A method and system for multi-line liquid flow control

Abstract

The application provides a multi-pipeline liquid flow control method and system. The method first collects context factors such as multi-pipeline flow, temperature, pressure, valve state, process arrangement, and external events, and constructs a multi-dimensional time series data set; then through multi-scale time window segmentation, the flow, pressure, and temperature time series characteristics at different time scales are extracted, and the transient and trend anomalies are identified through wavelet transform or multi-scale clustering. The weight is dynamically allocated by combining the context analysis algorithm, the weight and the abnormal factor are input into the detection model, and the cause analysis is completed; the multi-pipeline collaborative coupling modeling is used to distinguish single-pipeline abnormality and pipeline coupling abnormality, and to locate external causes and internal faults. According to the abnormal level and the coupling relationship, the intervention instruction is generated in stages, and the actuator is automatically driven to real-time regulate the pipeline parameters. The method can accurately identify and attribute multi-pipeline flow anomalies, reduce wafer defects caused by liquid supply fluctuations through sensitive detection and accurate intervention, and improve the yield of semiconductor production.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and in particular to a multi-pipeline liquid flow control method and system. Background Technology

[0002] In semiconductor manufacturing processes, key steps such as chemical mechanical polishing (CMP), wet cleaning, photoresist coating and development, and electrochemical deposition (ECD) all involve multi-channel, high-precision liquid delivery and distribution systems. These systems typically need to simultaneously supply multiple fluids, such as ultrapure water, chemical reagents, photoresist, polishing slurry, or electroplating solutions, to multiple process chambers or workstations. The stability of the flow rate, pressure, and temperature in each pipeline directly determines the uniformity of the wafer surface reaction, the precise control of feature dimensions, and the final chip yield.

[0003] In existing semiconductor manufacturing processes, technologies for monitoring and controlling multi-pipe flow anomalies generally employ empirical rule-based anomaly detection or fixed threshold warnings. This approach struggles to adapt to the diverse operating conditions across different pipelines caused by variations in medium type, flow state, and ambient temperature and pressure. Furthermore, for typical interactive effects in multi-flow systems (such as an anomaly in one pipeline leading to coordinated anomalies in multiple pipelines), existing systems lack the ability to dynamically capture and coordinately respond to anomaly evolution trends. They often only respond passively after an anomaly occurs, rather than proactively predicting and intervening. Simultaneously, the anomaly detection process has limited ability to distinguish between discrete anomalies and continuous, trend-based anomalies, easily leading to missed or false alarms, impacting the safety and stability of system operation.

[0004] Therefore, it is necessary to improve the multi-channel flow control technology in the semiconductor manufacturing process to overcome the shortcomings of the existing technology. Summary of the Invention

[0005] To overcome the problems existing in related technologies, one of the objectives of this invention is to provide a multi-pipeline liquid flow control method. This method is used in the multi-pipeline flow control of semiconductor manufacturing, and can achieve accurate identification and source attribution of multi-pipeline anomalies. Through highly sensitive anomaly detection and precise intervention, it can effectively reduce wafer defects caused by liquid supply fluctuations and improve production yield.

[0006] A multi-pipeline liquid flow control method, comprising: Flow rate, temperature, pressure, valve status, environmental variables, production line process arrangement, manual operation records and context factors of external events at different locations in multiple pipelines are obtained to construct a multidimensional time series raw dataset; Multi-scale time-series window segmentation is performed on the multi-dimensional time-series raw dataset. Based on the sliding window technique, time-series features of flow fluctuation, pressure change, and temperature shift are extracted at the second, minute, hour, and process cycle scales, respectively. By inputting multi-scale time series features into wavelet transform or multi-scale clustering models, transient anomaly and trend anomaly factors can be obtained to identify sudden and gradual changes in flow and related parameters at various scales. Based on context factor records, the context factor parsing algorithm is used to perform weighted labeling of production line process cycle, external events, and equipment status, and dynamically adjust the context factor weights. The context factor weights are then combined with transient anomaly factors and input into the flow anomaly detection model to achieve multi-dimensional environmental cause analysis. By utilizing the transient anomaly factor and context factor weights output by the flow anomaly detection model, a multi-pipe collaborative coupling modeling method is adopted to analyze flow anomaly events under multi-pipe joint conditions, determine whether the anomaly is a single-path anomaly or a coupled anomaly, and distinguish between external inducement and internal anomaly source based on this. Determine the anomaly level and coupling relationship of the multi-pipeline collaborative coupling model output, and generate graded intervention measures such as shutdown, voltage reduction, flow reduction or early warning based on the dynamic weight of context factors. The actuator is automatically triggered based on the intervention level command to adjust the pipeline flow and system pressure in real time.

[0007] In a preferred embodiment of the present invention, when adjusting the pipeline flow rate and system pressure in real time, the method further includes: Synchronously record changes in environment variables and context factors during execution; The traffic response curves, abnormal chain changes, and context factor weight changes after the intervention are archived in the global feature library, and the multi-scale anomaly detection, context factor analysis, and intervention strategies are continuously optimized through a self-learning algorithm.

[0008] In a preferred embodiment of the present invention, a multidimensional time-series raw dataset is constructed, including: The flow, temperature, and pressure signals collected by the sensors are time-synchronized to eliminate timestamp offsets caused by communication delays and obtain time-aligned raw sensor data sequences. The received valve status, equipment start / stop signals, and manual operation logs are structured and decoded to extract discrete control action events and their occurrence times, generating a standardized set of operation event records.

[0009] In a preferred embodiment of the present invention, constructing a multidimensional time-series raw dataset further includes: obtaining the process task arrangement table of the production line scheduling system, and combining it with maintenance notifications and power fluctuation records received by the external event reporting channel, performing event tagging processing to form an external context event sequence with time-series alignment tags.

[0010] In a preferred embodiment of the present invention, constructing a multidimensional time-series raw dataset includes: spatiotemporally fusing time-aligned sensor data sequences, standardized operation event record sets, and external context event sequences with time-alignment markers, and constructing a multidimensional time-series raw dataset containing physical quantities, operation quantities, and environmental quantities based on a unified time axis.

[0011] In a preferred embodiment of the present invention, multi-scale time-series window segmentation is performed on the multi-dimensional time-series raw dataset. Based on the sliding window technique, time-series features of flow fluctuations, pressure changes, and temperature shifts are extracted at the second, minute, hour, and process cycle scales, respectively, including: Normalize each dimension of the original multidimensional time series dataset, and map the data of different dimensions to the appropriate scales using a min-max scaling algorithm. The interval is used to obtain the normalized time-series input sequence; Four types of sliding windows are defined: second-level, minute-level, hour-level, and process cycle-level. The normalized time series input sequence is sliced ​​in layers to generate a set of subsequences covering different time spans.

[0012] In a preferred embodiment of the present invention, obtaining transient anomaly and trend anomaly factors includes: A non-subsampled dual-tree complex wavelet transform model is constructed to decompose the high-frequency detail components in the multi-scale time-series feature matrix layer by layer to obtain multi-layer wavelet coefficient tensors. Adaptive SURE threshold quantization is then performed on the detail coefficients of each layer to extract significant abrupt change energy segments and generate transient anomaly response maps. A multi-scale hierarchical clustering model based on dynamic time warping distance is constructed to measure the similarity between hourly and process cycle-level feature vectors, identify isolated sample points that deviate from normal clusters, obtain a candidate set of trend drift anomalies, and confirm the trend anomaly factor sequence through slope persistence test and confidence interval boundary crossing analysis.

[0013] In a preferred embodiment of the present invention, the multi-pipeline collaborative coupling modeling method includes: constructing a multi-pipeline topology modeling framework based on a graph neural network, treating each pipeline as a node, establishing directed edges based on physical connection relationships and fluid dynamic coupling strength, generating a pipeline coupling topology graph, and using transient anomaly factors and context factor weights as node attribute inputs, aggregating neighborhood information through graph convolution operations, and generating a node-level coupling state embedding vector for anomaly type discrimination and source attribution.

[0014] In a preferred embodiment of the present invention, the graded intervention measures include: When the severity of the anomaly is extremely high, a shutdown command is triggered; when it is high, a voltage reduction regulation is performed; when it is medium, a flow load reduction is implemented; when it is low, only an early warning command for observation is generated. The severity of the anomaly is generated based on a comprehensive score of the transient anomaly factor amplitude, the trend anomaly factor duration, and the number of coupled pipeline nodes. The self-learning algorithm includes: using the traffic response curve and abnormal chain changes after the intervention as feedback samples to update the normal operating condition benchmark and abnormal pattern template in the global feature library, and using incremental learning or online Bayesian update to adjust the threshold parameters of multi-scale anomaly detection, the weight distribution of context factor parsing, and the decision rules of the intervention strategy.

[0015] The second objective of this application is to provide a multi-pipeline liquid flow control system for implementing the multi-pipeline liquid flow control method described above.

