Cooling tower cleaning cycle decision-making method based on filler layer resistance characteristic online identification

Abstract

The invention provides a cooling tower cleaning period decision-making method based on filler layer resistance characteristic online identification, which comprises the following steps: acquiring multi-dimensional time sequence data such as differential pressure, water temperature, conductivity, turbidity and microbial concentration based on a multi-source sensor, and constructing a structured trend characteristic matrix through time synchronization, abnormal value elimination and statistical characteristic extraction; in combination with expert knowledge, constructing a multi-modal combined pollution knowledge graph, and learning pollution factor interaction weights by using a graph neural network to realize causal reasoning of combined pollution mechanisms such as scaling, biological slime and the like; according to the method, an abnormal causal chain path is efficiently matched in a knowledge graph by adopting improved A * search, causes and evolution stages are output, automatic attribution and intelligent cleaning suggestion optimization are realized, and the resistance abnormity diagnosis accuracy and the adaptability of a cleaning strategy are improved by continuously feeding back diagnosis results and dynamically expanding a knowledge graph structure.

Description

Technical Field

[0001] This invention relates to the field of cooling tower operation and maintenance optimization and intelligent diagnosis technology, and in particular to a cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics. Background Technology

[0002] In the current field of cooling tower operation and maintenance optimization and intelligent diagnosis, monitoring and anomaly attribution analysis of packing layer resistance characteristics has become one of the core technological directions for improving the energy efficiency of cooling tower operation and the intelligence of maintenance decisions. Mainstream industry solutions generally rely on industrial sensor networks to collect online monitoring data such as differential pressure, water temperature, conductivity, turbidity, and microbial concentration in the packing layer. This data is then combined with historical cleaning records and water quality test reports to model and analyze the contamination status of the cooling tower packing layer. In recent years, intelligent diagnostic technologies based on semantic reasoning, such as knowledge graphs, have been gradually introduced to assist in the causal analysis of resistance anomalies and the decision-making process for cleaning cycles, driving the industry's shift from traditional experience-based judgment to a combination of data-driven and knowledge-driven approaches. However, with the increasing complexity of industrial cooling tower operating environments, actual contamination of the packing layer is often not dominated by a single mechanism, but rather manifests as the interaction and coupling of multiple pollutants such as calcium carbonate, iron oxides, and organic biofilm. Under such complex and abnormal operating conditions, the causal attribution ability of a single pollutant factor is insufficient, and the isolated causal chains in the knowledge graph are insufficient to fully cover the nonlinear and dynamic evolution process formed by the synergistic effects of multiple pollutants, leading to ambiguity or decreased accuracy in attribution analysis results. For example, in the state of "calcium carbonate scaling superimposed with extracellular polymer encapsulation, iron oxide deposition, and coexistence of organic particles," existing models usually cannot characterize the mechanism by which their synergistic effects enhance pressure loss, or may omit important causal path nodes, thus affecting the credibility and scientific validity of diagnostic decisions. Furthermore, the static topology of traditional knowledge graphs makes it difficult for them to evolve with new operational experience. This results in the system's inability to expand the causal network in a timely manner when encountering new pollution patterns or unknown complex scenarios, leading to knowledge blind spots and slow response times. At the path reasoning level, existing methods often ignore the dynamic weight adjustment of the interaction intensity of different pollution factors, failing to provide fine-grained characterization of the linkage characteristics of multiple data sources and the stages of pollution evolution, thus limiting the depth and accuracy of attribution analysis. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, this invention provides a cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics.

[0004] The technical solution of this invention is implemented as follows: a cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics, comprising: S1: Collect multi-dimensional time-series data on differential pressure, water temperature, conductivity, turbidity, and microbial concentration during the operation of the cooling tower packing layer, and simultaneously record historical cleaning events and water quality test reports to form a structured operation dataset. S2: Perform time alignment and outlier removal preprocessing on the collected multidimensional time series data, extract the resistance growth trend feature vector based on the sliding window method, and generate a standardized trend feature matrix; S3: Construct an abnormal coupling node set based on the initial knowledge graph. The abnormal coupling node is used to characterize the coexistence state of multiple pollutants such as scaling and bio-sludge, and associates a dynamic weight parameter set with each coupling node. S4: Input the labeled composite fault cases into the graph neural network for training, learn the intensity parameters of the interaction enhancement effect of different pollutants, and update the dynamic weight configuration of each abnormal coupling node in the knowledge graph. S5: When the differential pressure index of the packing layer is detected to be rising continuously and accompanied by changes in specific water quality parameters, the abnormal pattern matching process is initiated to map the current trend feature vector to the candidate coupling path set in the knowledge graph. S6: Use the improved A* search algorithm to traverse all possible causal chain paths in the knowledge graph, calculate the confidence score of each path, and output the most likely combination of compound pollution causes and its evolution stage judgment results. S7: Each manually confirmed diagnosis result is injected into the training set as a new sample, triggering the knowledge graph structure update process and dynamically expanding the set of abnormally coupled nodes and their associated weight parameters. S8: Generate cleaning cycle optimization suggestions based on diagnostic results and feed the suggestions back to the cooling tower operation and maintenance control system to adjust the cleaning execution strategy and update the cleaning event records.

[0005] The cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics provided by this invention has the following beneficial effects: (1) Through the knowledge graph dynamic expansion mechanism of abnormally coupled nodes, this invention can systematically express and reason about the causal enhancement effect under multiple pollution complex states, realize the accurate location of the resistance increase caused by complex pollution, and greatly improve the accuracy and reliability of system diagnosis. (2) This invention introduces multi-dimensional time-series sensor data of differential pressure, water temperature, conductivity, turbidity, and microbial concentration, as well as historical cleaning events and water quality test reports. It adopts multiple processing methods such as sliding window and data cleaning and standardization to construct a trend feature matrix that reflects the actual operating status of the system in real time. When the resistance is continuously rising and specific water quality parameters change in a coordinated manner, it automatically matches various composite coupling paths in the knowledge graph, which effectively improves the real-time identification capability of resistance anomalies and the environmental adaptability under complex working conditions. Compared with traditional static threshold or expert experience rule systems, the diagnostic response speed and reasoning ability are greatly improved. (3) This invention innovatively uses a graph neural network (GNN) to train historical samples of compound pollution, automatically learns the interaction enhancement weights between different pollution coupling nodes, and dynamically updates the contribution of each node to the resistance anomaly under specific conditions. This mechanism effectively solves the problem of limited applicability of traditional rule bases or static weight models in nonlinear, multi-factor enhancement processes, and has higher attribution confidence. (4) This invention employs an improved A* search algorithm to prioritize the traversal of coupled paths in the knowledge graph, and combines multiple reliability assessments such as weights and historical case similarity, significantly improving the efficiency of causal reasoning and the speed of path reliability judgment. Even as the size of the knowledge graph dynamically expands with the increase of compound contamination cases, the system's reasoning response remains highly efficient and stable. This not only avoids the high computational resource consumption of full brute-force traversal but also effectively avoids problems such as missed detections and misjudgments. Attached Figure Description

