A fault diagnosis system for a pipe network

Abstract

The application relates to the field of fault detection and discloses a pipe network fault diagnosis system, which comprises a data acquisition module, a data preprocessing module, a data analysis module and a model development module. The data acquisition module is used for acquiring pipeline operation data from a SCADA system, a simulation environment, an intermediate database, API transmission and manual inspection data input. The data preprocessing module is used for cleaning, normalizing and structurally storing the collected data, including data availability inspection, basic statistical quantity calculation and box plot drawing. The data analysis module is used for constructing a scoring model according to the pipeline operation data and determining a data distribution type according to the scoring model in combination with a dynamic threshold value judgment mechanism. The model development module is used for selecting an optimal model according to data distribution characteristics, supporting multiple model integration, generating and evaluating multiple model results and outputting visualized model results by adopting a hybrid modeling strategy. The real-time early warning module is used for monitoring the operation state of the pipe network in real time and issuing early warning information based on diagnosis results and performing visualized display.

Description

Technical Field

[0001] This invention relates to the field of fault detection, and more specifically to a fault diagnosis system for pipeline networks. Background Technology

[0002] Oil pipelines are the invisible arteries that sustain a nation's energy lifeline; their safe and stable operation directly impacts the continuity of energy supply and national security strategy. However, current pipeline fault diagnosis technology faces numerous challenges, becoming a bottleneck restricting efficient energy transmission.

[0003] On the one hand, traditional diagnostic systems often rely on a single sensor or a simple threshold alarm mechanism. In complex environments, they are easily affected by noise factors such as electromagnetic interference and temperature fluctuations, leading to frequent false alarms. This not only consumes a lot of operation and maintenance resources, but may also weaken the staff's trust in the early warning system.

[0004] On the other hand, due to the lack of efficient data collection and analysis methods, there is a significant time delay in fault detection, making it difficult to achieve real-time monitoring and accurate location of faults. This results in delayed maintenance response, posing safety hazards to pipeline operation and potentially causing serious leakage accidents and ecological disasters. Summary of the Invention

[0005] The purpose of this invention is to provide a fault diagnosis system for pipeline networks to solve the above-mentioned technical problems.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A fault diagnosis system for a pipeline network, comprising:

[0008] The data acquisition module is used to obtain pipeline operation data from SCADA systems, simulation environments, intermediate databases, API transmissions, and manual inspection data entry.

[0009] The data preprocessing module is used to clean, normalize, and structure the collected data, including data availability testing, basic statistics calculation, and box plot drawing. The cleaning process uses statistical analysis and machine learning-based methods to identify and handle outliers and missing values. Normalization uses standardization or normalization functions to unify the data to the same scale range.

[0010] The data analysis module constructs a scoring model based on pipeline operation data, and then determines the data distribution type based on the scoring model combined with a dynamic threshold determination mechanism.

[0011] The model development module selects the optimal model based on data distribution characteristics, supports the integration of multiple models, generates and evaluates the results of multiple models, and outputs visualized model results using a hybrid modeling strategy.

[0012] The real-time early warning module is used to monitor the pipeline network operation status in real time and issue early warning information based on the diagnostic results, which is then displayed visually.

[0013] In a further embodiment, the model development module selects models using convolutional neural networks, recurrent neural networks and their variants, and gradient boosting tree models, and integrates models using stacked ensemble and hybrid models.

[0014] In a further embodiment, the pipeline operation data includes pressure, flow rate, temperature, vibration frequency, fluid density, corrosion rate, valve opening, pump speed, acoustic signal, and gas concentration parameters.

[0015] The process of constructing the scoring model is as follows:

[0016] The average values ​​of various parameters were calculated based on historical pipeline operation data. and standard deviation σ i This allows us to obtain the normal range of each parameter. i represents the i-th parameter;

[0017] Calculate each parameter x i degree of deviation d i ,when At that time, d i =0, when hour, when hour, The overall score S is obtained based on the weight allocation, and its expression is: θ i This represents the weight coefficients corresponding to each parameter, and n represents the total number of parameter terms.

[0018] By combining a dynamic threshold determination mechanism, the data distribution type and anomaly level are determined, including concentrated distribution, positive deviation from concentration, positive long tail, negative deviation from concentration, and negative long tail.

[0019] The failure probability is output based on the comprehensive score and the failure prediction value.

[0020] A further embodiment of the dynamic threshold determination mechanism includes:

[0021] Obtain the standard deviation threshold, interquartile range threshold, and data centrality threshold from historical pipeline operation data;

[0022] Based on the real-time range R of the data i =max(x i )-min(x i ) and historical data average range The threshold ratio is dynamically adjusted using the formula: Adjusted threshold = Initial threshold × [1 + k × (Real-time data range / Historical data average range)], which calculates the adjusted standard deviation threshold, interquartile range threshold, and data concentration threshold, where k is the adjustment coefficient, determined through experiments or experience.