[0016] The beneficial effects of this invention are as follows: This invention provides a multi-pipeline liquid flow control method and system. This method extracts flow rate, pressure, and temperature features at four scales: second, minute, hour, and process cycle. It can simultaneously capture millisecond-level transient pulses (such as water hammer effects caused by rapid valve opening and closing) and hourly / batch-level progressive drift (such as slow filter clogging and performance degradation of grinding fluid pumps). It overcomes the shortcomings of traditional single-threshold or fixed-sampling-frequency methods in distinguishing between process switching disturbances and actual faults. Through multi-scale clustering to identify transient and trend anomalies, it can effectively identify trend drift and automatically separate unknown fault modes without prior labels. It maintains high robustness in complex noise backgrounds, significantly reducing false alarms caused by normal process fluctuations in semiconductor production lines, while avoiding missed detections of slowly deteriorating faults. Furthermore, it incorporates context-based factor recording processing into a unified data system and uses attention mechanisms and online Bayesian updates to dynamically adjust the weights of each factor. The system can understand the current operating conditions, effectively solving the problem that static weights or fixed rules cannot adapt to temporary process changes and external disturbances, greatly improving the rationality of anomaly diagnosis. By treating each pipeline as a node and the fluid dynamic coupling strength as the edge weight, a pipeline coupling topology graph is constructed. Through graph convolution, neighborhood information is aggregated, accurately identifying whether anomalies are confined to a single pipeline or caused by multi-pipeline coupling. Combining external event weights further distinguishes between externally induced events (such as plant disturbances) and internal structural faults. This function overcomes the shortcomings of traditional single-pipeline monitoring in locating anomaly propagation paths, significantly improving fault tracing efficiency and avoiding unnecessary total shutdowns. Attached Figure Description

[0017] Figure 1 This is a flowchart of a multi-pipeline liquid flow control method provided in an embodiment of the present invention. Detailed Implementation

[0018] Preferred embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0019] Example 1 like Figure 1 As shown, this embodiment provides a multi-pipeline liquid flow control method, which includes the following steps: A multi-pipeline liquid flow control method, comprising: S1. Obtain flow rate, temperature, pressure, valve status, environmental variables, production line process arrangement, manual operation records and context factors of external events at different locations in multiple pipelines to construct a multi-dimensional time series raw dataset; S2. Perform multi-scale time-series window segmentation on the multi-dimensional time-series raw dataset, and extract time-series features of flow fluctuation, pressure change, and temperature shift at the second, minute, hour and process cycle scales respectively based on the sliding window technique. S3. Input the multi-scale time series features into the wavelet transform or multi-scale clustering model to obtain transient anomaly and trend anomaly factors, so as to identify the sudden and gradual changes in flow and related parameters at various scales. S4. Based on context factor records, the context factor parsing algorithm is used to perform weighted labeling on production line process cycle, external events, and equipment status, and the context factor weights are dynamically adjusted. The context factor weights and transient anomaly factors are then jointly input into the flow anomaly detection model to achieve multi-dimensional environmental cause analysis. In this application, based on context factor records, a context factor parsing algorithm is used to weight and label production line process cycles, external events, and equipment status in practical applications. The context factor weights are dynamically adjusted, allowing them to be jointly input into the flow anomaly detection model along with transient anomaly factors, thus achieving multi-dimensional environmental cause analysis. This process continuously updates the influence strength of each factor through an online learning mechanism, effectively overcoming the misjudgment problem caused by static weight settings. This enables the system to adaptively understand unconventional operations such as temporary maintenance and material switching.

[0020] S5. Using the transient anomaly factor and context factor weights output by the flow anomaly detection model, a multi-pipeline collaborative coupling modeling method is adopted to analyze flow anomaly events under multi-pipeline joint state, determine whether the anomaly is a single-path anomaly or a coupled anomaly, and distinguish between external inducement and internal anomaly source accordingly. S6. Determine the anomaly level and coupling relationship of the multi-pipeline collaborative coupling model output, and generate graded intervention measures such as shutdown, pressure reduction, flow reduction or early warning based on the dynamic weight of context factors; automatically trigger actuators based on intervention level instructions to adjust pipeline flow and system pressure in real time.

[0021] Even better, when adjusting pipeline flow and system pressure in real time, it also includes: S7. Synchronously record changes in environmental variables and context factors during the execution process; archive the traffic response curve, abnormal chain changes, and context factor weight changes after the intervention into the global feature library, and continuously optimize multi-scale anomaly detection, context factor analysis, and intervention strategies through self-learning algorithms.

[0022] Specifically, step S1 includes: The flow, temperature, and pressure signals collected by sensors deployed in a multi-pipeline system are time-synchronized to eliminate timestamp offsets caused by communication delays and obtain time-aligned raw sensor data sequences.

[0023] The valve status, equipment start / stop signals, and manual operation logs are structured and decoded to extract discrete control action events and their occurrence times, generating a standardized operation event record set.

[0024] Obtain the process task schedule from the production line scheduling system, combine it with maintenance notices and power fluctuation records, perform event tagging, and form an external context event sequence with time-sequence aligned markers.

[0025] The time-aligned sensor data, operation event set, and external context sequence are fused in a spatiotemporal dimension to construct a multi-dimensional time-series raw dataset containing physical quantities, operation quantities, and environmental quantities based on a unified time axis.

[0026] Furthermore, the construction of the multidimensional time-series raw dataset includes: The flow, temperature, and pressure signals collected by the sensors are time-synchronized to eliminate timestamp offsets caused by communication delays and obtain time-aligned raw sensor data sequences. The received valve status, equipment start / stop signals, and manual operation logs are structured and decoded to extract discrete control action events and their occurrence times, generating a standardized set of operation event records.

[0027] Obtain the process task schedule from the production line scheduling system, and combine it with maintenance notifications and power fluctuation records received from the external event reporting channel. Perform event tagging processing to form an external context event sequence with time-series alignment markers.

[0028] It also includes: spatiotemporal fusion of time-aligned sensor data sequences, standardized operation event record sets, and external context event sequences with time-aligned markers, and construction of a multidimensional time-series raw dataset containing physical quantities, operation quantities, and environmental quantities based on a unified time axis.

[0029] This application constructs a multi-dimensional time-series raw dataset, enabling multi-source context fusion in the application of this method. This overcomes the limitations of traditional methods that only monitor physical quantities, forming a three-dimensional perception system encompassing physics, operation, and environment, resulting in more comprehensive anomaly identification. Time synchronization processing eliminates communication delay deviations, ensuring time-series alignment of data from multiple nodes and improving the accuracy of subsequent coupled analysis. Structured decoding and event annotation standardize discrete operations and external events, facilitating the algorithm's automatic understanding of the operating context.

[0030] Spatiotemporal fusion constructs a unified timeline dataset, providing a high-quality, multi-dimensional data foundation for multi-scale analysis and anomaly tracing.

[0031] Furthermore, the multi-scale time-series window segmentation process is performed on the original multi-dimensional time-series dataset, and time-series features of flow fluctuations, pressure changes, and temperature shifts are extracted at the second, minute, hour, and process cycle scales based on the sliding window technique, including: Normalize each dimension of the original multidimensional time series dataset, and map the data of different dimensions to the appropriate scales using a min-max scaling algorithm. The interval is used to obtain the normalized time-series input sequence; Four types of sliding windows are defined: second-level, minute-level, hour-level, and process cycle-level. The normalized time series input sequence is sliced ​​in layers to generate a set of subsequences covering different time spans.

[0032] Furthermore, the acquisition of transient anomaly and trend anomaly factors includes: A non-subsampled dual-tree complex wavelet transform model is constructed to decompose the high-frequency detail components in the multi-scale time-series feature matrix layer by layer to obtain multi-layer wavelet coefficient tensors. Adaptive SURE threshold quantization is then performed on the detail coefficients of each layer to extract significant abrupt change energy segments and generate transient anomaly response maps. A multi-scale hierarchical clustering model based on dynamic time warping distance is constructed to measure the similarity between hourly and process cycle-level feature vectors, identify isolated sample points that deviate from normal clusters, obtain a candidate set of trend drift anomalies, and confirm the trend anomaly factor sequence through slope persistence test and confidence interval boundary crossing analysis.

[0033] Furthermore, the multi-pipeline collaborative coupling modeling method includes: constructing a multi-pipeline topology modeling framework based on graph neural networks, treating each pipeline as a node, establishing directed edges based on physical connection relationships and fluid dynamic coupling strength, generating a pipeline coupling topology graph, and using transient anomaly factors and context factor weights as node attribute inputs, aggregating neighborhood information through graph convolution operations, and generating node-level coupling state embedding vectors for anomaly type discrimination and source attribution.

[0034] This application determines the anomaly level and coupling relationship output by a multi-pipeline collaborative coupling model, and generates intervention level instructions based on the dynamic weight of context factors. These instructions include graded intervention measures such as shutdown, voltage reduction, flow reduction, and early warning pending observation, and are endowed with adaptive adjustment capabilities. This graded decision-making mechanism combines the severity of the anomaly with context tolerance to ensure that control actions are both timely and not excessive, fully guaranteeing a balance between system continuity and security.