[0006] Figure 1 This is a flowchart of the cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to the present invention; Figure 2 This is a sub-flowchart of the cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to the present invention; Figure 3 This is another sub-flowchart of the cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics, specifically including: S1: Collect multi-dimensional time-series data on differential pressure, water temperature, conductivity, turbidity, and microbial concentration during the operation of the cooling tower packing layer, and simultaneously record historical cleaning events and water quality test reports to form a structured operation dataset. S2: Perform time alignment and outlier removal preprocessing on the collected multidimensional time series data, extract the resistance growth trend feature vector based on the sliding window method, and generate a standardized trend feature matrix; S3: Construct an abnormal coupling node set based on the initial knowledge graph. The abnormal coupling node is used to characterize the coexistence state of multiple pollutants such as scaling and bio-sludge, and associates a dynamic weight parameter set with each coupling node. S4: Input the labeled composite fault cases into the graph neural network for training, learn the intensity parameters of the interaction enhancement effect of different pollutants, and update the dynamic weight configuration of each abnormal coupling node in the knowledge graph. S5: When the differential pressure index of the packing layer is detected to be rising continuously and accompanied by changes in specific water quality parameters, the abnormal pattern matching process is initiated to map the current trend feature vector to the candidate coupling path set in the knowledge graph. S6: Use the improved A* search algorithm to traverse all possible causal chain paths in the knowledge graph, calculate the confidence score of each path, and output the most likely combination of compound pollution causes and its evolution stage judgment results. S7: Each manually confirmed diagnosis result is injected into the training set as a new sample, triggering the knowledge graph structure update process and dynamically expanding the set of abnormally coupled nodes and their associated weight parameters. S8: Generate cleaning cycle optimization suggestions based on diagnostic results and feed the suggestions back to the cooling tower operation and maintenance control system to adjust the cleaning execution strategy and update the cleaning event records.

[0010] Step S1: Collect multi-dimensional time-series data on differential pressure, water temperature, conductivity, turbidity, and microbial concentration during the operation of the cooling tower packing layer, and simultaneously record historical cleaning events and water quality test reports to form a structured operational dataset. Specifically, this includes: S1.1: Based on industrial sensor networks, acquire multimodal monitoring data such as differential pressure signal, water temperature signal, conductivity signal, turbidity signal and microbial concentration signal during the operation of cooling tower packing layer, so as to realize real-time perception of the operating status of packing layer; The input conditions are real-time signals collected by differential pressure sensors, water temperature sensors, conductivity sensors, optical turbidity sensors and microbial concentration detection units during the operation of the cooling tower packing layer, and the transmission path is a wired / wireless sensor network deployed in the industrial site. A multi-node synchronous acquisition strategy is adopted (parameters: acquisition period τ=1s, network time synchronization protocol is IEEE1588) to unify the time base of various types of sensors and buffer the raw signals according to the device number and acquisition channel. Furthermore, through the analog signal conditioning module (parameters: filter type is second-order low-pass, cutoff frequency...), =50Hz), to achieve pre-filtering processing of the output of analog sensors such as differential pressure and water temperature, and to obtain a set of voltage signals with stable amplitude and no high-frequency interference; Furthermore, through a 12-bit analog-to-digital converter (parameter: sampling rate) =1kHz, input voltage range 0~5V), to realize the digitization of filtered analog signals and generate corresponding numerical monitoring data streams; Furthermore, through the RS-485 bus communication protocol (parameters: baud rate 9600bps, frame structure 8 data bits + 1 stop bit + no parity bit), the data from each digital acquisition unit is uploaded to the central data acquisition server, and raw multimodal data packets containing sensor identifiers and timestamps are generated. Through the above data acquisition link, the field measurement results obtained in the previous step are transformed into structured raw signal data of differential pressure, water temperature, conductivity, turbidity and microbial concentration, so as to realize real-time perception of all elements of the cooling tower packing layer operation status. For example, in an industrial circulating cooling system, the differential pressure sensor is configured with a range of 0~100kPa, an accuracy of ±0.25%FS, and a sampling period τ=1s; the water temperature sensor has a range of 0~100°C and an accuracy of ±0.1°C; the conductivity sensor has a range of 0~20mS / cm and an accuracy of ±0.5%FS; the optical turbidity sensor has a range of 0~1000NTU and an accuracy of ±2%FS; and the microbial concentration detection unit is based on the ATP bioluminescence method with a response time ≤5s. All sensors are connected to the signal conditioning module via shielded twisted-pair cables, with a low-pass filter cutoff frequency. =50Hz. The acquired signal is converted by a 12-bit ADC, sampling rate =1kHz, RS-485 bus uplink to the central acquisition server. The server synchronizes time with each acquisition terminal according to the IEEE 1588 protocol to ensure that the timestamp deviation of different types of signals is less than 1ms. In one operating cycle, the differential pressure data fluctuates within the range of [15.6, 18.9] kPa, the water temperature range is [32.4, 34.0] °C, the conductivity range is [4.21, 4.38] mS / cm, the turbidity range is [185, 230] NTU, and the microbial concentration range is [2.5×10^4, 3.1×10^4] CFU / mL. The multimodal raw data packets acquired through the above link are used for subsequent timestamp alignment and trend feature extraction processing, effectively ensuring full data coverage and high time synchronization accuracy. S1.2: Perform timestamp alignment processing on the acquired multimodal monitoring data to eliminate time offset caused by differences in sampling frequencies of different sensors and generate a multidimensional time-series data stream under a unified time reference; S1.3: Based on the sliding window algorithm, the aligned multidimensional time series data is segmented and the statistical features of each dimension of data within the time window are extracted, including mean, variance, maximum value, minimum value and rate of change, in order to construct a preliminary trend feature vector; The input is multi-dimensional time-series data of the cooling tower packing layer after timestamp alignment, including differential pressure, water temperature, conductivity, turbidity and microbial concentration parameters. The data type is a sequence of real number arrays under a unified time base. The sliding window segmentation method is adopted (parameters: window length (N hours), sliding step size). This enables the local truncation processing of multidimensional time-series data to obtain a set of local data fragments under time continuity. Furthermore, the formula for calculating the mean within a window is as follows: in, The mean of the window. For the i-th sampled value within the window, The number of sampling points is used to quantify the average level of each parameter and output the window mean feature group. Furthermore, through variance calculation methods (formula: in, Standard deviation, The window mean is used to measure data volatility and obtain the window variance feature set. Furthermore, by using an extreme value extraction algorithm (parameters: maximum value search, minimum value search), the peak and valley values ​​of each dimension parameter within the window period are detected, generating a set of maximum and minimum value features; Furthermore, the linear rate of change calculation method (formula: in, The rate of change is linear. and These are the end and start data values ​​within the window, respectively. and (for the corresponding timestamp), the trend slope of the parameter within the window is quantified to obtain the rate of change feature group; Through the above multi-parameter statistical feature extraction chain, the windowed time series data is transformed into a preliminary trend feature vector containing mean, variance, extreme values ​​and rate of change, thereby realizing a local trend quantitative expression of the operating status of the packing layer. For example, during the monitoring of an industrial cooling tower, a sliding window length of 4 hours and a step size of 30 minutes were set. Data on differential pressure (kPa), water temperature (°C), conductivity (μS / cm), turbidity (NTU), and microbial concentration (CFU / mL) after timestamp alignment were segmented. Within a certain detection window, the differential pressure sampling sequence length was 480 points, and the mean was calculated to obtain... kPa, variance is kPa², with a maximum value of kPa, minimum value kPa, the rate of change r is calculated using the formula. kPa / h. Average water temperature is ℃, variance is ℃², rate of change ℃ / h; mean conductivity is μS / cm, variance is μS² / cm²; mean turbidity was μS² / cm²; NTU, variance is NTU²; mean microbial concentration is CFU / mL. The feature vector of this window is stored in a trend feature database for subsequent historical cleansing event data alignment and online knowledge graph inference input; S1.4: Structure and label the key parameters in the historical cleaning event records and water quality test reports, and align the data based on timestamps and trend feature vectors to form a composite data sample containing cleaning event tags and water quality anomaly markers; S1.5: Perform standardization and normalization on the aligned composite data samples to eliminate the influence of different physical dimensions on subsequent modeling and generate a structured running dataset that can be used for knowledge graph training and inference.

