A Multimodal Sensing-Based Adaptive Intelligent Regulation and Optimization Method and System for Circulating Water

Abstract

This invention relates to the field of intelligent operation and control technology for industrial water treatment, and discloses a method and system for adaptive intelligent regulation and optimization of circulating water based on multimodal perception. One method includes: acquiring vibration signals, performance parameters, and pipeline topology connection data of pumps at each pumping station; constructing a spatiotemporal heterogeneous graph structure including equipment degradation states; analyzing the chaotic characteristics of each pump's vibration signals based on Lyapunov exponents to generate chaotic dynamic feature vectors; multiplying the chaotic synchronization error signal with degradation compensation coefficients to calculate personalized control commands for each pump, achieving flexible synchronous operation; and outputting a dynamically optimized collaborative control strategy based on real-time operation feedback to update graph network parameters and degradation models. This invention solves the problem of efficient collaborative operation of aging pump groups in complex pipeline networks.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and control technology for industrial water treatment, and more specifically, to a method and system for adaptive intelligent regulation and optimization of circulating water based on multimodal sensing. Background Technology

[0002] In the circulating water systems of large petrochemical parks, there are typically multiple pumps distributed across different pumping stations. These pumps have been in operation for anywhere from 2 to 15 years, with significant differences in their performance degradation. These pumping stations are interconnected through a complex pipe network, forming a highly coupled distributed network. During actual operation, due to the varying degrees of degradation among the pumps, their vibration characteristics differ considerably, leading not only to local resonance problems but also to cross-station disturbances propagating through the pipe network. Traditional independent compensation control methods focus only on optimizing individual pump operation, while simple synchronous control strategies ignore the heterogeneity between equipment, failing to simultaneously address both uneven equipment degradation and network coupling issues. This results in low overall operating efficiency of the circulating water system and a continuous accumulation of failure risks. Summary of the Invention

[0003] This invention provides a method and system for adaptive intelligent regulation and optimization of circulating water based on multimodal sensing, addressing system instability caused by uneven equipment degradation and network coupling in related technologies. It also addresses the technical problem of aging pump groups struggling to operate efficiently and collaboratively in complex pipe networks.

[0004] This invention provides a method for adaptive intelligent regulation and optimization of circulating water based on multimodal sensing, comprising:

[0005] Acquire vibration signal data, performance parameter data, and pipe network topology connection data of the circulating water system to generate a spatiotemporal heterogeneous graph structure that includes the equipment degradation state;

[0006] Analyze the historical performance degradation data of each water pump, and calculate the degradation degree index and remaining life prediction value of each water pump;

[0007] The specific method for analyzing the historical performance degradation data of each water pump is as follows: using the Weibull distribution model or the exponential degradation model, and fitting the degradation parameters through the maximum likelihood estimation method;

[0008] The degradation degree index is calculated as the ratio of the current performance value to the initial performance value, with a value range between 0 and 1. The smaller the value, the more severe the degradation.

[0009] The remaining life prediction value is obtained by solving the degradation curve equation. When the degradation degree index drops to the preset failure threshold, the equipment is determined to have reached a failure state.

[0010] Based on phase space reconstruction analysis of the chaotic characteristics of each pump vibration signal, the maximum Lyapunov exponent is calculated to generate dynamic feature vectors characterizing the degree of chaos in each pump vibration.

[0011] The degradation degree index and chaotic dynamic feature vector are used as node features and input into the graph neural network. The mutual influence relationship between nodes is learned through the graph attention mechanism, and the node coupling weight matrix considering the degradation difference is output.

[0012] Using the coupling weight matrix as environmental state information, a distributed reinforcement learning algorithm is used to learn the optimal control strategy and generate differentiated collaborative control parameters for each water pump.

[0013] Based on the chaotic synchronization error and degradation compensation coefficient between each water pump, the personalized control command of each water pump is calculated through the adaptive control law, and the control signal that realizes flexible synchronous operation is output.

[0014] Among them, the degradation degree index is defined as the ratio of the current performance value to the initial performance value, the degradation compensation coefficient is dynamically adjusted according to the degradation degree index, and the chaotic synchronization error is defined as the sum of the weighted state differences between a node and its neighboring nodes.

[0015] Furthermore, the vibration signal data includes vibration acceleration time-series data, vibration spectrum data, and vibration phase data; the performance parameter data includes flow rate data, head data, power data, and efficiency data; the pipeline network topology connection data includes a pipeline connection relationship matrix, pipeline length parameters, and pipeline resistance coefficient.

[0016] After normalizing the vibration signal data and performance parameter data, they are used as attribute features of the graph nodes. When converting the pipeline topology connection data into edge relationships of the graph, a weighted adjacency matrix is ​​constructed, and the edge weights are calculated exponentially based on the ratio of the pipeline length between nodes to the average pipeline length of the system.

[0017] Furthermore, the specific method for analyzing the chaotic characteristics of each pump vibration signal based on phase space reconstruction is as follows:

[0018] Phase space reconstruction is performed using a delayed coordinate embedding method. The embedding dimension is determined by the pseudo nearest neighbor method, and the delay time is determined by the first zero crossing of the autocorrelation function.

[0019] The maximum Lyapunov exponent is calculated using a small data-volume algorithm. The formula is the ratio of the logarithm of the change in the distance between neighboring trajectories in phase space to the time interval.

[0020] When the maximum Lyapunov exponent is greater than zero, it indicates that the system has chaotic characteristics and the neighboring trajectories diverge exponentially.

[0021] Furthermore, the graph neural network adopts a graph attention network architecture, and the attention coefficients are obtained by performing a linear transformation on the node features and then performing exponential normalization after passing them through an activation function;

[0022] The input to the graph neural network is a node feature matrix containing degradation index, remaining lifetime prediction and chaotic dynamic eigenvectors, and the output is a coupling weight matrix representing the strength of mutual influence between nodes.