[0035] Furthermore, in the tiered intervention measures: When the severity of the anomaly is extremely high, a shutdown command is triggered; when it is high, a voltage reduction regulation is performed; when it is medium, a flow load reduction is implemented; when it is low, only an early warning command for observation is generated. The severity of the anomaly is generated based on a comprehensive score of the transient anomaly factor amplitude, the trend anomaly factor duration, and the number of coupled pipeline nodes. In one implementation, the self-learning algorithm includes: using the flow response curve and abnormal chain changes after intervention as feedback samples to update the normal operating condition benchmark and abnormal pattern template in the global feature library; and adjusting the threshold parameters for multi-scale anomaly detection, the weight distribution of context factor analysis, and the decision rules for the intervention strategy using incremental learning or online Bayesian updates. The flow response curve, abnormal chain changes, and weight changes after intervention are archived, and the detection threshold, context weights, and intervention strategy are continuously optimized through self-learning algorithms (such as online Bayesian updates and incremental learning). The system can gradually adapt to the personalized operating condition drift of specific semiconductor production lines (such as equipment aging and chemical batch differences) as the operating time increases, overcoming the shortcomings of traditional models that "become rigid upon deployment," and achieving long-term stable operation and continuous performance improvement. In this application, the actuator is automatically triggered based on the intervention level command to adjust parameters such as pipeline flow and system pressure in real time, and simultaneously records changes in environmental variables and context factors during the execution process, providing closed-loop data support for subsequent feedback learning. This closed-loop control link achieves end-to-end automation from perception to execution, significantly reducing the delay of manual intervention and effectively improving emergency response efficiency.

[0036] The following is a detailed description of the multi-pipeline liquid flow control method claimed in this application: Step S1 is as follows: Step S1.1 Perform real-time synchronization processing on the flow, temperature, and pressure signals collected by the sensor network distributed on multiple pipelines to eliminate timestamp offsets caused by communication delays and obtain time-aligned original sensor data sequences.

[0037] The back-end control system receives heterogeneous physical quantity signals from a distributed industrial sensor network, including pulse frequency signals output by electromagnetic flowmeters, voltage signals converted from platinum resistance temperature sensors, and current signals transmitted by piezoresistive pressure transmitters. These signals exhibit nondeterministic timestamp offsets due to differences in transmission path length and network congestion. The system first initiates a high-precision clock synchronization protocol, calibrating the local clocks of each edge acquisition node using a network time protocol or a precise time protocol. Then, a time delay estimation algorithm based on cross-correlation functions is employed to calculate the hysteresis of each channel signal relative to a reference clock. Finally, linear interpolation resampling technology maps the non-uniformly sampled discrete data points onto a unified global time axis. This process eliminates phase distortion caused by communication link jitter, ensuring the comparability of the physical states of different pipeline locations at the same time. Ultimately, it outputs a raw sensor data sequence with strict time alignment properties, providing an accurate time series reference for subsequent multivariate coupling analysis.

[0038] After initial time alignment, the system further performs data cleaning to remove outliers caused by momentary sensor malfunctions. The Laida criterion is used to statistically examine the data within each time window, identifying and marking outliers exceeding three standard deviations. These marked outliers are not discarded directly; instead, they are filled in using Kalman filtering with valid data from adjacent time points to maintain the continuity and smoothness of the data sequence. This process not only repairs local data gaps but also effectively suppresses high-frequency noise interference with subsequent feature extraction, resulting in an output original sensor data sequence that possesses both temporal consistency and high data quality.

[0039] The final time-aligned raw sensor data sequence is encapsulated into a standardized data structure, including time index, sensor ID, physical quantity type, and numerical field, and stored in a high-speed circular buffer for downstream modules to access.

[0040] Step S1.2 Based on the industrial control protocol parsing module, the received valve status, equipment start / stop signals and manual operation logs are processed by structured decoding to extract discrete control action events and their occurrence times, and generate a standardized operation event record set.

[0041] The generated standardized operation event record set is stored in a time-series database in the form of key-value pairs, supporting efficient time-range queries and complex condition retrieval. This record set is linked to continuous sensor data sequences through a unified time axis index, enabling the system to trace back the operational context at any point in subsequent analysis.

[0042] Step S1.3: Use the production line scheduling system interface to obtain the process task arrangement table for the current cycle, and combine it with auxiliary information such as maintenance notifications and power fluctuation records received from the external event reporting channel to perform event tagging processing, forming an external context event sequence with time sequence alignment marks.

[0043] The system establishes data connections with the Manufacturing Execution System (MES) and Energy Management System (EMS) through an open application programming interface (API), pulling macro-level scheduling and external environment information in real time, including production batch plans, process changeover times, scheduled maintenance windows, and power grid quality monitoring reports. The system performs spatiotemporal registration of this non-real-time planning information with the real-time acquired physical data, using a timestamp matching algorithm to map the start and end times of process tasks, the duration of maintenance activities, and the occurrence periods of power fluctuations onto a unified timeline. By semantically tagging these time periods, the system assigns a clear operating context to each segment of time-series data, such as "high-load operation period," "downtime maintenance period," or "voltage sag disturbance period," ultimately forming an external context event sequence with precise time-series alignment markers, providing crucial clues for understanding the macro-level causes behind flow fluctuations.

[0044] Step S1.4 performs spatiotemporal fusion processing on the time-aligned continuous sensor data sequence, the standardized operation event record set, and the labeled external context event sequence to construct a multidimensional time-series raw dataset containing multi-source heterogeneous variables based on a unified time axis.

[0045] The system initiates a multi-source data fusion engine, using preprocessed time-aligned continuous sensor data sequences, discrete standardized operational event records, and semantically defined external context event sequences as input sources. It performs rigorous spatiotemporal registration based on a globally unified high-precision timeline. The fusion engine employs a columnar storage architecture, using continuous analog signals as the main data column and converting discrete event signals into indicator variables or one-hot encoded vectors as auxiliary data columns. These three data are then horizontally concatenated on the same time slice. By introducing a spatial topology index, the system maps sensor data from different pipeline locations to their corresponding valve operations and process area labels, ultimately constructing a dense, multi-dimensional time-series raw dataset where rows represent time steps and columns represent multi-source heterogeneous variables (physical quantities + operational quantities + environmental quantities). This achieves a deep fusion of the physical world state and the control logic background.

[0046] Step S1.5 performs integrity verification and missing pattern analysis on the constructed multidimensional time series raw dataset. Linear interpolation based on sliding window and forward preserve strategy are used to fill in locally missing data to ensure the availability and consistency of the dataset in subsequent analysis.

[0047] Specifically, the detailed implementation process of step S2 is as follows: Step S2.1 Normalize each dimension of the original multidimensional time series dataset and map the data of different dimensions to the [0,1] interval based on the minimum-maximum scaling algorithm to obtain the normalized time series input sequence, which provides a numerical stability basis for cross-scale feature extraction.

[0048] The system reads variable columns containing different physical dimensions such as flow rate, pressure, and temperature from the multidimensional time-series raw dataset. For each column, it independently calculates its global maximum and global minimum values ​​within the entire observation window, and then uses the minimum-maximum scaling formula to scale the raw data. Numerical values ​​are linearly mapped to a standardized closed interval [0,1]. This process eliminates the problem of significant differences in magnitude caused by variations in sensor ranges, prevents large numerical variables from dominating subsequent gradient descent or distance calculations, and preserves the relative trends and distribution patterns within the original data.

[0049] During normalization, the system first performs boundary value checks on each variable column to remove extreme outliers and prevent them from distorting the calculation of global maximum and minimum values. For variables with physical upper and lower limits, such as valve opening (0-100%), the control system directly uses the theoretical boundary as the scaling reference; for variables without fixed boundaries, such as instantaneous flow rate, the 99th percentile of the historical statistical distribution is used as the dynamic boundary to accommodate the slow drift of operating conditions.

[0050] The specific execution process of the min-max scaling algorithm is to process the original data... Convert to The formula is .in and These are the minimum and maximum values ​​of the variable within the current analysis window, respectively. The system applies this formula independently to each time series in the multidimensional dataset, generating a new set of normalized series. This process not only unifies the numerical range of the data but also makes the magnitudes of change between different physical quantities comparable.

[0051] Step S2.2 Based on the preset time granularity configuration parameters, four types of sliding windows are defined: second-level (1s step size / 10s window length), minute-level (30s step size / 5min window length), hour-level (10min step size / 1h window length), and process cycle-level (dynamically divided according to the single batch production cycle). The normalized time-series input sequence is sliced ​​in layers to generate a multi-scale subsequence set.

