognition neural network, extract the spatial dimension features, and output the primary feature map; S22. Input the primary feature map into the time dimension convolutional layer, extract the time series correlation features, and output the spatio-temporal fusion feature tensor; S23. Input the spatio-temporal fusion feature tensor into the modal classifier, calculate the matching probability values of each preset mode, and generate a modal confidence vector; S24. Perform a maximum value screening operation on the modal confidence vector: If the maximum value ≥ the first threshold and the standard deviation of the probability distribution ≤ the second threshold, output the dominant mode identifier corresponding to a single mode; If the maximum value < the first threshold or the standard deviation of the probability distribution > the second threshold, output the hybrid conflict mode identifier; Wherein the preset modes include: high-efficiency area operation mode, overload warning mode, low-frequency oscillation mode, and flow mismatch mode.

5. A multimodal pump unit dynamic optimization control method according to claim 1, characterized in that, The construction and invocation of the basic optimization strategy library include Receiving the dominant mode identifier as a strategy selection signal; When the strategy selection signal is the high-efficiency area operation mode: obtain the current total efficiency benchmark value of the pump group, calculate the efficiency deviation coefficient of each pump unit, generate a start-stop priority queue for the pump units based on the efficiency deviation coefficient, and output a basic optimization instruction including the target start-stop pump numbers; When the strategy selection signal is the overload warning mode: extract the real-time load rates of each pump unit, identify the overloaded pump units and the set of underloaded pump units, calculate the load transfer matrix, and output a basic optimization instruction including the load redistribution parameters; When the strategy selection signal is the low-frequency oscillation mode: obtain the set of main frequency phase angles of the pressure pulsation, calculate the phase compensation amount and the damping injection amount, and output a basic optimization instruction including the phase shift instruction; When the strategy selection signal is the flow mismatch mode: analyze the target flow - actual flow difference distribution, generate a flow correction gain coefficient vector, and output a basic optimization instruction including the gain coefficient.

6. The multimodal pump unit dynamic optimization control method according to claim 1, characterized in that, The cross-modal strategy collaboration steps specifically include: Extract the time-domain feature vector, frequency-domain feature vector, and time-series correlation feature vector from the pump group operation state tensor output in step S2, and combine them to generate a high-dimensional feature vector; Calculate the matching degree between the high-dimensional feature vector and the constraint conditions of each basic optimization strategy: calculate the energy efficiency adaptation degree of the high-efficiency area strategy, calculate the load balance degree of the overload strategy, calculate the phase compatibility degree of the oscillation strategy, calculate the flow response degree of the mismatch strategy, and output a strategy compatibility metric set; Perform normalization weighting based on the strategy compatibility metric set: when there is a single metric value ≥ 0.8, set the weight of this strategy to 0.7, and when all metric values < 0.8, allocate the weight coefficients proportionally to generate a strategy weight vector; Linearly superimpose the basic optimization strategy instructions according to the strategy weight vector: extract the start-stop priority queue of the high-efficiency area strategy, extract the load transfer matrix of the overload strategy, extract the phase compensation amount of the oscillation strategy, extract the flow correction gain of the mismatch strategy, and fuse them according to the weight coefficients to generate a hybrid optimization instruction.

7. A multi-modal pump unit dynamic optimization control method according to claim 1, characterized in that The specific steps of S4 include S41. Determine the type of the optimization instruction output in step S3: If it is a basic optimization instruction, execute the linear conversion path: Extract the start-stop priority queue, load transfer matrix, phase compensation amount, and flow correction gain in the basic optimization instruction; Convert each parameter into a frequency modulation amount, a phase shift amount, and a power compensation amount through a linear mapping function; Combine to generate the first frequency converter parameter set; If it is a hybrid optimization instruction, execute the non-linear conversion path: Analyze the policy weight vector and the set of basic policy parameters in the hybrid optimization instruction; Input the policy weight vector into the pre-trained neural network transducer; Generate a frequency modulation waveform, a phase compensation curve, and a power response surface through non-linear transformation; Integrate to generate the second frequency converter parameter set; S42. Perform physical dimension alignment processing on the converted frequency converter parameter set: Extract the spatial position identifier of the pump group, and reorganize the parameter set into a three-dimensional parameter matrix according to the spatial position identifier, where: The first dimension: the spatial index of the pump unit, The second dimension: the time window sequence, The third dimension: the parameter type channel; S43. Output the reorganized three-dimensional parameter matrix as the frequency converter parameter set for loading and execution in step S5.