[0023] During training, the actual vibrational coupling relationships between nodes in historical running data are used as supervision labels. The loss function includes mean squared error loss and graph structure regularization loss to suppress spurious coupling between unconnected nodes.

[0024] Furthermore, the distributed reinforcement learning algorithm employs a multi-agent deep deterministic policy gradient algorithm, configuring an independent Actor network and Critic network for each agent;

[0025] The state space of reinforcement learning includes the degradation degree, vibration state, and coupling weights of each pump; the action space is the speed adjustment of each pump.

[0026] The reward function comprehensively considers system energy consumption, vibration suppression effect, and flow stability. It combines these factors through a weighted summation method, and the optimization objective is to maximize the cumulative discount reward.

[0027] Furthermore, the formula for calculating the chaotic synchronization error is the sum of the product of the difference in state variables between a node and its neighboring nodes that are directly connected in the pipeline topology and the corresponding coupling weight.

[0028] The degradation compensation coefficient is calculated by adding 1 to the product of the compensation intensity coefficient and the difference in degradation degree, i.e. (in This is the degradation compensation coefficient. For water pumps The degree of normalization degradation, (To compensate for the strength coefficient), used to enhance control over severely degraded equipment;

[0029] The personalized control command is the product of control gain, degradation compensation coefficient, and chaotic synchronization error.

[0030] Furthermore, the method for converting control commands into actual executable pump control signals is as follows: after limiting the control command value through a hyperbolic tangent function, multiply it by the maximum frequency adjustment amplitude, and add the reference operating frequency to obtain the target operating frequency;

[0031] Target frequency values ​​that exceed the safe operating frequency range of the water pumps need to be truncated to boundary values ​​and sent to the frequency converters of each water pump via the communication interface to perform speed adjustment.

[0032] This invention provides a circulating water adaptive intelligent regulation and optimization system based on multimodal sensing, comprising:

[0033] The data acquisition module is used to acquire vibration signal data, performance parameter data, and pipeline topology connection data of the circulating water system.

[0034] The degradation analysis module is used to analyze historical performance degradation data and calculate degradation degree indicators and remaining life prediction values;

[0035] The chaos analysis module is used to analyze the chaotic characteristics of vibration signals and generate chaotic dynamic feature vectors.

[0036] The coupling modeling module is used to learn the coupling weight matrix between devices through a graph neural network.

[0037] The collaborative optimization module is used to generate differentiated collaborative control parameters through distributed reinforcement learning;

[0038] The control execution module is used to calculate personalized control commands and output water pump control signals.

[0039] The beneficial effects of this invention are as follows:

[0040] The adaptive intelligent control and optimization method for circulating water based on multimodal perception provided by this invention uses a degradation-aware graph neural network to model the propagation impact of equipment heterogeneity in the pipeline network, accurately quantifying the interaction relationships between pumps with different degrees of degradation. Through Lyapunov exponential analysis and chaotic synchronous control, it achieves flexible and coordinated operation of equipment with varying vibration characteristics. Furthermore, it utilizes a distributed reinforcement learning algorithm to optimize the collaborative decision-making of the entire network, generating personalized control strategies that adapt to equipment differences. Therefore, it overcomes the system instability factors caused by uneven equipment degradation and network coupling, solving the technical problem of inefficient coordinated operation of aging pump groups in complex pipeline networks. Attached Figure Description

[0041] Figure 1 This is a flowchart of a circulating water adaptive intelligent regulation and optimization method based on multimodal sensing according to the present invention;

[0042] Figure 2 This is a line graph comparing the time-domain waveforms of water pump vibration signals at different degradation levels according to the present invention;

[0043] Figure 3 This is a line graph of the pump performance degradation trajectory prediction curve of the present invention;

[0044] Figure 4 This is a heatmap of the coupling weight matrix considering degradation differences according to the present invention;

[0045] Figure 5This is a bar chart comparing and analyzing the chaotic dynamics characteristics of this invention;

[0046] Figure 6 This is a bar graph comparing the system performance before and after adaptive control according to the present invention.

[0047] Figure 7 This is a diagram showing the topology and coupling relationship of the circulating water system network of the present invention. Detailed Implementation

[0048] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0049] At least one embodiment of the present invention discloses an adaptive intelligent regulation and optimization method for circulating water based on multimodal sensing, such as... Figure 1 As shown, it includes the following steps:

[0050] Step 1: Acquire multi-source heterogeneous data of the circulating water system and generate a spatiotemporal heterogeneous graph structure;

[0051] The vibration signal data, performance parameter data, and pipeline topology connection data of each pumping station are obtained. The vibration signal data and performance parameter data are used as attribute features of graph nodes. The pipeline topology connection data is converted into edge relationships of the graph, generating a spatiotemporal heterogeneous graph structure that includes the equipment degradation state.

[0052] It should be noted that the vibration signal data includes vibration acceleration time series data, vibration spectrum data, and vibration phase data; the performance parameter data includes flow rate data, head data, power data, and efficiency data; and the pipeline network topology connection data includes the pipeline connection relationship matrix, pipeline length parameters, and pipeline resistance coefficient.

[0053] Furthermore, the sampling frequency of the vibration acceleration time series data is set to the range of 1000Hz to 10000Hz, and the sampling duration is no less than 10 seconds; the vibration spectrum data is obtained through Fast Fourier Transform (FFT), with a frequency resolution of no less than 0.1Hz; the value range of the vibration phase data is... Radius is used to extract instantaneous phase information from vibration time-series signals through Hilbert transform.