[0052] Based on a pre-defined multi-resolution analysis strategy, the system instantiates four sliding window generators with different time granularity attributes, corresponding to four time scales: second-level, minute-level, hour-level, and process cycle-level. The second-level window uses a 1-second sliding step and a 10-second window length to capture high-frequency transient pulses; the minute-level window is set with a 30-second step and a 5-minute window length, focusing on short-term operational responses; the hour-level window spans a 1-hour interval with a 10-minute step, aiming to identify medium- to long-term drift trends; and the process cycle-level window dynamically binds to the time span of production batches, adaptively adjusting its length according to the start and end times of work orders to cover the complete process cycle. The system uses these windows to perform overlapping or non-overlapping hierarchical slicing of the normalized time-series input sequence, truncating continuous long sequences into numerous local subsequence fragments with specific time spans. These fragments are ultimately aggregated into a multi-scale subsequence set covering millisecond-level mutations to day-level evolutions, providing structured data units for hierarchical feature extraction.

[0053] Step S2.3 For each type of scale subsequence, apply a statistical feature extraction algorithm to calculate the mean, variance, slope, and kurtosis to obtain a primary feature vector set describing the intensity of flow fluctuations, the rate of pressure change, and the direction of temperature shift.

[0054] The system traverses each data segment in the multi-scale subsequence set, invoking the built-in statistical feature calculation engine to quantify the numerical distribution and trends within each segment. For each subsequence, the system first calculates its first moment (mean) to characterize the average level within that period, calculates the second central moment (variance) to measure the dispersion and fluctuation intensity of the data, uses the least squares method to fit a linear regression equation to extract the slope parameter to reflect the rate and direction of change of the variable, and calculates the fourth standardized moment (kurtosis) to assess the sharpness of the data distribution and the probability of extreme values. These statistics are combined into a fixed-dimensional primary feature vector, where the mean reflects the steady-state level, the variance reveals the magnitude of noise or disturbance, the slope indicates the drift trend, and the kurtosis captures sudden shocks. Together, they constitute a digital fingerprint describing the intensity of flow fluctuations, the rate of pressure change, and the direction of temperature shifts, providing a basic criterion for subsequent anomaly pattern recognition.

[0055] Step S2.4 In the process cycle-level subsequence, the Fourier transform algorithm is introduced to perform frequency domain decomposition processing on the periodic operating mode, extract the energy proportion of the dominant frequency component, and generate periodic characteristic parameters that reflect the steady-state operating characteristics.

[0056] The control system is specifically designed for process cycle-level subsequences. It applies a Fast Fourier Transform (FFT) algorithm to convert the time-domain signal to the frequency domain, performing spectral analysis on long sequences containing complete production batches. By calculating the signal's power spectral density, the system identifies several dominant frequency peaks in the spectrum. These peaks correspond to the pump's rotational frequency, the reciprocating compressor's stroke frequency, or the resonant frequency caused by periodic valve switching. The system calculates the proportion of each dominant frequency component in the total signal energy and monitors the amplitude stability of the fundamental frequency and harmonic components, encoding the extracted frequency domain features as periodic characteristic parameters. These parameters quantitatively describe the system's periodicity under steady-state operation, effectively identifying frequency drift caused by mechanical wear or harmonic distortion caused by load changes, thus compensating for the limitations of pure time-domain statistical features in capturing periodic patterns.

[0057] Step S2.5 re-aggregates the primary feature vector groups extracted at each scale with the periodic feature parameters according to the time index to form a multi-scale temporal feature matrix with a hierarchical structure, which serves as the input data basis for the next stage of anomaly factor identification.

[0058] During the aggregation process, the control system addresses the mismatch in temporal resolution between features at different scales. Second- and minute-level features have high temporal density, while hour- and process cycle-level features are relatively sparse. The system employs nearest-neighbor interpolation or forward padding strategies to extend low-frequency features onto the high-frequency time axis, ensuring that each time point possesses a complete multi-scale feature description. For example, a given second not only possesses the statistical features of that second but also inherits the feature labels of the current minute, current hour, and current batch. This approach guarantees the continuity of the feature matrix in the temporal dimension, enabling subsequent time-series models to capture cross-scale dependencies.

[0059] The detailed implementation process of step S3 is as follows: Step S3.1 Construct a non-subsampled dual-tree complex wavelet transform (NS-DTCWT) model. Based on its multi-resolution decomposition capability, perform layer-by-layer decomposition processing on the high-frequency detail components in the multi-scale time-series feature matrix to enhance the sensitivity to transient pulse-type anomalies and obtain multi-layer wavelet coefficient tensors.

[0060] The system initializes a non-downsampled dual-tree complex wavelet transform algorithm engine, loads pre-defined bioorthogonal wavelet basis functions and filter banks, and inputs high-frequency detail components from the multi-scale time-series feature matrix as the input signal into the decomposition process. This model utilizes the translation invariance of the dual-tree structure to perform multi-level iterative decomposition of the input signal. In each level of decomposition, two sets of wavelet coefficients—real and imaginary—are generated simultaneously, corresponding to different phase information of the signal. Through progressive downsampling (maintaining resolution in non-downsampling mode) and filtering operations, the system separates the original signal into sub-band signals of different frequency bands, focusing on preserving and amplifying transient details in the high-frequency range. The multi-layer wavelet coefficients generated during the decomposition process are organized into a three-dimensional tensor structure, with dimensions corresponding to the decomposition level, time index, and real / imaginary channel, respectively. This tensor finely characterizes the local energy distribution of the signal in the time-frequency plane, significantly enhancing the sensitivity to transient pulse-like anomalies such as short-duration pulses and spike interference, providing a high-resolution time-frequency representation for subsequent extraction of transient anomaly response maps.

[0061] In the layer-by-layer decomposition process, the control system first performs low-pass and high-pass filtering on the input signal to obtain approximate and detail components. Unlike traditional wavelet transform, the non-downsampling mode does not perform decimation after each layer of decomposition, but instead uses upsampling filter coefficients to maintain the signal length. This means that the wavelet coefficient sequence of each layer has the same time length as the original signal, thus maintaining accurate time positioning capability.

[0062] The obtained multi-layer wavelet coefficient tensor is the core data structure for subsequent analysis. Each element in the tensor represents the energy intensity of the signal at a specific time and frequency scale. The real coefficients reflect the amplitude changes of the signal, while the imaginary coefficients reflect the phase changes. The combination of the two can more accurately describe the morphology of transient events.

[0063] The multi-layer wavelet coefficient tensor generated in this step not only removes background noise interference but also highlights singularities and discontinuities in the signal. Compared to simple time-domain difference or frequency-domain analysis, wavelet transform can simultaneously locate anomalies in both time and frequency dimensions, making it particularly suitable for processing common non-stationary and nonlinear transient signals in pipeline systems. This high-dimensional tensor data provides a rich source of information for subsequent threshold quantization and anomaly map generation, enabling the system to keenly capture fleeting fault signs. The continuous wavelet transform formula is as follows: in, The input is a multi-scale temporal feature signal. The selected mother wavelet function, Indicates complex conjugation. It is the scale factor (the reciprocal of the corresponding frequency). This represents the translation factor (corresponding to the time position). In the non-downsampled dual-tree complex wavelet transform, this integration process is discretized using two sets of parallel filter banks, generating real and imaginary coefficients respectively, thereby constructing a complex wavelet coefficient tensor with translation invariance. This formula describes the basic principle of the continuous wavelet transform. In the discrete implementation of this embodiment, recursive filtering and upsampling operations are used to approximate this continuous integration effect, ensuring high-precision decomposition of transient signals.

[0064] Step S3.2 performs threshold quantization on each level of detail coefficients in the multi-layer wavelet coefficient tensor, uses the adaptive SURE thresholding algorithm to suppress background noise interference, extracts significant abrupt change energy segments, and generates transient anomaly response maps.

[0065] The system traverses the detailed coefficient matrix of each layer in the multi-layer wavelet coefficient tensor and applies an adaptive thresholding algorithm based on Stein's Unbiased Risk Estimation (SURE) for noise suppression and feature enhancement. For the coefficient distribution of each layer, the system first estimates the variance level of the background noise, and then calculates the optimal soft or hard threshold that minimizes the mean square error. The system sets coefficients with absolute values ​​less than the threshold to zero, treating them as background noise and filtering them out; it retains coefficients with absolute values ​​greater than the threshold, considering them as significant abrupt energy components. The sparse coefficient matrix after threshold quantization is reconstructed into a binarized or grayscale energy distribution map, i.e., a transient anomaly response map. This map clearly outlines the spatiotemporal trajectory of high-energy bursts on the time-to-ground scale plane, eliminating the influence of random fluctuations and retaining only statistically significant transient anomalies, providing intuitive evidence and quantitative basis for judging sudden faults such as sudden increases in flow and pressure shocks.

[0066] Step S3.3 Construct a multi-scale hierarchical clustering model based on dynamic time warping distance, perform similarity measurement on hourly and process cycle-level feature vectors, identify isolated sample points that deviate from normal clusters, and obtain a candidate set of trend drift anomalies.

[0067] Step S3.4 performs slope persistence test and confidence interval boundary crossing analysis on the candidate set of trend drift anomalies to determine whether they constitute a significant trend evolution trajectory and outputs the trend anomaly factor sequence.