[0011] Step S2: Perform time alignment and outlier removal preprocessing on the collected multidimensional time-series data, extract the resistance growth trend feature vector based on the sliding window method, and generate a standardized trend feature matrix. Specifically, this includes: S2.1: Perform timestamp alignment processing on the multi-dimensional time-series data of differential pressure, water temperature, conductivity, turbidity and microbial concentration collected during the operation of the cooling tower packing layer to eliminate the data asynchrony problem caused by the difference in the acquisition frequency of each sensor and obtain a time-synchronized multi-source monitoring dataset. S2.2: Based on the time-synchronized multi-source monitoring dataset, an outlier detection method based on Z-score is used to identify and remove outliers in each dimension of the data to eliminate the negative impact of sensor drift or transient interference on subsequent feature extraction and obtain cleaned multi-dimensional time-series data. S2.3: Perform sliding time window slicing processing on the cleaned multidimensional time series data, construct a time window sequence based on the window length N hours and the sliding step size Δt, in order to capture the dynamic trend characteristics of the change in the resistance of the packing layer and generate a set of windowed time series data segments; S2.4: Perform feature extraction on each time window data in the windowed time series data segment set, and construct feature vectors using statistical indicators such as mean, variance, maximum value, minimum value and linear regression slope to quantify the changing trend of each parameter within the window period and obtain the original trend feature vector set; The input data is a set of windowed time-series data segments generated in step S2.3. Each segment contains multi-dimensional synchronous monitoring data such as differential pressure, water temperature, conductivity, turbidity, and microbial concentration. The arithmetic mean calculation method (parameter: time series length L of each dimension) is used to quantify the central trend of each monitoring parameter within the current window and form mean feature components. Furthermore, through variance calculation methods (parameter: mean) The total number of samples (L) is used to characterize the dispersion of data in each dimension and obtain the volatility characteristic components. Furthermore, an extreme value search algorithm (parameter: window sequence) is used to extract the maximum and minimum values ​​of each monitoring parameter, construct extreme value feature components, and capture peak or valley information in the operating state. Furthermore, the least squares linear regression algorithm (parameters: time index t, parameter values) is used. ), calculate the slope of change of each dimension of data within the window period, and generate trend direction and rate feature components; Furthermore, by employing the feature vector concatenation method, the mean, variance, extreme values, and slope components are sequentially combined into the original trend feature vector, thereby achieving a full characterization of the changes in multidimensional parameters within the window period. Through the above feature extraction algorithm chain, the windowed time series data fragments obtained in the previous step are transformed into a set of original trend feature vectors, so as to realize a quantitative feature description of the operating status of the filler layer on the time segment scale. For example, setting the window length to 6 hours and the sliding step Δt to 1 hour, the differential pressure data series (unit: kPa) within a certain window yields a mean of 2.35, a variance of 0.12, a maximum value of 2.58, and a minimum value of 2.14. The slope is calculated using the least squares method, resulting in a slope of 0.021 kPa / h. The water temperature series within the same window yields a mean of 30.5℃, a variance of 0.8, a maximum value of 31.0℃, and a minimum value of 29.8℃, with a slope of 0.05℃ / h. Vector concatenation of the parameter results forms the original trend feature vector [2.35, 0.12, 2.58, 2.14, 0.021, 30.5, 0.8, 31.0, 29.8, 0.05, ...]. Verification shows that this vector effectively reflects the changing trends of the packing layer resistance and related water quality parameters within the window period, providing accurate input data for the standardization of S2.5. S2.5: Perform standardization and normalization on each dimension of the original trend feature vector set, and use the Min-Max scaling method to map each feature to the [0,1] interval to eliminate the influence of different physical dimensions on subsequent knowledge graph path matching, and generate a standardized trend feature matrix as input data for the abnormal pattern matching module.