[0054] Since vibration signal data and performance parameter data have different physical dimensions, they need to be normalized before being used as graph node attribute features. The min-max normalization method is used to uniformly scale all types of data to the range of 0 to 1, eliminating the impact of dimensional differences on subsequent graph structure analysis.

[0055] Furthermore, the specific formula for calculating min-max normalization is as follows:

[0056]

[0057] in, The original data values, and These are the minimum and maximum values ​​of this type of data in historical statistics, respectively. The normalized value has a range of values. .

[0058] Furthermore, the specific method for converting pipeline network topology connection data into graph edge relationships is as follows: First, based on the pipeline connection relationship matrix... Determine the connection relationships between nodes, where Indicates water pump and water pump There are pipe connections between them. Indicate no connection; then construct a weighted adjacency matrix. The formula for calculating edge weights is: ,in For nodes To the node The length of the pipe, The average length of all pipes in the system is used, and the edge weight calculation formula gives nodes that are closer to each other a greater coupling weight; finally, the weighted adjacency matrix is... The edge relationships of the graph are input into the subsequent graph neural network.

[0059] Step 2: Analyze historical performance degradation data and output degradation degree indicators and remaining lifetime prediction values;

[0060] The historical performance degradation data of each water pump was analyzed using a degradation trajectory modeling algorithm. By fitting the performance parameter change curves over time, the real-time degradation degree index and remaining life prediction value of each water pump were calculated.

[0061] It should be noted that the degradation trajectory modeling algorithm adopts the Weibull distribution model or the exponential degradation model, and fits the degradation parameters through the maximum likelihood estimation method. The degradation degree index is defined as the ratio of the current performance value to the initial performance value, and the remaining lifetime prediction value is obtained by extrapolating the degradation curve to the failure threshold.

[0062] Furthermore, the formula for calculating the degree of degradation index is as follows:

[0063]

[0064] in, For water pumps At any moment The degree of degradation index, This is the current performance value. These are the initial performance values. The range of values ​​is The smaller the value, the more severe the degradation.

[0065] Furthermore, the failure threshold is defined as the state when the performance parameter drops to 60% of its initial value.

[0066] That is when When the equipment is determined to have reached a failure state, the remaining life prediction value is given. By solving the equation of the degradation curve get.

[0067] The aforementioned Weibull distribution model takes as input time-series data of historical performance parameters of each pump (including the time sequence of changes in parameters such as efficiency, flow rate, and head) and outputs as shape and scale parameters of the degradation model; the exponential degradation model takes as input time-series data of performance parameters and outputs as degradation rate parameters and initial performance baseline values.

[0068] Furthermore, the formula for the degenerate trajectory of the Weibull distribution model is:

[0069]

[0070] in, For water pumps At any moment Performance values, These are the initial performance values. This is the scale parameter (feature lifetime). For shape parameters, The maximum degradation rate (valued at 0.4, corresponding to a failure threshold of 60%); parameters and The likelihood function is obtained through maximum likelihood estimation. ,in Let Weibull probability density function be used. This represents the number of historical observation points.

[0071] Furthermore, the degradation trajectory formula for the exponential degradation model is:

[0072]

[0073] in, For degradation rate parameter; parameter The data is obtained by performing linear regression after taking the logarithm of the observed data. The calculation formula is as follows: ,in and The first The time and performance values ​​of each observation. and These are the mean values ​​for time and logarithmic performance, respectively.

[0074] In this embodiment of the application, in order to improve the accuracy of degradation modeling, a Bayesian update mechanism can also be introduced to dynamically correct the degradation model parameters using real-time monitoring data.

[0075] The aforementioned Bayesian update mechanism takes the prior parameter distribution of the degenerate model and the real-time monitoring data at the current moment as input, calculates the posterior parameter distribution using the Bayesian formula, and outputs the updated model parameter mean and variance.

[0076] Furthermore, the specific calculation method for Bayesian updates is as follows: assuming a degradation rate parameter... The prior distribution is a normal distribution. ,exist Obtain new performance observations at all times Then, the mean and variance of the posterior distribution are updated using the following formula:

[0077]

[0078] in, This is an estimate of the degradation rate calculated based on current observations. To observe the noise variance (obtained through statistical analysis of historical data, typically taking the value of...), ), the updated and As the prior distribution parameters for the next update.

[0079] Step 3: Analyze the chaotic characteristics of the vibration signal and generate chaotic dynamic feature vectors;

[0080] Based on the Lyapunov exponent analysis, the chaotic characteristics of the vibration signals of each water pump are analyzed. Through phase space reconstruction and maximum Lyapunov exponent calculation, dynamic feature vectors characterizing the degree of chaos in the vibration of each water pump are generated.

[0081] It should be noted that the phase space reconstruction uses a delayed coordinate embedding method, the embedding dimension is determined by the pseudo nearest neighbor method, and the delay time is determined by the first zero crossing of the autocorrelation function; the maximum Lyapunov exponent is calculated using a small data volume algorithm, and the formula is:

[0082] Furthermore, the embedding dimension The range of values ​​is The optimal value is determined iteratively using the false nearest neighbor method, and the iteration stops when the proportion of false nearest neighbors falls below 5%; delay time. The time step corresponding to the first zero crossing of the autocorrelation function is preferred. If the autocorrelation function does not cross zero within the analysis time window, the time step is used when the autocorrelation function decreases to its initial value. The time step corresponding to approximately 0.368 is used as the delay time.

[0083]

[0084] in, The maximum Lyapunov index, Let be the distance between neighboring trajectories in phase space. and These are the start time and the end time, respectively.