[0068] The control system receives a candidate set of trend drift anomalies and performs rigorous statistical hypothesis testing on each candidate sample point and its neighborhood data to eliminate false anomalies caused by random fluctuations. First, the system performs piecewise linear regression on the feature sequences within the time window of the candidate point, calculates the slope value of each segment, and tests the sign consistency and magnitude significance of the slope, performing a slope persistence test to confirm the existence of a unidirectional, continuous evolution trend. Second, the system constructs dynamic confidence intervals (e.g., 95% confidence bands) based on historical normal data, monitoring whether candidate feature values ​​repeatedly exceed the upper or lower bound of the confidence interval, performing boundary crossing analysis. Only samples that simultaneously satisfy a continuously non-zero slope and continuously cross confidence boundaries are confirmed as true trend anomalies. The system concatenates these confirmed anomaly points in chronological order, assigns anomaly intensity scores, and finally outputs a trend anomaly factor sequence. This sequence quantitatively describes the rate and direction of system performance degradation, providing a basis for predictive maintenance decisions. The slope persistence test aims to verify the stability of the changing trend. The system divides the time series near the candidate point into several overlapping sub-windows, fits a straight line to each sub-window, and calculates the slope. If the slope of multiple consecutive sub-windows If all values ​​are greater than zero (or all values ​​are less than zero), and their absolute values ​​exceed the statistical threshold for noise levels, then a continuous upward (or downward) trend is considered to exist.

[0069] Confidence interval boundary crossing analysis focuses on assessing the degree of deviation. The system uses a sliding window to calculate the mean of historical normal data. and standard deviation Construct dynamic confidence intervals ,in This represents the critical value corresponding to the confidence level (e.g., 1.96 corresponds to 95%). The system monitors whether candidate feature values ​​are continuous. points (e.g.) The value falls outside the range. This continuous crossing mechanism effectively filters out occasional spikes, ensuring that only sustained deviations are considered trend anomalies.

[0070] In determining whether a significant trend trajectory exists, the system integrates the results of slope tests and boundary crossings. If, over a certain period, the characteristic value not only changes monotonically but also steadily moves out of the confidence band, the system determines this to be a significant trend trajectory. The system records the starting point, ending point, average rate of change, and maximum deviation of this trajectory. These parameters constitute the core content of the trend anomaly factor. For example, if the flow characteristic value decreases at a constant slope for five consecutive hours and falls below the lower limit, the system will generate a high-intensity trend anomaly factor, indicating a possible serious leak or pump efficiency degradation.

[0071] The output trend anomaly factor sequence is a time series, with each element containing a timestamp, anomaly type (rising / falling), anomaly intensity (a comprehensive score based on slope and deviation), and confidence level. This sequence complements the previous transient anomaly response map; the former focuses on sudden shocks, while the latter focuses on long-term degradation. The trend anomaly factor sequence is directly input into the subsequent anomaly fusion module, helping the system to comprehensively understand the health status of the pipeline system, especially those chronic faults that are difficult to detect but pose significant risks, thus realizing a shift from "post-event alarm" to "pre-event warning."

[0072] Step S3.5 performs spatiotemporal alignment and fusion processing on the transient anomaly response map and the trend anomaly factor sequence to generate a joint anomaly factor matrix, where each element represents the comprehensive anomaly degree identifier at a specific time scale location.

[0073] The generated joint anomaly factor matrix is ​​a dense three-dimensional array with values ​​normalized to [0,1], where 0 represents complete normality and 1 represents extreme anomaly. High-value regions in the matrix clearly indicate the time and frequency range of anomaly occurrences. For example, a high value at a certain time in the high-frequency layer indicates a transient impact; a high value in the low-frequency layer with a long duration indicates a trend-based fault; and high values ​​across the entire frequency band suggest a potential catastrophic accident. The system also performs morphological filtering on the matrix to remove isolated noise points and connect adjacent anomaly regions to form coherent blocks of anomaly events.

[0074] This joint anomaly factor matrix serves as a unified interface for subsequent processing, shielding the complexity of various underlying detection algorithms and providing a standardized anomaly metric for upper-layer applications. It not only tells the system "there is an anomaly," but also, through the matrix's structure, tells the system "what kind of anomaly" (transient or trend-based, high-frequency or low-frequency). This comprehensive anomaly representation improves the accuracy of subsequent contextual analysis and root cause analysis, enabling the system to more intelligently understand the nature of the anomaly and providing data support for developing precise intervention strategies.

[0075] The detailed implementation process of step S4 is as follows: Step S4.1 Construct a context factor encoder to perform one-hot encoding and embedding mapping on the production line process cycle stage, valve switching status, and external event type to generate a low-dimensional dense context feature vector, which serves as the basic representation for context factor parsing.

[0076] Step S4.2 Design a context factor weight allocation module based on attention mechanism. Use historical anomaly samples to train the module to learn the influence strength of transient anomaly response under different context conditions and output the initial static weight distribution.

[0077] The system constructs a weight allocation network based on a scaled dot product attention mechanism. It uses context feature vectors as queries and predefined anomaly pattern prototypes as keys and values ​​to quantify the potential contribution of different contextual factors to the current transient anomaly response. This module calculates the correlation score between the context vector and each anomaly pattern through a multilayer perceptron nonlinear transformation, and generates a probability distribution-form weight vector after normalization using a Softmax function. The module is trained offline under supervised supervision using a massive historical anomaly sample dataset. The loss function is defined as the cross-entropy between the weight distribution and the expert-annotated true cause distribution, forcing the network to learn to assign higher attention to relevant contextual factors under specific operating conditions (such as "high load + valve half-open"). After training convergence, the module can output an initial static weight distribution reflecting the influence strength of each factor for any input context state. This distribution reveals which environmental factors are most likely to induce flow anomalies under general statistical laws, providing a priori knowledge benchmark for real-time dynamic adjustment.

[0078] The core of the attention mechanism lies in dynamic focusing. The system will dynamically focus the encoded contextual feature vector. Mapped to query vector Simultaneously, a set of learnable abnormal pattern prototype vectors are used as the key matrix. Sum matrix The formula for calculating attention score is: .in, The correlation between the current context and each anomaly pattern was calculated. To prevent gradient vanishing during scaling, Softmax converts scores into probability weights. This mechanism allows the model to automatically identify the most critical contextual factors in the current operating condition. For example, in "cleaning mode," the model automatically assigns higher weight to "cleaning agent dosing events," while in "normal production" it focuses more on "pump status."

[0079] S4.3 introduces an online Bayesian update mechanism, which dynamically corrects the weights of context factors based on real-time incoming abnormal feedback signals, adjusts the influence weights of each factor under the current working condition based on posterior probability estimation, and generates a dynamic context weight vector.

[0080] The system integrates an online Bayesian inference engine, using the initial static weight distribution output from the previous step as the prior probability. Furthermore, real-time monitored anomaly feedback signals (such as the actual type and intensity of the anomaly and the confirmation results) will be used as likelihood evidence. Using Bayes' theorem The system calculates the posterior probability of each context factor as a true cause in real time. When an abnormal feature is detected to be highly correlated with a specific context factor (e.g., a sudden pressure drop immediately following valve closure), the likelihood function value increases, significantly boosting the posterior weight of that factor; conversely, if the observed data does not match expectations, the corresponding weight is reduced. This mechanism employs a recursive update algorithm, eliminating the need to retrain the entire model and requiring only a small number of new samples for rapid convergence, thus achieving millisecond-level dynamic correction of context factor weights. The resulting dynamic context weight vector reflects the actual driving force of each environmental factor on the anomaly at the current moment, overcoming the limitation of static models in adapting to temporary changes in operating conditions (such as sudden maintenance or temporary production changes), and ensuring the real-time adaptive capability of cause analysis.

[0081] Step S4.4 performs weighted fusion processing on the dynamic context weight vector and the transient anomaly factor to construct a context-aware anomaly enhancement input term, which is used to improve the robustness of the traffic anomaly detection model to external disturbances.

[0082] Step S4.5 outputs the context-modulated anomaly enhancement input and dynamic context weight vector, which serve as key inputs to the traffic anomaly detection model to support subsequent anomaly source attribution analysis.

[0083] The implementation process of step S5 is as follows: Step S5.1 Construct a multi-pipeline topology modeling framework based on graph neural networks. Treat each pipeline as a graph node, establish a set of directed edges based on physical connection relationships and fluid dynamic coupling strength, and generate a pipeline coupling topology graph.

[0084] The system initializes a Graph Neural Network (GNN) computational framework, abstracting each independent pipeline in the industrial fluid transport network as a node in a graph structure. Node attributes include the pipeline's real-time flow rate, pressure, temperature, and local anomaly factors. Based on the pipeline network engineering design drawings and physical connection logic, the system establishes directed edges between connected pipeline nodes, with the edge direction determined by the fluid flow direction. The system calculates the coupling strength between adjacent pipelines (such as pressure transmission coefficient and flow distribution ratio) based on fluid dynamics principles, quantifying it as edge weights, thereby constructing a weighted directed pipeline coupling topology graph. This graph not only explicitly represents the physical connection structure of the pipeline network but also implicitly encodes the interaction mechanisms at the fluid dynamics level, providing a precise topological data structure foundation for subsequent analysis of anomaly propagation paths and synergistic effects within the pipeline network using message passing mechanisms.