[0012] Step S3: Construct a set of abnormal coupling nodes based on the initial knowledge graph. These abnormal coupling nodes characterize the coexistence of multiple pollutants such as scaling and biofilm, and each coupling node is associated with a set of dynamic weight parameters. For example... Figure 2 As shown, it specifically includes: S3.1: Based on the historical operation data of the cooling tower packing layer and expert experience, typical composite pollution patterns are extracted, including combinations such as the coexistence of scaling and biological slime, and the coupling of particle deposition and organic matter adhesion. A feature set of composite pollution patterns is generated to construct the semantic basis of abnormal coupling nodes. S3.2: For each pollution combination in the feature set of compound pollution patterns, define abnormal coupling nodes based on the terminology of the industrial control field, add node entities in the initial knowledge graph in a graph structuring manner, and use standardized pollution factor combination naming rules for node naming to achieve a formal expression of compound pollution state. The input condition is the set of features of the composite pollution pattern generated in S3.1, which includes typical combination types such as coexistence of scaling and biological slime, and coupling of particulate deposition and organic matter attachment, and has pollution factor combination information that can be used as a basis for node naming. An industrial control terminology parsing method (parameters: domain terminology library Version=3.4, semantic matching threshold=0.85) is used to map the names of pollutants in the feature set of composite pollution patterns to standard terms in control engineering, thereby obtaining semantically consistent pollution combination expressions. Furthermore, the pollution combination standardization algorithm (parameter: naming rule template ID=CT-FILL-COUPLE-01) is used to perform structured naming standardization on the mapped pollution combination and generate node naming strings, such as "calcium carbonate-iron oxide-extracellular polymer composite deposition", to ensure that the same pollution pattern in data from different sources has a unique name; Furthermore, a graph-structured node generation method is adopted (parameters: node type = abnormally coupled node, node attribute fields = list of contaminants, combination type, source data index) to embed standardized named contaminant combinations into the initial knowledge graph structure and generate corresponding node entity objects; Furthermore, a bidirectional index link between node objects and pollution factor combinations is achieved through a node attribute binding algorithm (parameter: binding field set = {name string, pollution factor list}) to support subsequent adjacent node traversal and causal path construction based on factor queries; By standardizing node construction and naming, the features of the compound pollution pattern in the previous step are transformed into knowledge graph nodes with unique names, semantic consistency and structured attributes, realizing a formal expression of the compound pollution state and providing bindable objects for the extraction of environmental conditions in S3.3. For example, in an industrial circulating cooling system, the feature set of composite pollution patterns includes the type "CaCO3 scaling and EPS biofilm adhesion coupling," where CaCO3 represents calcium carbonate scaling and EPS represents extracellular polymeric substances (EPS). After mapping using a control domain terminology library, the semantic expression "calcium carbonate-extracellular polymeric scaling-slime coupling" is obtained. The naming rule template CT-FILL-COUPLE-01 is called to convert it into the standard node name "calcium carbonate-iron oxide-extracellular polymeric composite deposition," where iron oxide is added to the combination name because its content exceeds the limit as shown by regular water testing. The node generation method inserts this name as the unique identifier of the node into the initial knowledge graph. The node attributes include the list of pollutants {CaCO3, Fe2O3, EPS}, the combination type "inorganic-organic coupling," and the source data index ID=00482. An attribute binding algorithm establishes a bidirectional index between the node and the list of pollutants, through... Implement index tracing. During the verification phase, when querying the "extracellular polymeric factor," the system can directly locate the coupling node and display its connection relationship with the differential pressure anomaly path in the graph view, ensuring the traceability and naming uniqueness of the causal chain construction; S3.3: Based on the multimodal monitoring data collected during the operation of the cooling tower, extract the environmental condition characteristics corresponding to each composite pollution mode, including parameters such as water temperature, turbidity, conductivity and microbial concentration, to generate the initial configuration of the dynamic weight parameter set, so as to reflect the intensity difference of the interaction of pollutants under different operating conditions. Based on multimodal monitoring data such as water temperature, turbidity, conductivity and microbial concentration collected during the operation of the cooling tower, the feature mapping analysis method (parameters: monitoring data matrix M, set of composite pollution modes C) is used to separate the set of environmental condition features that are highly correlated with each composite pollution mode from the real-time operation data. Furthermore, thresholds are selected using multidimensional statistical regression methods (parameters: linear regression, stepwise regression). This allows for the estimation of weights for various environmental conditions and the generation of an initial coefficient vector describing the influence of water temperature, turbidity, conductivity, and microbial concentration on the intensity of interactions between pollutants. Furthermore, principal component analysis (parameter: cumulative contribution rate threshold 90%) is used to reduce the dimensionality of the weight coefficient vector and generate initial projection values ​​of dynamic weight parameters that retain the maximum variance information in the low-dimensional space. Furthermore, by using a normalization mapping algorithm (parameters: Min-Max scaling, interval [0,1]), the initial projection values ​​of the dimensionality-reduced weight parameters are standardized, and a set of dimensionless dynamic weight parameters that can be directly used for the initialization of knowledge graph node parameters is generated. By extracting environmental features and standardizing weights, the results of the previous step in defining the compound pollution pattern are transformed into dynamic weight parameter data that reflect the differences in the intensity of the interaction between pollutants under different operating conditions, thereby realizing the initialization of the operating status features of abnormally coupled nodes in the knowledge graph. For example, during the long-term operation of an industrial cooling tower, the average differential pressure was 4.5 kPa, the water temperature range was 25℃-35℃, the turbidity range was 60 NTU-120 NTU, the conductivity was 1800 μS / cm-2400 μS / cm, and the microbial concentration was 1500 cfu / mL-3500 cfu / mL. N=500 time-series samples were collected and input into a feature mapping analysis method. The correlation coefficient between water temperature and the scaling-biosludge pattern was 0.82, and the correlation coefficient between turbidity and the pattern was 0.88. Multidimensional statistical regression was used to calculate the weighting coefficients, using the following formula: in, Let be the weighting coefficient of the j-th type of environmental condition characteristics. Let j be the standardized value of the j-th feature in sample i. The total number of samples is given. Principal component analysis yielded the contribution rates of the first two principal components, which were mapped to the initial projection values ​​of low-dimensional dynamic weight parameters [0.73, 0.64]. After processing with Min-Max normalization (interval [0, 1]), a standardized parameter set [0.95, 0.84] was obtained, which was then bound to the corresponding "calcium carbonate-extracellular polymer" node in the knowledge graph. During the verification period, the Pearson correlation coefficient between the weight change curve of this node and the differential pressure growth trend under different seasonal conditions remained above 0.91, significantly improving the accuracy of the attribution analysis of complex pollution. S3.4: For each abnormally coupled node, based on the causal reasoning model in the field of industrial control, establish a causal mapping relationship between it and the change of packing layer resistance, and construct a causal path between the resistance growth trend and the state of complex pollution, so as to enhance the causal explanation ability of complex faults in the knowledge graph. S3.5: Based on the dynamic weight parameter set and causal path model, a graph neural network algorithm is used to initialize and train the parameters of the abnormally coupled nodes to obtain the dynamic weight configuration corresponding to each coupled node. The configuration is then associated and bound with the corresponding coupled node to support the path confidence assessment in subsequent online attribution analysis. Based on the established dynamic weight parameter set and causal path model, a graph neural network algorithm is adopted (parameter: node embedding dimension). =64, training batch =128, learning rate lr=0.001, number of iterations epoch=500), to achieve parameter initialization training for abnormally coupled nodes; Furthermore, a node feature encoding algorithm (parameters: standardized vectors of environmental condition features, including water temperature, conductivity, turbidity, and microbial concentration) is used to realize a structured representation of the input layer features of each anomaly-coupled node, and the initial node state matrix is ​​obtained. ; Furthermore, a message passing mechanism based on adjacency matrix is ​​adopted (parameter: constructing adjacency matrix A using the connection relationships between composite contaminants in the graph, and introducing a dynamic weight parameter set during message propagation). This enables the modeling of interactions between pollutants at different nodes and generates an intermediate state matrix. ; Furthermore, the node state is iteratively optimized using a graph convolution update formula, which is: in, This is the updated node state matrix. This is the current node's state matrix. It is an adjacency matrix. For trainable weight matrix, The ReLU activation function is used, and this formula enables node state updates based on topology and dynamic weights. Furthermore, the mean squared error loss function is used to optimize the parameters of the node output. The loss function is: in, This is the mean squared error loss value. For the sample size, The actual weight label of the node. The W parameters are iteratively updated using gradient descent to minimize the loss, given the predicted value. By using the output of the trained graph neural network, the dynamic weight configurations corresponding to each abnormal coupling node are associated and bound with the node ID, thereby realizing the initial values ​​of the parameters required for path confidence assessment in subsequent online attribution analysis. By using the above graph neural network initialization training method, the causal path and environmental condition features of the previous step are transformed into interpretable dynamic weight vectors, which can accurately quantify the interaction intensity of abnormally coupled nodes under different working conditions and improve the reasoning credibility of knowledge graphs under complex compound pollution modes. For example, in an air conditioning cooling water system operation scenario, the node embedding dimension is selected. The adjacency matrix A is 64, constructed from the co-occurrence relationships of scaling factors and biofilm factors, with an edge weight of 0.75 for scaling-biofilm and 0.60 for scaling-particle deposition. The input layer environmental conditions are water temperature (0.65), conductivity (0.72), turbidity (0.58), and microbial concentration (0.80), all values ​​normalized to Min-Max. Node state updates use ReLU activation; after 500 iterations, the MSE loss decreases from the initial 0.024 to 0.003. The final output dynamic weight configuration for an anomalous coupled node, "CaCO3-Fe2O3-EPS composite deposition," is 0.82, corresponding to the path confidence calculation formula in online attribution analysis. in, =0.91 represents the similarity of trend features. =0.82 is the dynamic weight. =0.88 represents the support of historical samples, and the calculated confidence level C=0.65776 significantly improves the accuracy and reliability of attributing complex pollution.