[0085] Furthermore, the maximum Lyapunov index The value of can be positive, zero, or negative: when When the system exhibits chaotic characteristics, the neighboring trajectories diverge exponentially; when When , it indicates that the system is in a critical state; when This indicates that the system is stable and the adjacent trajectories converge. In the vibration signal analysis of a circulating water system, under normal operating conditions... The typical range of values ​​is ,when The time was determined to be a strong chaotic oscillation, requiring enhanced control and intervention.

[0086] The aforementioned small-data-volume algorithm takes the reconstructed phase space trajectory matrix and the initial neighboring point pairs as inputs, and outputs the maximum Lyapunov exponent value by tracking the separation velocity of neighboring trajectories in phase space.

[0087] Furthermore, the specific steps for phase space reconstruction are as follows:

[0088] The first step is to analyze the vibration acceleration time series data. Delay time and embedding dimension Construct the delayed coordinate vector, the first The coordinates of the phase space points are:

[0089]

[0090] The second step is to calculate the autocorrelation function. The first step is to determine the delay time corresponding to the first zero crossing; the third step is to determine the candidate embedding dimension. (Incrementing from 2), calculate the nearest neighbor distance for each phase space point. Distance to the second nearest neighbor ,when (threshold) When the value is 15, it is determined to be a false nearest neighbor. The proportion of false nearest neighbors is counted, and the optimal embedding dimension is determined when the proportion is less than 5%.

[0091] Furthermore, the specific steps for calculating the maximum Lyapunov exponent using the small data algorithm are as follows: First, select an initial point in the reconstructed phase space. Search for its neighboring points satisfy (neighborhood radius) (Take 1% of the phase space diameter); second step, calculate the initial distance. After evolution time The distance after The third step is to calculate the local Lyapunov exponent. Fourth, repeat the above process for all point pairs that meet the conditions, and take the average of the local Lyapunov exponents to obtain the maximum Lyapunov exponent. ,in This represents the number of valid point pairs.

[0092] Furthermore, evolutionary time The value of needs to be determined based on the system's dynamic characteristics, and the typical range is 5 to 20 sampling intervals. ( , The sampling interval for vibration data is 1000 Hz. (seconds) If the evolution time is too short, the distance change will not be obvious, affecting the calculation accuracy. If the evolution time is too long, the neighboring trajectories will go out of the linear range, resulting in calculation errors.

[0093] Step 4: Input the degradation degree index and the chaotic dynamics feature vector into the graph neural network, and output the coupling weight matrix;

[0094] The degradation degree index obtained in step 2 and the chaotic dynamic feature vector obtained in step 3 are used as node features and input into the graph neural network. The mutual influence relationship between nodes is learned through the graph attention mechanism, and the inter-node coupling weight matrix considering degradation differences is calculated.

[0095] Before concatenating the degradation index, remaining lifetime prediction, and chaotic dynamics feature vector into node features, it is necessary to normalize the remaining lifetime prediction and chaotic dynamics feature vector. The standardization method is used to convert each feature component into a distribution with a mean of 0 and a standard deviation of 1, thereby eliminating the influence of different feature dimensions.

[0096] Furthermore, the calculation formula for the standardization method is as follows:

[0097]

[0098] in, These are the original eigenvalues. This is the mean of the feature across all nodes. The standard deviation of this feature across all nodes. The standardized feature values ​​follow a distribution with a mean of 0 and a standard deviation of 1.

[0099] Furthermore, the learnable weight matrix The dimension parameters are: The input feature dimensions consist of a 1-dimensional degradation index, a 1-dimensional remaining lifetime prediction, and a 1-dimensional Lyapunov index, totaling [number missing]. ), Output feature dimension (typically takes the value of ) The negative slope parameter of the LeakyReLU activation function is set to 0.2.

[0100] The input layer of the aforementioned graph neural network receives Feature matrix of each node ,in For the number of water pumps, For node features, each node feature is composed of a degradation degree index, a predicted remaining lifetime, and a chaotic dynamics feature vector; the output layer is a coupling weight matrix. , representing the strength of mutual influence between nodes. The training method employs a supervised learning model, using the actual vibrational coupling relationships between nodes in historical operational data as supervisory labels, and updating network parameters through a backpropagation algorithm; the loss function is a combination of mean squared error loss and graph structure regularization loss.

[0101] Furthermore, supervised labels (real coupled weights) The method to obtain ) is as follows:

[0102] First, synchronous vibration time-series data of each water pump during historical periods were collected, and node values ​​were calculated. and nodes Cross-correlation coefficient of vibration signals ,in For water pumps exist The amplitude of vibration at time t. The average vibration amplitude, The number of sampling points is used; then, the absolute value of the cross-correlation coefficient is taken and normalized to obtain the true coupling weight. By collecting historical data from multiple time windows, a sample set of coupling relationships containing different degradation states is constructed for training graph neural networks.

[0103]

[0104] in, For the predicted coupling weights, For the actual coupling weights, This is the set of actual connected edges in the pipeline network topology. The first term is the regularization coefficient, and the second term is used to suppress spurious coupling between non-connected nodes.

[0105] Furthermore, the regularization coefficient The range of values ​​is Typical value A larger value indicates stronger suppression of coupling between non-connected nodes; the coupling weight matrix elements The range of values ​​is , representing a node For nodes The normalization effect strength, Indicates no coupling. This indicates the strongest coupling.

[0106] Furthermore, the training parameters for the graph neural network are set as follows: the learning rate uses the Adam optimizer, with an initial learning rate of 0.001, which decays to 0.95 times the original rate every 50 training epochs; the training batch size is set to 32 samples; the total number of training epochs is set to 200 epochs, and training is stopped early when the validation set loss does not decrease for 20 consecutive epochs; the network parameters are initialized using the Xavier initialization method to ensure the consistency of the output variance of each layer.