[0085] The pipeline coupling topology graph generated in this step is a dynamic graph structure. As valves are opened and closed or pumps are started and stopped, the connectivity and edge weights of the graph are updated in real time. The system uses an adjacency matrix or adjacency list to store the graph structure and leverages sparse matrix techniques to optimize storage and computation efficiency. This topology graph serves as input to the GNN, enabling the model to naturally aggregate information from neighboring nodes during message passing, thereby detecting abnormal propagation across pipelines.

[0086] By constructing this topology modeling framework, the system elevates the monitoring of isolated single pipelines to a level of network-wide collaboration. The pipeline coupling topology diagram provides indispensable structural prior knowledge for accurately identifying single-path and coupling anomalies and distinguishing between internal and external causes, which is a key step in achieving systematic and global fault diagnosis.

[0087] Step S5.2 takes the transient anomaly factor and context factor weights corresponding to each pipeline as node attribute input, performs neighborhood information aggregation processing through graph convolution operation, extracts local anomaly propagation mode features, and generates node-level coupling state embedding vector.

[0088] The system concatenates the transient anomaly factor of each pipeline with a dynamic context weight vector, using this as the initial feature attribute input to the graph convolutional network (GCN). During the forward propagation of the network, each node aggregates the feature information of its neighboring nodes (i.e., physically connected upstream or downstream pipelines) through graph convolution operators. Specifically, the node uses a normalized adjacency matrix to perform a weighted summation of the features of its neighboring nodes and updates its own hidden state through a nonlinear activation function. This process simulates the physical propagation mechanism of anomalous signals in the fluid network, ensuring that each node's new feature representation not only includes its own local anomaly information but also incorporates the collaborative state from surrounding pipelines. After stacking multiple layers of graph convolutions, the system extracts local anomaly propagation pattern features rich in topological context information and compresses them into a high-dimensional node-level coupled state embedding vector. This vector implicitly encodes the diffusion path and impact range of the anomaly in the pipeline network, providing a deep feature representation for determining whether the anomaly has cross-pipeline linkage.

[0089] The core formula for graph convolution operations is: ,in To add a self-loop adjacency matrix, For degree matrix, For the first The node feature matrix of the layer This is a learnable weight matrix. Through this operation, information flows in the graph, at intervals... The nodes of the layer are passing through Information can be exchanged after each convolution. In this embodiment, 2-3 convolutional layers are sufficient to cover most local coupling relationships, enabling nodes to perceive the status of directly connected and indirectly affected pipelines.

[0090] Neighborhood information aggregation processing enables the system to capture the interconnected effects of a single leak. For example, when the main pipe pressure is abnormal, all nodes in the branch pipes will sense this change through aggregation, and their embedding vectors will exhibit a high degree of similarity. Conversely, if the leak is unique to a particular branch, only the embedding vectors of that node and its immediate neighbors will change significantly, while other distant nodes will remain unchanged. This differentiated characteristic response is key to distinguishing between single-path leaks and coupled anomalies.

[0091] The generated node-level coupling state embedding vectors are dense vectors that form specific clustering structures in the vector space, where similar anomaly propagation patterns are mapped to nearby regions. The system uses these vectors for subsequent clustering analysis and consistency comparison to identify which pipeline anomalies occur synchronously and which occur independently.

[0092] Through this step, the system deeply integrates the topological structure with anomaly features, generating a coupled state representation with powerful expressiveness. These embedded vectors not only contain rich local information but also reflect global collaborative relationships, providing a feature foundation for subsequent anomaly type discrimination. This enables the system to unravel complex network interactions and accurately pinpoint the essence of anomalies.

[0093] Step S5.3 performs cluster analysis and anomaly consistency comparison on the node-level coupling state embedding vector. If multiple adjacent pipelines exhibit highly synchronized abnormal responses within the same time period, it is determined to be a coupling abnormal event; otherwise, it is determined to be a single-path abnormal event.

[0094] The system performs unsupervised clustering analysis (such as K-Means or DBSCAN) on the generated node-level coupled state embedding vectors, grouping pipeline nodes with similar embedding representations into the same cluster. Simultaneously, the system calculates the correlation coefficient of anomalous responses of adjacent nodes over time series and performs anomaly consistency comparison. If a cluster contains multiple physically adjacent pipeline nodes, and these nodes all have anomaly scores exceeding a threshold and exhibiting highly consistent trends (correlation coefficient > 0.8) within the same time window, the system classifies the event as a coupled anomaly, indicating that the anomaly source may be located at a common upstream node or caused by a systemic disturbance. Conversely, if only a single node or its anomaly response is significantly out of sync with its neighbors (low correlation), it is classified as a single-path anomaly, indicating that the fault is confined to within that pipeline (e.g., local blockage or instrument malfunction). This logical discrimination process fully utilizes the dual constraints of topological proximity and state synchronization, effectively defining the scope of anomaly propagation and classifying event types.

[0095] Through this discrimination process, the system outputs a clear anomaly type label (coupled / single-path), pointing the way for subsequent root cause analysis. If it is a coupled anomaly, the focus shifts to the common parent node or the external environment; if it is a single-path anomaly, the focus is on the equipment and components within the pipeline itself. This intelligent classification significantly improves the efficiency and targeting of fault diagnosis.

[0096] Step S5.4 Combines the external event weight components in the dynamic context weight vector to determine whether the coupling anomaly occurs during a specific external disturbance. If there is a strong correlation, it is attributed to an externally induced anomaly; otherwise, it is attributed to an internal structural fault.

[0097] Based on the determination that the event is a coupled anomaly, the system further retrieves the "external event" component (such as the weight values ​​of labels for power grid fluctuations, raw material changes, extreme weather, etc.) from the dynamic context weight vector to analyze the overlap between the anomaly's time window and the active period of the external event. If the time of the coupled anomaly's outbreak highly overlaps with the time of the high-weight external event label, and statistical tests show a significant positive correlation between the two (e.g., the Granger causality test passes), the system attributes the anomaly to an externally induced anomaly, considering it as an external disturbance transmitted through the pipeline network causing multiple synchronous fluctuations. Conversely, if the external event has a low weight or the time is mismatched, the system infers that the anomaly originates from a structural fault within the pipeline network (e.g., a main pipeline rupture or a shared pump unit failure). This attribution logic deeply integrates environmental context with topological coupling analysis, realizing the qualitative determination of the anomaly's root cause.

[0098] The introduction of external event weighting components provides important contextual clues for attribution analysis. The system checks whether high-weight events such as "voltage sag" or "raw material switching" exist at the time of the anomaly. If so, and these events are known to cause flow and pressure fluctuations, the system tends to classify them as external causes.

[0099] The determination of strong correlation employs time series analysis. The system calculates the cross-correlation function between the external event indication sequence and the anomaly intensity sequence, identifying the maximum correlation peak and its lag time. If the peak is significant and the lag time conforms to the laws of physical propagation, a strong correlation is confirmed. Furthermore, the system considers the matching degree between the event duration and the anomaly duration to ensure the rigor of the attribution logic.

[0100] When the cause is attributed to an externally induced anomaly, the system generates a corresponding label and suggests mild measures such as "observation and waiting" or "adjusting process parameters," because external disturbances are usually temporary. However, when the cause is attributed to an internal structural fault, a higher-level alarm is triggered, recommending immediate shutdown and repair, as these types of faults are often destructive and do not recover on their own.

[0101] Through this step, the system completes the final identification of the source of the anomaly, enhancing its intelligence and practical value, and making operation and maintenance decisions more scientific and accurate.

[0102] Step S5.5 outputs the abnormality type discrimination results and source attribution labels, including categories such as "single-path internal abnormality" and "multi-path coupling external inducement", for subsequent intervention strategy generation module to call.

[0103] The specific implementation process of step S6 is as follows: Step S6.1 Define a four-level anomaly severity rating system, and perform comprehensive scoring processing based on the amplitude of transient anomaly factors, the duration of trend anomaly factors, and the size of the coupling range to generate an anomaly severity index.

[0104] The system constructs a four-level anomaly severity rating system, comprising "Low Risk (Level IV)," "Medium Risk (Level III)," "High Risk (Level II)," and "Extremely High Risk (Level I)," serving as the benchmark framework for intervention decisions. The system reads in real-time the instantaneous amplitude of transient anomalies, the duration of trend anomalies, and the number of pipeline nodes involved in coupled anomalies (coupling range), and constructs a comprehensive scoring function using multi-attribute utility theory. This function performs a normalized weighted summation of the three indicators, where amplitude reflects the impact intensity, duration reflects the degree of deterioration, and coupling range reflects the breadth of impact. The calculated comprehensive score is mapped to an anomaly severity index ranging from 0 to 100, with higher indices indicating greater risk. The system sets three dynamic thresholds to divide the index into four levels, thereby quantitatively defining the severity of the anomaly and providing a quantitative decision-making benchmark for matching differentiated intervention strategies, ensuring that the intervention intensity matches the risk level.