[0013] Step S4: The labeled composite fault cases are input into a graph neural network for training to learn the intensity parameters of the interaction enhancement effect of different pollutants and update the dynamic weight configuration of each abnormal coupling node in the knowledge graph. Figure 3 As shown, it specifically includes: S4.1: Perform feature encoding on the labeled composite fault cases to convert multidimensional operational data and anomaly attribution labels into structured input vectors that can be processed by graph neural networks; S4.2: Based on the node embedding mechanism of graph neural network, the initial representation learning of abnormally coupled nodes in the knowledge graph is performed to generate a low-dimensional semantic embedding vector containing the interaction relationship of pollution factors. Based on the structured input vector processed by feature encoding, a graph neural network node embedding mechanism (parameters: node type label, edge type label, initial feature matrix) is used to initialize the vectorized representation of abnormally coupled nodes in the knowledge graph; Furthermore, a local neighborhood subgraph is generated on the knowledge graph structure using a random walk sampling algorithm (parameters: sampling length L, sampling number K), which explicitly encodes the associated context structure of nodes into the sampling sequence to capture the adjacency relationship between pollutants. Furthermore, the Skip-gram embedding training method (window size W, number of negative samples N) is used to perform context prediction modeling on the sampled neighborhood sequence to achieve low-dimensional vector clustering of semantically similar nodes; Furthermore, by using a weighted normalized adjacency matrix diffusion method (the weight matrix consists of node degree and edge type weights), the semantic features and structural features of nodes are fused to generate a preliminary set of embedding vectors. Furthermore, a spectral domain graph convolution algorithm (convolution order p, feature dimension d) is used to perform smoothing filtering on the initial embedding vector in the Laplacian matrix feature space, preserving highly relevant features and suppressing noise, to obtain the final low-dimensional semantic embedding vector; By using the node embedding mechanism described above, the feature encoding results of the previous step are transformed into low-dimensional vectors that can simultaneously represent the interaction relationship between pollutants and the topology of the knowledge graph, thus achieving the expected technical effect of providing structured input for subsequent attention message passing and interaction modeling. S4.3: An attention-based message passing algorithm is used to perform information propagation and interaction modeling between pollutants in a graph neural network in order to identify causal enhancement effects under compound pollution patterns. S4.4: Normalize the pollution interaction intensity parameters output by the graph neural network to generate an interpretable dynamic weight configuration, which is used to characterize the influence of different composite pollution types on the change of packing layer resistance. S4.5: Update the learned dynamic weight configuration to the corresponding abnormal coupling nodes in the knowledge graph to enhance the knowledge graph's causal reasoning ability and path completeness for complex resistance anomalies.

[0014] Step S5: When a continuous increase in the differential pressure index of the packing layer is detected, accompanied by a linked change in specific water quality parameters, the abnormal pattern matching process is initiated to map the current trend feature vector to the candidate coupling path set in the knowledge graph. Specifically, this includes: S5.1: Perform sliding window mean filtering on the real-time data sequence output by the differential pressure sensor of the cooling tower packing layer to eliminate high-frequency noise interference and obtain a smoothed differential pressure change trend sequence, which is used to determine whether the resistance shows a continuous upward trend. S5.2: Based on the multidimensional parameters of water temperature, conductivity, turbidity and microbial concentration obtained by the water quality monitoring system, the Pearson correlation coefficient matrix is ​​used to analyze the linear correlation between them and the differential pressure change trend, and to identify the set of water quality linkage parameters that are significantly related to the increase in resistance. S5.3: Perform Z-score standardization on the identified set of water quality linkage parameters to eliminate the differences in the dimensions of different parameters, and extract the feature pattern vector under the current operating conditions based on the time series clustering algorithm to characterize the current operating status; S5.4: Input the standardized trend feature vector into the feature matching layer of the multimodal anomaly coupled knowledge graph, use the cosine similarity algorithm to calculate its matching degree with each candidate coupling path in the knowledge graph, and filter out the set of candidate paths with matching degree higher than the set threshold. S5.5: Perform path feasibility assessment on the selected candidate coupling path set, calculate its confidence score under the current environmental conditions based on the dynamic weight parameters of each path in the knowledge graph, and generate an initial path priority list for subsequent causal chain search. The path feasibility evaluation algorithm is applied to the candidate coupling path set output from step S5.4 (parameters: candidate path set size M, dynamic weight parameter matrix). Environmental condition feature vector This enables reliable measurement of each path under the current operating state of the environment; Furthermore, through a dynamic weight parameter mapping method (parameter: node weight) Edge weight The environmental condition correction coefficient β is used to adapt the overall path weight to environmental conditions, and the path environmental weighting coefficient matrix is ​​obtained. ; Furthermore, a confidence score formula was used to calculate the quantitative evaluation value for each path. The formula is: in, Let be the dynamic weight of the i-th node. This is a correction factor based on environmental conditions. This represents the path length. Furthermore, a priority generation method based on threshold filtering is used (parameter: confidence threshold). Rating list To retain only those with confidence scores higher than [a certain threshold] The set of paths, and generate an initial path priority list. ; The path feasibility assessment algorithm transforms the candidate coupled path set from the previous step into path priority data optimized and sorted according to the current operating conditions, thereby providing high-confidence, low-search-space initial input for the causal chain search in step S6. For example, in a cooling tower operation and maintenance system, suppose the size M of the candidate coupling path set is 5, and the dynamic weight parameter matrix... The weights of each node range from 0.6 to 0.9, and the environmental condition feature vector... The parameters included water temperature of 28℃, conductivity of 2.3 mS / cm, turbidity of 4.5 NTU, and microbial concentration of 120 CFU / mL. An environmental weighting coefficient matrix was calculated using a dynamic weighted parameter mapping method. The scaling nodes were weighted at 0.85 with a correction coefficient β = 0.95, and the biofilm nodes were weighted at 0.78 with a correction coefficient β = 0.92. The scores for the five paths were calculated using the confidence scoring formula, with the highest score reaching 0.87. A confidence threshold was then set. The initial score was set to 0.8. After filtering, three paths with scores of 0.87, 0.85, and 0.81 were retained, forming an initial path priority list sorted in descending order of score. And serve as input for the S6 step to improve the A* search.