[0107] Step 5: Generate differentiated collaborative control parameters based on the coupling weight matrix;

[0108] Using a distributed reinforcement learning algorithm, the coupling weight matrix obtained in step 4 is used as the environmental state information. The optimal control strategy is learned through the multi-agent deep deterministic policy gradient (MADDPG) algorithm, generating differentiated collaborative control parameters for each water pump.

[0109] It should be noted that the state space of reinforcement learning includes the degradation degree, vibration state, and coupling weight of each pump; the action space is the speed adjustment of each pump; and the reward function comprehensively considers system energy consumption, vibration suppression effect, and flow stability.

[0110] Since the degradation degree, vibration state and coupling weight in the state space have different dimensions and numerical ranges, each state component needs to be standardized before being input into the reinforcement learning network, and converted into a distribution with a mean of 0 and a standard deviation of 1.

[0111] Furthermore, state standardization employs the same Z-score standardization method as in step 4, with the calculation formula being: ,in These are the original values ​​of the state components. and These are the mean and standard deviation of the state component in the historical data, respectively.

[0112] The three indicators in the reward function—system energy consumption, vibration suppression effect, and flow stability—have different physical dimensions and need to be normalized separately before being combined into a comprehensive reward value through weighted summation. The weight coefficients are determined based on the importance of each optimization objective in the actual application scenario.

[0113] Furthermore, the formula for calculating the comprehensive reward function is as follows:

[0114]

[0115] in, , and These are the normalized energy consumption reward, vibration suppression reward, and flow stability reward (all normalized to...). scope),

[0116] , and For the weighting coefficients, satisfying and In typical application scenarios, the value is taken as , , .

[0117] Furthermore, energy consumption rewards The calculation method is as follows:

[0118] First, calculate the total system power. ,in For water pumps The real-time power is then obtained through a normalization formula. Convert energy consumption into reward value ( and These are the minimum and maximum total power values ​​from historical statistics, respectively. The negative sign indicates that the lower the power, the higher the reward.

[0119] Furthermore, vibration suppression reward The calculation method is as follows: First, calculate the root mean square of the vibration amplitude of each water pump. Then calculate the average vibration level of the system. Finally, the normalization formula is used. Convert vibration level into reward value ( and These are the minimum and maximum average vibration levels from historical statistics, respectively.

[0120] Furthermore, the time window in the calculation of the root mean square of vibration amplitude Set to be the same as the control cycle, i.e. (Typical values ​​range from 1 to 5 seconds), corresponding to the number of sampling points. ( The sampling frequency is typically 1000 Hz, which allows for real-time reflection of the vibration level within the current control cycle.

[0121] Furthermore, traffic stability rewards The calculation method is as follows:

[0122] First, calculate the total system traffic. ,in For water pumps Real-time traffic, then calculate traffic deviation. ( (for the target flow), finally normalized using the formula Convert traffic deviation into reward value ( (The maximum acceptable traffic deviation) The smaller the traffic deviation, the higher the reward.

[0123] Furthermore, the specific method for reflecting the time dimension in the optimization objective of reinforcement learning is as follows: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] At any moment The optimization objective is to maximize the cumulative discount reward starting from the current moment. The objective function is defined as:

[0124]

[0125] in, For intelligent agents strategy The corresponding long-term optimization goal, For intelligent agents At any moment Instant rewards received This is the discount factor (typically 0.99). To plan the time range (typically 100 time steps). This is a multi-agent joint strategy; the cumulative discount reward objective function reflects the cumulative optimization of future rewards over time, with a discount factor... Its purpose is to make the weight of recent rewards greater than that of long-term rewards, so as to avoid the optimization objective from diverging.

[0126] Furthermore, the constraints on the optimization objective include: action constraints. ,in The maximum single-step speed adjustment range (typically 5Hz) is set to ensure smooth control action; state constraints require a deviation in the total system flow rate. ( To ensure flow stability, the maximum allowable flow deviation is set at 5% of the target flow rate; vibration constraints require the root mean square of vibration for each pump. ( The vibration safety threshold is determined according to the equipment specifications to ensure the safe operation of the equipment.

[0127] The aforementioned multi-agent deep deterministic policy gradient algorithm configures an independent Actor network and Critic network for each agent. The Actor network adopts a three-layer fully connected neural network structure, with the first layer receiving the agent's input. Local state (Including the pump's degradation level, vibration state, and neighbor coupling weights), the output is passed through the ReLU activation function to the second hidden layer (256 dimensions), the second hidden layer is then passed through the ReLU activation function to the third hidden layer (128 dimensions), and finally the action is output through the Tanh activation function. (i.e., speed regulation); the Critic network uses a four-layer fully connected neural network structure.

[0128] The first layer receives the joint state of all agents. Joint actions of all intelligent agents The concatenated vector is then passed through three hidden layers (512, 256, and 128 dimensions respectively) after passing through the ReLU activation function, and finally output as a scalar. Indicates the intelligent agent Action value estimation. The training method adopts an asynchronous training mode. The Actor network of each agent is updated through the policy gradient method, and the optimization policy is to maximize the expected reward estimated by the Critic network; the Critic network is updated through temporal difference learning, and the optimization policy is to minimize the Bellman error.

[0129] Furthermore, the local state dimension The components are: a 1-dimensional degradation degree index, a 1-dimensional root mean square vibration amplitude, and coupling weights with neighboring nodes. Dimension (number of neighbors), therefore For a typical pipeline topology (each node connects to an average of 3 neighbors). Approximately 5; Action Dimensions , representing the speed adjustment of a single water pump; the dimension of the input concatenation vector of the Critic network is . .