8. A multi-modal pump unit dynamic optimization control method according to claim 2, characterized in that Step S5 specifically includes: S51. Analyze the frequency converter parameter set output in step S4, and extract the parameter subset corresponding to the target pump unit spatial position identifier; S52. According to the parameter type channel identifier of the parameter subset: If it is a frequency modulation quantity, generate a PWM waveform loading instruction; if it is a phase offset quantity, generate a time delay compensation instruction; if it is a power compensation quantity, generate an IGBT trigger pulse sequence; Combine them into an executable code block for the programmable logic controller; S53. Inject the executable code block into the runtime environment of the programmable logic controller of the target pump unit to drive the actuator to generate a mechanical adjustment action; S54. Real-time collect the feedback signals of the actuator: the actual change in motor speed, the valve opening adjustment value, the instantaneous pressure response value, and construct an execution response triple; S55. Dynamically compare the execution response triple with the boundary threshold of the target parameter envelope: If all parameters ∈ [lower threshold, upper threshold], it is determined to be in a convergence state, If any parameter exceeds [lower threshold, upper threshold], perform trajectory correction: Calculate the parameter deviation gradient value, generate a parameter compensation quantity based on the gradient value, superimpose the parameter compensation quantity on the parameter subset in step S51, and trigger a new round of execution loading; S56. When the convergence state lasts for more than 3 control cycles, send a closed-loop confirmation signal to step S6.

9. A multimodal pump unit dynamic optimization control method according to claim 8, characterized in that The said step S6 specifically includes: S61. Receive the execution response triple output in step S54, and aggregate it into a system response matrix according to the time window sequence, where the row dimension of the matrix corresponds to the pump unit spatial index, and the column dimension corresponds to the response parameter type; S62. Extract the following policy effectiveness evaluation elements from the system response matrix: the deviation of the actual energy efficiency improvement rate from the target value, the standard deviation of the load balance degree, the pressure oscillation attenuation rate, the flow tracking response delay, and calculate the policy effectiveness evaluation index based on the element weighting; S63. Calculate the index deviation degree between the policy effectiveness evaluation index and the target reference value, and the target reference value is set according to the historical optimal operation data of the pump group; S64. When the index deviation degree ≤ the preset value, perform model fine-tuning: Extract the set of convolution kernel parameters of the multi-modal recognition neural network in step S2, calculate the parameter gradient based on the index deviation degree, and update the convolution kernel parameters in the gradient direction to generate a fine-tuned model; S65. When the index deviation degree > preset value, perform policy reconstruction: Obtain the decision tree topology of the basic optimization policy library, parse the failure path nodes in the system response matrix, reconstruct the decision tree branch structure based on the node failure frequency, and output the reconstructed decision tree topology; S66. Load the fine-tuned model or the reconstructed decision tree topology into the operating environment for the next control cycle to call.

10. A multi-modal pump unit dynamic optimization control system, based on the multi-modal pump unit dynamic optimization control method according to any one of claims 1-9, characterized in that, Including, A multi-source sensing acquisition module, configured to collect the current waveform data, pressure pulsation spectrum, and flow time series of each pump unit in the pump group system in real time through a distributed sensor array; A time-frequency fusion module, configured to perform time-frequency domain fusion processing on the current waveform data, pressure pulsation spectrum, and flow time series to generate a pump group operation state tensor; A modal recognition module, configured to input the pump group operation state tensor into a pre-constructed multi-modal recognition neural network, and output the current dominant modal identifier and modal confidence vector; A policy selection module, configured to activate the basic optimization policy library to output basic optimization instructions when the maximum value in the modal confidence vector exceeds the preset threshold, and perform cross-modal policy collaboration to generate hybrid optimization instructions when it is recognized as a mixed conflict mode or the maximum value in the modal confidence vector does not exceed the preset threshold; An instruction conversion module, configured to dynamically select a conversion path according to the type of optimization instruction to generate a frequency converter parameter set; An execution control module, configured to load the frequency converter parameter set into the programmable logic controller of the target pump unit, and drive the actuator to make the pump group operation trajectory converge within the target parameter envelope; An adaptive optimization module, configured to collect the system response matrix to calculate the policy effectiveness evaluation index, and trigger the fine-tuning of the convolution kernel parameters of the modal recognition module or the reconstruction of the decision tree topology of the basic optimization policy library according to the index deviation degree.

Images