[0015] Step S6: An improved A* search algorithm is used to traverse all possible causal chain paths in the knowledge graph, calculate the confidence score of each path, and output the most likely combination of compound pollution causes and its evolutionary stage judgment. Specifically, this includes: S6.1: Based on the current differential pressure trend feature vector of the packing layer and the abnormal coupling nodes in the knowledge graph, semantic matching is performed to generate a set of candidate causal paths to determine possible combinations of pollution causes, which serve as the initial search space for the A* search algorithm; Based on the initial path priority list obtained by S5.5 processing and the current differential pressure trend feature vector of the packing layer, a semantic matching algorithm (parameter: bidirectional matching based on embedding vector similarity) is used to perform vector space mapping between the feature vector and the abnormally coupled nodes in the knowledge graph, thereby locking the node candidate set in the high-dimensional semantic space. Furthermore, by using the node relationship parsing method (parameters: node attribute constraints and edge type filtering rules), the causal relationship between candidate nodes can be quickly filtered, and a set of primary path chains that meet the current working condition constraints can be obtained. Furthermore, a structured path generation method is adopted (parameter: path length limit). (constrained by the diversity of pollutants), the primary path chain set is constructed to generate a candidate causal path set with complete causal start and end definitions; Furthermore, by combining the temporal patterns of feature vectors with the composite pollution evolution history curves stored in the knowledge graph, pattern matching is performed (algorithm: Dynamic Time Warping DTW) to calculate the matching score of candidate paths, and the matching score is used as one of the path heuristic parameters in the subsequent search space. Furthermore, a path semantic consistency verification method is adopted (parameter: edge type semantic consistency threshold). Invalid paths caused by node attribute conflicts or incorrect causal chain directions are eliminated to form a final set of candidate causal paths that meet the input requirements of search A. Through the above semantic matching and path filtering processing methods, the initial path priority list of S5.5 is transformed into candidate causal path data with high semantic relevance, structural consistency and historical working condition consistency, so as to achieve the efficient initial search space construction effect of A search. For example, under the operating conditions of a certain cooling tower system, the current differential pressure trend feature vector has a dimension of 5, including the normalized average differential pressure (0.78), average water temperature (0.65), average conductivity (0.72), average turbidity (0.81), and average microbial concentration (0.69). The node embedding vector also has a dimension of 5, and cosine similarity is used for similarity calculation. Under this operating condition, semantic matching calculation yields a similarity of 0.92 with the node "calcium carbonate-iron oxide-extracellular polymeric composite deposition," which is higher than the set threshold of 0.85, thus entering the candidate node set. Edge type filtering is performed on the candidate node set, retaining edges with a water quality parameter path correlation > 0.7, constructing two primary path chains. Path length constraints are then applied. =4 and the constraint of pollutant diversity (containing at least two different types of pollutants), three candidate causal paths that meet the conditions are obtained. In DTO pattern matching, the matching distance between candidate path 1 and the historical curve is 0.15, path 2 is 0.21, and path 3 is 0.35. Path 1 and path 2 are retained in the final set according to the matching priority. After semantic consistency verification, an edge with an incorrect causal chain direction in path 2 is removed, and finally only path 1 enters the S6.2 extension stage; S6.2: Perform path expansion operation on each path in the candidate causal path set, and combine the dynamic weight parameters between pollutants in the knowledge graph to generate a candidate sequence of causal chains after path expansion, so as to simulate the evolution process of compound pollution. S6.3: Each candidate causal chain path is evaluated based on an improved heuristic function, which integrates path length, dynamic weight coefficient and historical diagnostic similarity factor to improve search efficiency and attribution accuracy; S6.4: An improved A* search algorithm is used to traverse and score all candidate causal chain paths, and a confidence score is calculated for each path. The confidence score is generated based on a weighted average of path matching degree, dynamic weight contribution degree and historical sample support. The input conditions are the candidate sequences of causal chains evaluated by S6.3 and their corresponding dynamic weight parameters, in which the path matching degree, weight contribution degree and historical sample support have been formed into a structured numerical matrix. An improved A* search algorithm (parameters: initial state = candidate path set, heuristic function = comprehensive path length, dynamic weight coefficient and historical diagnostic similarity factor) is adopted to achieve traversal and heuristic scoring of all candidate causal chain paths in the knowledge graph; Furthermore, by using the path matching degree calculation method (input: current trend feature vector and path feature vector, algorithm: cosine similarity), the matching degree of each causal chain is quantified, and a matching degree index matrix is ​​obtained; Furthermore, by using the dynamic weight contribution calculation formula, the weight parameters of nodes in each path are weighted and summed according to the intensity of interaction to characterize the contribution intensity of the combination of pollution factors to the resistance trend. Furthermore, the frequency of a path appearing in historical records is quantified by using a historical sample support evaluation method (input: historical composite fault sample set, algorithm: sample matching frequency statistics). The three types of indicators mentioned above are combined into a path confidence score using a weighted combination formula. The formula is defined as: in, For path matching degree, For dynamic weight contribution, For historical sample support, , , The weighting coefficients are adjusted based on expert experience and model validation. By improving the step-by-step expansion of the A* search algorithm and calculating the above-mentioned fusion scoring formula, the candidate path feature data of the previous step is transformed into a confidence score list, thereby realizing a quantitative evaluation of the credibility of all causal chain paths under the current working conditions. For example, in an industrial cooling tower operation scenario, the candidate path set has 50 paths, the initial dynamic weight parameter range is [0.3, 0.8], and the historical sample database contains 500 composite fault cases. The parameter configuration is as follows: =0.4、 =0.35、 =0.25. During the matching degree calculation, the cosine similarity threshold was set to 0.85, and paths with a matching degree below this threshold were removed. The dynamic weight contribution is obtained by summing the node weights and dividing by the number of nodes, with an average value of approximately 0.55. Historical sample support statistics show that the highest support path occurred at a frequency of 0.72. In the fusion calculation, a certain path had a matching degree of 0.9, a weight contribution of 0.6, and historical support of 0.7. Substituting these values ​​into the formula... The confidence score obtained was approximately 0.755. This score ranked first in the ranking and was used in S6.5 for the final attribution output. The validation results show that this step can effectively improve the diagnostic reliability and reduce the false positive rate under complex contamination conditions. S6.5: Sort the confidence scores of all paths, select the causal chain path with the highest confidence as the final attribution result, and output the judgment result of the combination of complex pollution causes and its evolution stage as the input basis for the cleaning cycle decision module.