[0130] The loss function of the Actor network is:

[0131]

[0132] in, For intelligent agents Actor network, This serves as an experience replay buffer; the loss function of the Critic network is:

[0133]

[0134] in, For the target Q value, and The target networks are Critic and Actor, respectively. As a discount factor, For intelligent agents Instant rewards.

[0135] In this embodiment of the application, in order to accelerate the convergence of reinforcement learning, experience replay and target network techniques can be used, and a priority sampling mechanism can be introduced to improve sample utilization efficiency.

[0136] The aforementioned experience replay input is the state transition tuple generated by the interaction between the agent and the environment. Tuples are stored in a fixed-capacity replay buffer. During training, batches of data are randomly sampled from the buffer, and the output is used as training sample batches for network updates. The target network technique generates a target network by copying the parameters of the Actor and Critic networks. The target network parameters slowly track the main network parameters using a soft update method. The operation formula is as follows: ,in Main network parameters, For the target network parameters, This is for updating coefficients.

[0137] Furthermore, update coefficients The range of values ​​is Typical value to Between these values, a smaller value indicates a slower update speed for the target network and higher training stability; the replay buffer capacity is set to... There are 1 state transition tuple, and the batch sampling quantity is set to 256.

[0138] The aforementioned priority sampling mechanism takes the temporal differential error of each sample in the playback buffer as input, and calculates the sampling priority based on the error magnitude. (in For the sample The timing difference error, To avoid zero probability (using small constants), the sampling probability is proportional to the priority. The power of the power is used to output a batch of sampled samples with importance weights.

[0139] Furthermore, the parameters for priority sampling are set as: small constants. Priority index The range of values ​​is Typical value ,when Degenerates into uniform sampling when Time is full priority sampling; discount factor The range of values ​​is Typical value .

[0140] Furthermore, the training parameters for the reinforcement learning network were set as follows: the learning rate for both the Actor and Critic networks was set to 0.0001, and the Adam optimizer was used; training employed an online learning mode, with network parameters updated after each environment interaction in each control cycle; the exploration strategy used Ornstein-Uhlenbeck noise, with noise intensity parameters... , The noise intensity decreases linearly to 10% of its initial value as training progresses; the total number of training steps is set to... The training is considered convergent when the average reward over 1000 consecutive steps reaches the target threshold (set to 90% of the overall reward).

[0141] Step 6: Calculate the personalized control command and output the water pump control signal;

[0142] The chaotic synchronization error signal between each water pump is multiplied by the corresponding degradation compensation coefficient, and the personalized control command of each water pump is calculated by the adaptive control law, and the control signal that realizes flexible synchronous operation is output.

[0143] It should be noted that the chaotic synchronization error is defined as:

[0144]

[0145] in, For water pumps Synchronization error, For water pumps State variables, The coupling weights are used; the degradation compensation coefficient is dynamically adjusted according to the degradation degree index, and the compensated control law is as follows:

[0146] Furthermore, neighbor set Defined as being related to water pumps in the pipe network topology. The set of all directly connected pump nodes, with the number of neighboring nodes denoted as . The range of values ​​is ,in The total number of water pumps in the system; state variables Use water pump The vibration velocity or vibration displacement signal, after normalization, has a value range of [value missing]. Synchronization error The physical meaning of water pump The weighted average state deviation of the pumps from its neighboring pumps.

[0147] Furthermore, the calculation and execution of control commands adopt a fixed control cycle. Typical values ​​range from 1 to 5 seconds, meaning every [time period]. The system status is reacquired, synchronization error is calculated, control commands are updated and sent to the frequency converter. The selection of the control cycle needs to balance the system response speed and control stability. Too short a control cycle may cause control oscillation, while too long a control cycle will reduce the system's ability to respond to disturbances.

[0148]

[0149] in, For control commands, To control the gain, This is the degradation compensation coefficient.

[0150] The aforementioned adaptive control law input is the chaotic synchronization error of each water pump. The degradation degree index and the collaborative control parameters obtained in step 5 are used to dynamically calculate the degradation compensation coefficient. (in For water pumps The degree of normalization degradation, (To compensate for the strength coefficient), the output is a personalized speed control command for each water pump.

[0151] Furthermore, control gain The range of values ​​is The value is determined based on the dynamic response characteristics of the water pump through trial and error or the Ziegler-Nichols tuning method, with a typical value being [value to be filled in]. Compensation strength coefficient The range of values ​​is Typical value ,when Degradation compensation is not performed when When full compensation is achieved, the degradation compensation coefficient is... The range of values ​​is The larger the value, the stronger the control over the degraded equipment.

[0152] Control commands It needs to be converted into an actual executable water pump control signal. The specific conversion method is as follows: map the control command value to the frequency setting range of the frequency converter, convert the control command into the corresponding motor operating frequency value through linear transformation, and then send it to the frequency converter controller of each water pump through the communication interface to perform speed regulation.

[0153] Furthermore, the linear transformation formula from control command to frequency is as follows:

[0154]

[0155] in, For water pumps Target operating frequency, The base operating frequency (determined based on system traffic requirements, with a range of values) is [value missing]. Hz), This is the maximum frequency adjustment range (typically 10Hz). The hyperbolic tangent function is used to limit the amplitude of control commands to... Range, target operating frequency The range of values ​​is Hz, frequency values ​​that exceed the range must be truncated to the boundary value to protect the motor safety.