[0105] The design of the comprehensive scoring function fully considers the physical meaning of each indicator. The magnitude weight is usually the largest because drastic transients often mean immediate danger; the duration weight is the second largest, reflecting the cumulative effect of chronic failures; and the coupling range weight is used to amplify the scoring of systemic risk.

[0106] The generation of the severity index is performed in real time, updating at set time intervals (e.g., every 5 seconds). The system monitors the index's trend, and if the index rises rapidly, even before reaching a high threshold, an alert will be triggered, indicating that risk is accumulating. This dynamic assessment mechanism enables the system to keenly detect the evolution of the situation, allowing for early detection and intervention.

[0107] The boundaries of the four-level grading system are not fixed; the system can be adaptively adjusted based on historical accident data and expert experience.

[0108] Step S6.2 performs rule engine matching processing on the abnormal severity index and key factors (such as valve status weight and external event weight) in the dynamic context weight vector, and determines the corresponding intervention strategy template by looking up the table.

[0109] The system incorporates a configurable rule engine that uses the calculated severity index and key contextual factors (such as high-weight items like "valve operation," "external power," and "manual maintenance") from the dynamic context weight vector as joint input conditions. The rule engine maintains a multi-dimensional decision lookup table, where row indices represent severity levels and column indices represent the state combinations of key contextual factors. The system searches the table based on the current input, locating a unique cell with a predefined applicable intervention strategy template. For example, when the severity is "high risk" and "external power weight" is high, the "reduced voltage operation - waiting for recovery" template is matched. Through this lookup mechanism, the system can quickly select the most suitable intervention plan from a vast pool of preset strategies, achieving transparency and configurability in the decision-making logic.

[0110] Step S6.3 Select specific control actions according to the intervention strategy template: trigger a shutdown command when the score is extremely high risk; perform pressure reduction regulation when it is high risk; implement flow reduction when it is medium risk; and generate an early warning and observation command when it is low risk.

[0111] Step S6.4 adds priority tags and time constraints to the generated intervention level instructions, and encapsulates them into standard MODBUS / TCP control messages via the industrial communication bus to ensure that the instructions can be correctly parsed and responded to by downstream actuators.

[0112] The system encapsulates selected control action instructions, attaching unique priority tags (e.g., urgent, high, medium, low) and time constraints (e.g., "execute immediately," "effective within 5 minutes," "execute on next shutdown"). Priority tags are used for arbitration when multiple instructions are executed concurrently, ensuring critical instructions are executed first. Time constraints define the validity window of the instruction; expired instructions are automatically invalidated to prevent erroneous actions from outdated instructions. Priority tags are managed using a preemptive scheduling mechanism. When a high-priority instruction arrives, a low-priority instruction being executed is paused or overwritten. The system maintains an instruction queue, sorted by priority, ensuring the most important control actions are processed first. Time constraint checks are performed continuously before instruction issuance and during execution; if a timeout is detected, the system automatically cancels the instruction and logs the result.

[0113] Step S6.5 outputs an intervention level instruction stream with priority and context dependency markers, realizing a closed-loop connection from anomaly identification to decision output.

[0114] The system outputs encapsulated intervention commands in chronological order as a continuous stream of intervention level commands. Each command carries a priority tag, a time constraint, and a context-dependent flag that triggered it (e.g., "triggered by external voltage fluctuations"). The output of the command stream signifies the completion of the entire process of anomaly identification, analysis, and decision-making, achieving a closed-loop connection from perception to action. The presence of context-dependent flags ensures that each command is traceable, enhancing the system's interpretability and traceability, and ensuring that the closed-loop control system possesses the ability to continuously evolve.

[0115] Step S6.6: Based on the intervention level command, the actuator is automatically triggered to adjust parameters such as pipeline flow and system pressure in real time. The intervention level command stream is input into the PLC control unit, which executes the PID control algorithm to perform closed-loop control processing on the opening of the electric regulating valve or the speed of the pump group, thereby realizing the regulation of pipeline flow and system pressure.

[0116] The detailed implementation process of step S7 is as follows: Step S7.1 Synchronously record changes in environment variables and context factors during execution.

[0117] Environmental variables include macroscopic environmental indicators that may affect fluid behavior, such as ambient temperature, humidity, plant power parameters (voltage, current, frequency), and workshop cleanliness.

[0118] Context factors: These are the current process stage (e.g., feeding, reaction, termination), key valve status, manual operation records (e.g., operator confirmation alarm, manual adjustment), external events (e.g., maintenance notification, power fluctuation), and the real-time values ​​of each factor in the dynamic context weight vector.

[0119] Actuator response parameters include valve opening command value and actual feedback value, pump speed set value and actual value, overshoot and settling time during the adjustment process.

[0120] Step S7.2 archives the traffic response curve, abnormal chain changes, and context factor weight changes after the intervention into the global feature library.

[0121] Archived data is stored in a structured format in a global feature library. This feature library employs a hierarchical storage strategy: recent data (e.g., 30 days) is stored on high-speed SSDs for rapid learning, while historical data is migrated to large-capacity HDDs or cloud storage for long-term analysis and macro trend mining.

[0122] Step S7.3 continuously optimizes multi-scale anomaly detection, context factor analysis, and intervention strategies through a self-learning algorithm.

[0123] Once the global feature library accumulates to a certain size (e.g., every 100 intervention events or every 24 hours), the system triggers the self-learning engine, employing the following strategies to optimize the core model: (1) Optimize the multi-scale anomaly detection module Adaptive thresholding: The archived normal response curves (i.e., data segments showing system stabilization after intervention) are used as new "normal samples" to recalculate the statistical distributions (mean, standard deviation, quantiles) of each feature in the multi-scale time-series feature matrix. For the adaptive SURE threshold in wavelet transform, an incremental update formula is used: in It is a locally optimal threshold calculated based on new samples. The learning rate is (0.01-0.1). This is the original optimal threshold. Similarly, for the DTW clustering model, the center vector of normal clusters and the cutoff distance of outliers are updated.

[0124] Feature weight adjustment: By analyzing the correlation between the flow response curves before and after intervention and features at various scales, the weights of features with strong predictive power are increased, while redundant or noisy features are weakened. Online gradient descent or random forest feature importance assessment can be used.

[0125] (2) Optimize the context factor parsing module Incremental training of attention mechanism parameters: The archived context factor weight change sequence is used as a supervision signal (i.e., the true cause distribution) to perform mini-batch incremental training on the attention weight allocation network in step S4. The loss function is defined as the cross-entropy between the static weights output by the network and the actual dynamic weights in the archive. Stochastic gradient descent (SGD) with momentum is used to update the network parameters.

[0126] Online Bayesian Prior Update: Update the prior probability distribution in the Bayesian model using archived context factor-anomaly type pairing data. and likelihood function For example, if the "valve operation" factor frequently accompanies the "transient pulse" anomaly in multiple abnormal events, then the prior weight of this factor and the accuracy of the likelihood estimation are increased. The update formula uses analytical recursion under conjugate priors: in For the parameters of the Dirichlet distribution, Factors in archived data Counting as real triggers.

[0127] (3) Optimize intervention strategies The strategy rule engine utilizes reinforcement learning: Intervention actions (shutdown, voltage reduction, load shedding, early warning) are treated as the action space, while the stabilization time of the post-intervention traffic response curve, overshoot, and whether production interruption occurs are used as reward signals (negative costs). Batch reinforcement learning algorithms (such as Fitted Q-Iteration or decision tree fitting) are used to optimize the decision lookup table in the rule engine. The optimization objective is to maximize the cumulative reward, i.e., to seek intervention actions that result in faster recovery and fewer side effects under the same abnormal conditions.

[0128] Severity rating coefficient adjustment: By utilizing the correlation between the abnormal severity index in the archived data and the actual hazard results (such as whether it leads to scrap or equipment damage), the weight coefficients in the rating function (transient amplitude weight, duration weight, coupling range weight) are refitted using ridge regression or Bayesian linear regression to make the rating results more consistent with the actual risk level.

[0129] Self-learning execution mode: The system supports two learning modes—online learning (real-time updates after each intervention, suitable for production lines with frequent changes) and offline batch learning (unified updates after accumulating a certain amount of data, suitable for production lines with higher stability). After the updated model parameters are verified by simulation or A / B testing, they are distributed to each online module through the hot deployment interface without downtime.

[0130] Step S7.4, Self-learning effect verification and model version management After each self-learning iteration, the system automatically calculates the changes in key performance indicators (KPIs) on the historical data holdout set, including: anomaly detection accuracy, false alarm rate, false negative rate, anomaly tracing accuracy, average recovery time after intervention, and number of production interruptions.

[0131] If the new model significantly outperforms the old model, the system will automatically switch to the new model and save the old model as a rollback version. All model versions and their performance metrics are recorded in the model repository, supporting manual review and manual rollback.

[0132] Example 2 This embodiment provides a multi-pipeline liquid flow control system, which is used to implement the multi-pipeline liquid flow control method described above.