[0016] Step S7: Each manually confirmed diagnosis result is injected into the training set as a new sample, triggering the knowledge graph structure update process and dynamically expanding the set of abnormally coupled nodes and their associated weight parameters. Specifically, this includes: S7.1: Perform structured analysis on the manually confirmed diagnostic results, extract the composite pollution types, abnormal coupling nodes and their associated weight parameter change trends, in order to generate standardized composite fault sample descriptions. S7.2: Based on the generated standardized composite fault sample description, perform incremental update operation on the graph neural network training dataset, inject new samples into the existing training set to expand the sample space and enhance the model's ability to learn composite contamination patterns. S7.3: Based on the updated training set, the graph neural network model is incrementally trained to optimize the connection weights between abnormally coupled nodes, thereby improving the completeness of the expression of causal paths and the accuracy of reasoning in the knowledge graph. S7.4: Based on the output of the graph neural network after incremental training, perform dynamic expansion operation of the knowledge graph structure, add or merge abnormal coupling nodes of compound pollution type, and update their corresponding dynamic weight parameter set. Based on the output dataset of the graph neural network after incremental training, a node structure parsing algorithm (parameters: node embedding vector set, node confidence parameter threshold) is used to realize the structured parsing and classification of the existing knowledge graph node set and generate a list of nodes to be expanded. Furthermore, through the composite pollution pattern similarity calculation method (parameters: embedding vector cosine similarity, pattern feature weight coefficient), semantic clustering analysis of existing abnormally coupled nodes is realized, and the node clusters to be merged and their aggregation feature descriptions are obtained; Furthermore, through the node merging and addition rule engine (parameters: pollution type naming convention, node relationship constraints), the system performs a structure merging operation on nodes with similarity exceeding a set threshold in the clustering results, and performs a node addition operation on pollution combinations that do not appear in the graph but have been verified by new samples, generating a node set with updated structure. Furthermore, through a dynamic weight parameter reconfiguration algorithm (parameters: interaction strength value of training output, environmental condition correction coefficient), the weight parameters associated with newly added or merged nodes are replaced and calibrated, and a weight configuration matrix is ​​generated for subsequent causal path confidence calculation. Furthermore, by using the graph structure index update method (parameters: node identifier mapping table, relation edge weight table), the node set and weight configuration matrix are synchronously written into the knowledge graph storage structure, so that the latest topology structure can be directly called by the runtime inference engine; By using node structure parsing and dynamic weight parameter reconfiguration, the incremental training output of the previous step is transformed into an expanded node set and updated weight configuration of the knowledge graph, thereby enhancing the expressive power of the composite pollution pattern and improving the inference accuracy. For example, in an industrial circulating cooling system, the output of the incrementally trained graph neural network contains 12 newly added composite contamination pattern embedding vectors, each with a dimension of 128, corresponding to different combinations of contaminants. The node structure parsing algorithm sets the node confidence parameter threshold to 0.85, selecting 8 high-confidence nodes for the expansion process. The composite contamination pattern similarity calculation uses the embedding vector cosine similarity method, with a similarity threshold set to 0.92. The results show that 3 pairs of nodes meet the merging conditions. The node merging and new rule engine, based on the contaminant naming rules, identifies "CaCO3-Fe2O3-EPS composite deposition" and "calcium carbonate-iron oxide-extracellular polymer composite deposition" as the same pattern and merges them. The dynamic weight parameter reconfiguration uses the following calculation formula: in, This is an environmental condition correction factor, with a value of 0.95. To train the output interaction strength matrix, the updated weight matrix is ​​calculated. The interaction coefficients of various pollution factors were significantly improved. The graph structure index update method synchronously writes the five newly added nodes and their relation edge weights into the Neo4j database, and the inference engine can directly call these updated contents in the next round of online attribution analysis; S7.5: Perform consistency checks and causal path integrity verification on the expanded knowledge graph to ensure that the newly added abnormal coupling nodes are consistent with the existing graph structure in logic and semantics, so as to improve the system's adaptability to complex working conditions and diagnostic reliability.

[0017] Step S8: Based on the diagnostic results, generate cleaning cycle optimization suggestions and feed them back to the cooling tower operation and maintenance control system to adjust the cleaning execution strategy and update the cleaning event records. Specifically, this includes: S8.1: Based on the results of the combined causes of complex pollution and the judgment of its evolution stage from the attribution analysis of the resistance trend of the packing layer, the fuzzy inference algorithm is used to calculate the sensitivity weight of different pollution types to the cleaning response, so as to generate a priority sequence of cleaning response. S8.2: Perform multi-objective optimization modeling on the cleaning response priority sequence. Inputs include cleaning cost, energy consumption constraints, cooling tower operating load curve and ambient temperature prediction data. Use NSGA-II multi-objective genetic algorithm to solve for the Pareto optimal cleaning cycle set. S8.3: Execute a filtering mechanism based on operation and maintenance strategy rules on the Pareto optimal cleaning cycle set, and generate a unique recommended cleaning cycle by combining the current cooling tower operating status, equipment maintenance window and historical cleaning cycle execution records; S8.4: The recommended cleaning cycle is encapsulated in OPC UA protocol format and transmitted to the cooling tower operation and maintenance control system via industrial Ethernet interface to trigger the cleaning execution module to start the cleaning process; S8.5: After the cleaning is completed, collect the cleaning event execution feedback data, including the actual cleaning time, cleaning water consumption, turbidity index of discharged wastewater and differential pressure recovery value of the packing layer, and update the structured cleaning event record database.