[0156] The circulating water system of a large petrochemical park consists of 5 centrifugal pumps distributed across 3 pumping stations (Pumping station A includes pumps P1 and P2, pumping station B includes pumps P3 and P4, and pumping station C includes pump P5), interconnected by 6 main pipelines to form a complex network. The pumps have been in operation for 2, 5, 8, 11, and 15 years respectively, exhibiting significant differences in their degree of degradation. The system's target flow rate is... For a flow rate of m³ / h, it is necessary to achieve coordinated operation of heterogeneous equipment and suppress vibration propagation while ensuring stable flow rate.

[0157] In step 1, vibration signal data and performance parameter data of each water pump are acquired. Table 1 shows the basic information and initial operating status of each water pump.

[0158] Table 1. Basic information and initial operating status of each water pump:

[0159]

[0160] Vibration acceleration signals from each water pump were collected, with a sampling frequency of 5000Hz and a sampling duration of 10 seconds. Table 2 shows the statistical characteristic values ​​of the vibration signals (normalized to the range of 0 to 1).

[0161] Table 2 Statistical characteristics of vibration signals (normalized):

[0162]

[0163] Based on the pipeline network topology, a pipeline connection matrix and a weighted adjacency matrix are constructed. The pipeline connections are: P1-P2 (pipe length 150m), P1-P3 (pipe length 280m), P2-P4 (pipe length 320m), P3-P4 (pipe length 180m), P3-P5 (pipe length 240m), P4-P5 (pipe length 200m). The average pipeline length of the system is calculated. m, the weighted adjacency matrix is ​​shown in Table 3.

[0164] Table 3 Weighted Adjacency Matrix :

[0165]

[0166] Figure 2 The vibration acceleration time-domain signals of five water pumps with different service lives were displayed within a 10-second sampling window.

[0167] Figure 7 The topological connections and coupling strength of the five water pumps in the pipeline network are shown.

[0168] In step 2, the degradation trajectory is fitted using a Weibull distribution model based on historical performance data. Table 4 shows the degradation model parameters, the calculated degradation degree indices, and the predicted remaining lifetime.

[0169] Table 4 Degradation analysis results:

[0170]

[0171] Figure 3 This paper demonstrates the prediction of performance degradation trajectories for five water pumps based on the Weibull distribution model.

[0172] In step 3, the vibration signal is reconstructed in phase space and the maximum Lyapunov exponent is calculated. The optimal embedding dimension of each pump is determined by the pseudo nearest neighbor method, the delay time is determined by the autocorrelation function, and the Lyapunov exponent is calculated using a small data-volume algorithm. Table 5 shows the results of the chaotic dynamics analysis.

[0173] Table 5. Results of Chaotic Dynamics Characteristic Analysis:

[0174]

[0175] Figure 5 The chaotic dynamic characteristic parameters of five water pumps were compared, including the maximum Lyapunov exponent, embedding dimension, and root mean square value of vibration.

[0176] In step 4, the degradation level index, the predicted remaining lifetime, and the Lyapunov index are used as node features and input into the graph attention network. After training, the graph neural network outputs a coupling weight matrix that considers the degradation differences, as shown in Table 6.

[0177] Table 6. Coupling weight matrix considering degradation differences :

[0178]

[0179] Figure 4 The coupling weight matrix between the five water pumps, learned by the Graph Attention Network (GAT), is shown.

[0180] In step 5, the coupling weight matrix obtained in step 4, the degradation degree of each pump, and the vibration state are used as the state space for reinforcement learning. The MADDPG algorithm is then used to optimize the cooperative operation strategy of each pump. After training, the MADDPG algorithm outputs differentiated control parameters that consider global cooperative optimization, including the reference operating frequency of each pump. (P1 to P5 are 48.5Hz, 47.2Hz, 46.8Hz, 45.1Hz, and 44.3Hz respectively) and control gain (Set to 1.0 uniformly).

[0181] In step 6, based on the cooperative control parameters obtained in step 5, the chaotic synchronization error and degradation compensation coefficient between each pump are calculated, and personalized control commands are generated through an adaptive control law. First, the synchronization error is calculated based on the vibration state variables. Then calculate the degradation compensation coefficient. (Set the compensation strength coefficient) Then through the control law The control commands are calculated and finally converted using a frequency conversion formula. (set up The target operating frequency is obtained by calculating the control command frequency (Hz). Table 7 shows the calculation and output results of the control command.

[0182] Table 7: Calculation and Output Results of Control Commands

[0183]

[0184] After the control command was executed, the system achieved flexible synchronous operation. Each pump received differentiated frequency regulation according to its own degree of degradation. The severely degraded pumps (P4, P5) received larger positive regulation to compensate for performance loss, while the less degraded pumps (P1) received negative regulation to reduce load. The overall system flow rate remained stable at the target value. At a flow rate of approximately 7985 m³ / h (with an actual total flow rate of 7985 m³ / h and a deviation rate of 0.19%), the average vibration level of the system was reduced by 34.6% compared to traditional synchronous control.

[0185] Figure 6 The control effects of traditional synchronous control and adaptive intelligent control methods were compared.