[0133] Specifically, the system includes: a data acquisition and preprocessing module, a multi-scale feature extraction module, an anomaly detection and factor extraction module, a context factor analysis module, a multi-channel collaborative coupling analysis module, a hierarchical intervention decision-making module, and a closed-loop self-learning module.

[0134] More specifically: The data acquisition and preprocessing module is responsible for acquiring multi-source heterogeneous data from the industrial site and performing preprocessing operations such as time synchronization, cleaning, and fusion.

[0135] The multi-scale feature extraction module receives the multi-dimensional time-series raw dataset from the data acquisition module, performs normalization, sliding window slicing, statistical feature calculation and frequency domain analysis, and outputs a multi-scale time-series feature matrix.

[0136] The anomaly detection and factor extraction module is based on wavelet transform and multi-scale clustering model to extract transient anomaly factors and trend anomaly factors from the multi-scale time series feature matrix.

[0137] The context factor parsing module performs weighted labeling on production line process cycles, external events, and equipment status, and dynamically adjusts the context factor weights.

[0138] The multi-pipeline collaborative coupling analysis module uses graph neural networks (GNNs) to model the joint state of multiple pipelines, identify single-path anomalies and coupling anomalies, and distinguish between externally induced and internally induced anomaly sources.

[0139] The tiered intervention decision-making module generates tiered intervention instructions based on the severity of the anomaly and dynamic weights within the context. Specifically, this includes: Severity scoring unit: Defines four levels of anomaly severity (low risk, medium risk, high risk, and very high risk), and performs a weighted comprehensive score based on the amplitude of transient anomaly factors, the duration of trend anomaly factors, and the number of coupled pipeline nodes to generate an anomaly severity index of 0-100.

[0140] The rules engine has a built-in configurable multidimensional decision lookup table that combines severity index with dynamic context weights (such as valve status weights and external event weights) to match the corresponding intervention strategy template.

[0141] Action Mapper: Selects specific control actions based on the strategy template: Extremely high risk - shutdown command; High risk - voltage reduction command; Medium risk - flow reduction command; Low risk - warning and observation command.

[0142] Command encapsulation and communication unit: The generated intervention level commands are attached with priority tags (urgent, high, medium, low) and time constraints (immediate execution, delayed effect, etc.), encapsulated into standard MODBUS / TCP messages, and sent to the field actuators via industrial Ethernet.

[0143] The closed-loop self-learning module records the system response after the intervention is implemented and continuously optimizes the parameters and strategies of each module through a self-learning algorithm.

[0144] This system integrates data acquisition, multi-scale feature extraction, anomaly detection, context parsing, multi-pipeline collaborative analysis, hierarchical intervention, and self-learning into a complete closed-loop architecture. The modules collaborate efficiently through standardized interfaces, avoiding the data fragmentation and difficult linkage issues between multiple independent subsystems (such as separate SCADA, alarms, and PLC logic) in traditional solutions, thus achieving full-process automation from raw signals to control commands.

[0145] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations.

Claims

1. A method for controlling liquid flow in a multi-pipeline system, characterized in that, include: Flow rate, temperature, pressure, valve status, environmental variables, production line process arrangement, manual operation records and context factors of external events at different locations in multiple pipelines are obtained to construct a multidimensional time series raw dataset; Multi-scale time-series window segmentation is performed on the original multi-dimensional time-series dataset. Based on the sliding window technique, time-series features of flow fluctuation, pressure change, and temperature shift are extracted at the second, minute, hour, and process cycle scales, respectively. By inputting multi-scale time series features into wavelet transform or multi-scale clustering models, transient anomaly and trend anomaly factors can be obtained to identify sudden and gradual changes in flow and related parameters at various scales. Based on context factor records, the context factor parsing algorithm is used to perform weighted labeling of production line process cycle, external events, and equipment status, and dynamically adjust the context factor weights. The context factor weights are then combined with transient anomaly factors and input into the flow anomaly detection model to achieve multi-dimensional environmental cause analysis. By utilizing the transient anomaly factor and context factor weights output by the flow anomaly detection model, a multi-pipe collaborative coupling modeling method is adopted to analyze flow anomaly events under multi-pipe joint conditions, determine whether the anomaly is a single-path anomaly or a coupled anomaly, and distinguish between external inducement and internal anomaly source based on this. Determine the anomaly level and coupling relationship of the multi-pipeline collaborative coupling model output, and generate graded intervention measures such as shutdown, voltage reduction, flow reduction or early warning based on the dynamic weight of context factors. Automatically triggers actuators to regulate pipeline flow or system pressure.

2. The multi-pipeline liquid flow control method according to claim 1, characterized in that: When adjusting pipeline flow and system pressure in real time, it also includes: Synchronously record changes in environment variables and context factors during execution; The traffic response curves, abnormal chain changes, and context factor weight changes after the intervention are archived in the global feature library, and the multi-scale anomaly detection, context factor analysis, and intervention strategies are continuously optimized through a self-learning algorithm.

3. The multi-pipeline liquid flow control method according to claim 1, characterized in that: The construction of the multidimensional time-series raw dataset includes: The flow, temperature, and pressure signals collected by the sensors are time-synchronized to eliminate timestamp offsets caused by communication delays and obtain time-aligned raw sensor data sequences. The received valve status, equipment start / stop signals, and manual operation logs are structured and decoded to extract discrete control action events and their occurrence times, generating a standardized set of operation event records.

4. The multi-pipeline liquid flow control method according to claim 3, characterized in that: The construction of the multidimensional time-series raw dataset also includes: obtaining the process task arrangement table of the production line scheduling system, and combining it with the maintenance notifications and power fluctuation records received by the external event reporting channel, performing event labeling processing to form an external context event sequence with time-series alignment tags.

5. The multi-pipeline liquid flow control method according to claim 4, characterized in that: The construction of the multidimensional time-series raw dataset includes: spatiotemporal fusion of time-aligned sensor data sequences, standardized operation event record sets, and external context event sequences with time-aligned markers, and constructing a multidimensional time-series raw dataset containing physical quantities, operation quantities, and environmental quantities based on a unified time axis.

6. The multi-pipeline liquid flow control method according to any one of claims 1-5, characterized in that: The process involves performing multi-scale time-series window segmentation on the original multi-dimensional time-series dataset, and extracting time-series features of flow fluctuations, pressure changes, and temperature shifts at the second, minute, hour, and process cycle scales based on the sliding window technique. Normalize each dimension of the original multidimensional time series dataset, and map the data of different dimensions to the appropriate scales based on the min-max scaling algorithm. The interval is used to obtain the normalized time-series input sequence; Four types of sliding windows are defined: second-level, minute-level, hour-level, and process cycle-level. The normalized time series input sequence is sliced ​​in layers to generate a set of subsequences covering different time spans.

7. The multi-pipeline liquid flow control method according to claim 1, characterized in that: The factors for obtaining transient anomalies and trend anomalies include: A non-subsampled dual-tree complex wavelet transform model is constructed to decompose the high-frequency detail components in the multi-scale time-series feature matrix layer by layer to obtain multi-layer wavelet coefficient tensors. Adaptive SURE threshold quantization is then performed on the detail coefficients of each layer to extract significant abrupt change energy segments and generate transient anomaly response maps. A multi-scale hierarchical clustering model based on dynamic time warping distance is constructed to measure the similarity between hourly and process cycle-level feature vectors, identify isolated sample points that deviate from normal clusters, obtain a candidate set of trend drift anomalies, and confirm the trend anomaly factor sequence through slope persistence test and confidence interval boundary crossing analysis.

8. The multi-pipeline liquid flow control method according to claim 1, characterized in that: The multi-pipeline collaborative coupling modeling method includes: constructing a multi-pipeline topology modeling framework based on graph neural networks, treating each pipeline as a node, establishing directed edges based on physical connection relationships and fluid dynamic coupling strength, generating a pipeline coupling topology graph, and using transient anomaly factors and context factor weights as node attribute inputs, aggregating neighborhood information through graph convolution operations, and generating node-level coupling state embedding vectors for anomaly type discrimination and source attribution.

9. The multi-pipeline liquid flow control method according to claim 2, characterized in that: In the tiered intervention measures: When the severity of the anomaly is extremely high, a shutdown command is triggered; when it is high, a voltage reduction regulation is implemented; when it is medium, a flow load reduction is implemented; when it is low, only an early warning command for observation is generated. The severity of the anomaly is generated based on a comprehensive score of the transient anomaly factor amplitude, the trend anomaly factor duration, and the number of coupled pipeline nodes. The self-learning algorithm includes: using the traffic response curve and abnormal chain changes after the intervention as feedback samples to update the normal operating condition benchmark and abnormal pattern template in the global feature library, and using incremental learning or online Bayesian update to adjust the threshold parameters of multi-scale anomaly detection, the weight distribution of context factor parsing, and the decision rules of the intervention strategy.

10. A multi-pipeline liquid flow control system, characterized in that: The system is used to implement the multi-pipeline liquid flow control method as described in any one of claims 1-9.

Images