[0018] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0019] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics, characterized in that, Includes the following steps: S1: Collect multi-dimensional time-series data on differential pressure, water temperature, conductivity, turbidity, and microbial concentration during the operation of the cooling tower packing layer, and simultaneously record historical cleaning events and water quality test reports to form a structured operation dataset. S2: Perform time alignment and outlier removal preprocessing on the multidimensional time series data, extract the resistance growth trend feature vector based on the sliding window method, and generate a standardized trend feature matrix; S3: Construct an abnormal coupling node set based on the initial knowledge graph. The abnormal coupling node is used to represent the coexistence state of multiple pollutants, and a dynamic weight parameter set is associated with each coupling node. S4: Input the labeled composite fault cases into the graph neural network for training, learn the intensity parameters of the interaction enhancement effect of different pollutants, and update the dynamic weight configuration of each abnormal coupling node in the knowledge graph. S5: When the differential pressure index of the packing layer is detected to be rising continuously and accompanied by changes in specific water quality parameters, the abnormal pattern matching process is initiated to map the current trend feature vector to the candidate coupling path set in the knowledge graph. S6: The improved A* search algorithm is used to traverse all possible causal chain paths in the knowledge graph, calculate the confidence score of each path, and output the most likely combination of compound pollution causes and its evolution stage judgment results.

2. The cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to claim 1, characterized in that, Following step S6, the following is also included: S7: Each manually confirmed diagnosis result is injected into the training set as a new sample, triggering the knowledge graph structure update process and dynamically expanding the set of abnormally coupled nodes and their associated weight parameters. S8: Generate cleaning cycle optimization suggestions based on diagnostic results, and feed the suggestions back to the cooling tower operation and maintenance control system to adjust the cleaning execution strategy and update the cleaning event record.

3. The cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to claim 1, characterized in that, Step S1 specifically includes: Based on industrial sensor networks, multimodal monitoring data such as differential pressure, water temperature, conductivity, turbidity, and microbial concentration are acquired during the operation of the cooling tower packing layer. The multimodal monitoring data is subjected to timestamp alignment processing to generate a multidimensional time-series data stream under a unified time reference; The multi-dimensional time-series data stream under the unified time benchmark is segmented, and the statistical features of each dimension of data within the time window are extracted to construct a preliminary trend feature vector. The key parameters in historical cleaning event records and water quality test reports are structured and labeled, and the data is aligned based on timestamps and trend feature vectors to form a composite data sample containing cleaning event tags and water quality anomaly markers. The composite data sample is subjected to standardization and normalization processing to generate a structured runtime dataset.

4. The cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to claim 3, characterized in that, The multimodal monitoring data acquisition includes the use of differential pressure sensors, water temperature sensors, conductivity sensors, optical turbidity sensors, and microbial concentration detection units. Each sensor acquires data through a wired or wireless industrial sensor network with a 1-second acquisition cycle. The data is acquired using the IEEE 1588 time synchronization protocol, equipped with a second-order low-pass filter (cutoff frequency 50Hz) and a 12-bit ADC (sampling rate 1kHz, input voltage range 0~5V), and transmitted to the central data acquisition server via an RS-485 bus.

5. The cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to claim 1, characterized in that, Step S2 specifically includes: The time-series data of differential pressure, water temperature, conductivity, turbidity and microbial concentration collected during the operation of the cooling tower packing layer are time-stamped and aligned to obtain a time-synchronized multi-source monitoring dataset. Based on the time-synchronized multi-source monitoring dataset, outlier identification and removal are performed on the data of each dimension to obtain cleaned multi-dimensional time-series data. Perform sliding time window slicing on the cleaned multidimensional time series data to generate a set of windowed time series data fragments; Perform feature extraction on each time window data in the windowed time series data segment set to obtain the original trend feature vector set; Standardization and normalization are performed on each dimension of the original trend feature vector set to map each feature to a unified interval, thereby generating a standardized trend feature matrix.

6. The cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to claim 5, characterized in that, Step S2 further includes segmenting the aligned multidimensional time series data according to a set window length and step size based on the sliding window algorithm, extracting the mean, variance, maximum value, minimum value and trend slope statistical features within the window, and forming a set of trend feature vectors of a unified dimension after all parameters are normalized by Min-Max.

7. The cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to claim 1, characterized in that, Step S3 specifically includes: Based on historical operating data of cooling tower packing layers and expert experience, typical composite pollution patterns are extracted, a feature set of composite pollution patterns is generated, and the semantic basis of abnormal coupling nodes is constructed. For each pollution combination in the feature set of the composite pollution pattern, abnormal coupling nodes are defined based on the terminology of the industrial control field. Node entities are added to the initial knowledge graph in a graph-structured manner, and the node naming adopts the standardized naming rules for pollution factor combinations. Based on the multimodal monitoring data collected during the operation of the cooling tower, the environmental condition characteristics corresponding to each composite pollution mode are extracted and used to generate the initial configuration of the dynamic weight parameter set. For each abnormal coupling node, based on the causal reasoning model in the field of industrial control, a causal mapping relationship between it and the change of packing layer resistance is established, and a causal path model between the resistance growth trend and the state of compound pollution is constructed. Based on the dynamic weight parameter set and the causal path model, a graph neural network algorithm is used to initialize and train the parameters of the abnormally coupled nodes to obtain the dynamic weight configuration corresponding to each coupled node, and the configuration is associated and bound with the corresponding coupled node.

8. The cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to claim 1, characterized in that, Step S4 specifically includes: Feature encoding is performed on the labeled composite fault cases to convert multidimensional operational data and anomaly attribution labels into structured input vectors that can be processed by graph neural networks; Based on the node embedding mechanism of graph neural network, the initial representation learning of abnormally coupled nodes in the knowledge graph is performed to generate a low-dimensional semantic embedding vector containing the interaction relationship of pollution factors. An attention-based message passing algorithm is used to model the information propagation and interaction between pollutants in a graph neural network, and to identify causal enhancement effects under compound pollution patterns. The pollution interaction intensity parameters output by the graph neural network are normalized to generate an interpretable dynamic weight configuration. The dynamic weight configuration is then updated to the corresponding abnormally coupled node in the knowledge graph.

9. The cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to claim 1, characterized in that, Step S5 specifically includes: A sliding window mean filter is applied to the real-time data sequence output by the differential pressure sensor of the cooling tower packing layer to obtain a smoothed differential pressure change trend sequence. Based on the multidimensional parameters obtained from the water quality monitoring system, we analyze their linear correlation with the differential pressure change trend and identify the set of water quality linkage parameters that are significantly related to the increase in resistance. The water quality linkage parameter set is standardized, and the feature pattern vector under the current operating condition is extracted based on the time series clustering algorithm. The standardized trend feature vector is input into the feature matching layer of the multimodal anomaly coupled knowledge graph. The cosine similarity algorithm is used to calculate its matching degree with each candidate coupling path in the knowledge graph, and the set of candidate paths with matching degree higher than the set threshold is selected. Perform path feasibility assessment on the candidate coupling path set, calculate the confidence score of each path under the current environmental conditions based on the dynamic weight parameters of each path in the knowledge graph, and generate an initial path priority list.

10. The cooling tower cleaning cycle decision method based on online identification of packing layer resistance characteristics according to claim 1, characterized in that, The confidence score is generated by weighting path matching degree, dynamic weight contribution degree and historical sample support.

Images