[0186] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for adaptive intelligent regulation and optimization of circulating water based on multimodal sensing, characterized in that, Includes the following steps: Acquire vibration signal data, performance parameter data, and pipe network topology connection data of the circulating water system to generate a spatiotemporal heterogeneous graph structure that includes the equipment degradation state; Analyze the historical performance degradation data of each water pump, and calculate the degradation degree index and remaining life prediction value of each water pump; The specific method for analyzing the historical performance degradation data of each water pump is as follows: using the Weibull distribution model or the exponential degradation model, and fitting the degradation parameters through the maximum likelihood estimation method; The degradation degree index is calculated as the ratio of the current performance value to the initial performance value, with a value range between 0 and 1. The smaller the value, the more severe the degradation. The remaining life prediction value is obtained by solving the degradation curve equation. When the degradation degree index drops to the preset failure threshold, the equipment is determined to have reached a failure state. Based on phase space reconstruction analysis of the chaotic characteristics of each pump vibration signal, the maximum Lyapunov exponent is calculated to generate dynamic feature vectors characterizing the degree of chaos in each pump vibration. The degradation degree index and chaotic dynamic feature vector are used as node features and input into the graph neural network. The mutual influence relationship between nodes is learned through the graph attention mechanism, and the node coupling weight matrix considering the degradation difference is output. Using the coupling weight matrix as environmental state information, a distributed reinforcement learning algorithm is used to learn the optimal control strategy and generate differentiated collaborative control parameters for each water pump. Based on the chaotic synchronization error and degradation compensation coefficient between each water pump, the personalized control command of each water pump is calculated through the adaptive control law, and the control signal that realizes flexible synchronous operation is output. Among them, the degradation degree index is defined as the ratio of the current performance value to the initial performance value, the degradation compensation coefficient is dynamically adjusted according to the degradation degree index, and the chaotic synchronization error is defined as the sum of the weighted state differences between a node and its neighboring nodes.

2. The adaptive intelligent regulation and optimization method for circulating water based on multimodal sensing according to claim 1, characterized in that, The vibration signal data includes vibration acceleration time-series data, vibration spectrum data, and vibration phase data; the performance parameter data includes flow rate data, head data, power data, and efficiency data; the pipeline network topology connection data includes pipeline connection relationship matrix, pipeline length parameters, and pipeline resistance coefficient. After normalizing the vibration signal data and performance parameter data, they are used as attribute features of the graph nodes. When converting the pipeline topology connection data into edge relationships of the graph, a weighted adjacency matrix is ​​constructed, and the edge weights are calculated exponentially based on the ratio of the pipeline length between nodes to the average pipeline length of the system.

3. The adaptive intelligent regulation and optimization method for circulating water based on multimodal sensing according to claim 1, characterized in that, The specific method for analyzing the chaotic characteristics of vibration signals of various water pumps based on phase space reconstruction is as follows: Phase space reconstruction is performed using a delayed coordinate embedding method. The embedding dimension is determined by the pseudo nearest neighbor method, and the delay time is determined by the first zero crossing of the autocorrelation function. The maximum Lyapunov exponent is calculated using a small data-volume algorithm. The formula is the ratio of the logarithm of the change in the distance between neighboring trajectories in phase space to the time interval. When the maximum Lyapunov exponent is greater than zero, it indicates that the system has chaotic characteristics and the neighboring trajectories diverge exponentially.

4. The adaptive intelligent regulation and optimization method for circulating water based on multimodal sensing according to claim 1, characterized in that, The graph neural network adopts a graph attention network architecture, and the attention coefficients are obtained by linearly transforming the node features and then performing exponential normalization after passing them through an activation function. The input to the graph neural network is a node feature matrix containing degradation index, remaining lifetime prediction and chaotic dynamic eigenvectors, and the output is a coupling weight matrix representing the strength of mutual influence between nodes. During training, the actual vibrational coupling relationships between nodes in historical running data are used as supervision labels. The loss function includes mean squared error loss and graph structure regularization loss to suppress spurious coupling between unconnected nodes.

5. The adaptive intelligent regulation and optimization method for circulating water based on multimodal sensing according to claim 1, characterized in that, The distributed reinforcement learning algorithm employs a multi-agent deep deterministic policy gradient algorithm, configuring an independent Actor network and Critic network for each agent. The state space of reinforcement learning includes the degradation degree, vibration state, and coupling weights of each pump; the action space is the speed adjustment of each pump. The reward function comprehensively considers system energy consumption, vibration suppression effect, and flow stability. It combines these factors through a weighted summation method, and the optimization objective is to maximize the cumulative discount reward.

6. The adaptive intelligent regulation and optimization method for circulating water based on multimodal sensing according to claim 1, characterized in that, The formula for calculating the chaotic synchronization error is the sum of the product of the difference in state variables between a node and its neighboring nodes that are directly connected in the pipeline topology and the corresponding coupling weight. The degradation compensation coefficient is calculated by adding 1 to the product of the compensation intensity coefficient and the difference in degradation degree, i.e. ,in This is the degradation compensation coefficient. For water pumps The degree of normalization degradation, The strength coefficient is used to compensate for the reduced stress and enhance control over severely degraded equipment. The personalized control command is the product of control gain, degradation compensation coefficient, and chaotic synchronization error.

7. The adaptive intelligent regulation and optimization method for circulating water based on multimodal sensing according to any one of claims 1 to 6, characterized in that, The method for converting control commands into actual executable pump control signals is as follows: after limiting the control command value through a hyperbolic tangent function, multiply it by the maximum frequency adjustment amplitude, and add the reference operating frequency to obtain the target operating frequency; Target frequency values ​​that exceed the safe operating frequency range of the water pumps need to be truncated to boundary values ​​and sent to the frequency converters of each water pump via the communication interface to perform speed adjustment.

8. A multimodal sensing-based adaptive intelligent regulation and optimization system for circulating water, used to execute the multimodal sensing-based adaptive intelligent regulation and optimization method for circulating water as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire vibration signal data, performance parameter data, and pipe network topology connection data of the circulating water system. The degradation analysis module is used to analyze historical performance degradation data and calculate degradation degree indicators and remaining life prediction values; The chaos analysis module is used to analyze the chaotic characteristics of vibration signals and generate chaotic dynamic feature vectors. The coupling modeling module is used to learn the coupling weight matrix between devices through a graph neural network. The collaborative optimization module is used to generate differentiated collaborative control parameters through distributed reinforcement learning; The control execution module is used to calculate personalized control commands and output water pump control signals.

Images