[0023] The final data distribution type is determined based on the adjusted standard deviation threshold, interquartile range threshold, and data centrality threshold; then, the comprehensive score output by the scoring model is compared with the preset scoring interval.

[0024] When the overall score is lower than the score range and the data distribution is concentrated, the anomaly level is low; when the overall score is higher than the score range and the data distribution is positively or negatively deviating from the concentration, the anomaly level is high; otherwise, the anomaly level is medium.

[0025] A further approach involves determining the data distribution type using the following logic:

[0026] When the absolute value of the difference between the mean and the median is less than or equal to the median, and the standard deviation and interquartile range are both lower than the corresponding standard deviation threshold and interquartile range threshold, the data type is determined to be a concentrated distribution.

[0027] When the data shows positive skewness and the proportion of data exceeding the upper bound is higher than the concentration threshold, the data type is determined to be positively skewed from concentration.

[0028] When the data is positively skewed, but the proportion of data exceeding the upper bound is lower than the concentration threshold, the data type is determined to be positive long-tailed.

[0029] When the data shows negative skewness and the proportion of data below the lower bound is higher than the concentration threshold, the data type is determined to be negatively skewed from concentration.

[0030] When the data shows negative skewness, but the proportion of data below the lower bound is lower than the concentration threshold, the data type is determined to be negative long-tailed.

[0031] In a further embodiment, the operation of the real-time early warning module includes:

[0032] Based on the cumulative changes in historical pipeline operation data, the comprehensive score output by the scoring model, and the anomaly level, the fault prediction value is obtained through the fault early warning analysis model. Then, the fault early warning level is divided according to the comparison between the fault prediction value and the fault prediction interval. The fault early warning levels are blue, yellow, and red from low to high.

[0033] Differentiated early warning measures are implemented for different fault levels, including audible and visual alarms, SMS notifications, and automatic shutdown.

[0034] In a further embodiment, the expression for the fault early warning analysis model is:

[0035]

[0036] In the formula, σ(S) represents the transformation result of the Sigmoid function on the comprehensive score S. ρ1, ρ2, and ρ3 represent exponential coefficients, C represents the cumulative change in historical pipeline operation data, ta and tb represent the start and end times respectively, and x i (t) represents the value of the i-th pipeline operating parameter at time t, ψ i Let L represent the weighting factor corresponding to the change in the i-th pipeline operating parameter, L represent the anomaly level (1 for low level, 2 for medium level, and 3 for high level), Γ represent the intermediate impact index, and ξ represent the intermediate impact index. j ω represents the comprehensive score characteristic value obtained from historical data statistics. j represents the weight factor of the j-th Sigmoid function variant, m represents the number of Sigmoid function variants, and P represents the fault prediction value;

[0037] When P exceeds the fault prediction range, the fault warning level is red; when P falls within the fault prediction range, the fault warning level is yellow; when P is below the fault prediction range, the fault warning level is blue.

[0038] In a further embodiment, the model development module includes the following models:

[0039] The residual model is suitable for data that are centrally distributed or positively deviate from the central distribution.

[0040] Isolation forests are suitable for data with a positive long-tail distribution.

[0041] Support vector machines are suitable for data that are negatively skewed from the central distribution or have a negative long-tailed distribution.

[0042] The beneficial effects of this invention are:

[0043] (1) Data is collected through multiple channels, including SCADA system, simulation environment, intermediate database, API transmission and manual inspection data entry, to obtain more comprehensive pipeline operation information; in the data preprocessing stage, outliers and missing values ​​are identified and processed using statistical analysis and machine learning methods, and data scale is unified by standardization or normalization functions to improve data quality; when constructing the scoring model, the deviation of multiple parameters such as pressure and flow rate from the normal range is comprehensively considered, and the data distribution type and anomaly level are determined by combining dynamic threshold judgment mechanism, which can more accurately assess the probability of failure, effectively reduce false alarms, and greatly improve the accuracy of fault diagnosis;

[0044] (2) The data acquisition module acquires pipeline operation data in real time, and the data analysis module quickly builds a scoring model and determines the data distribution type and anomaly level. The real-time early warning module obtains the fault prediction value through the fault early warning analysis model based on the cumulative change of historical pipeline operation data, the comprehensive score output by the scoring model and the anomaly level. Based on the comparison between the predicted value and the fault prediction range, the fault early warning level is divided. It can issue early warning information such as audible and visual alarms and SMS notifications in the early stage of the fault. For serious faults, it can also automatically shut down, realizing real-time monitoring and rapid early warning of pipeline faults, effectively avoiding serious leakage accidents and ecological disasters caused by failure to detect faults in time.

[0045] (3) Select the optimal model based on the data distribution characteristics. For example, the residual model is suitable for data with a concentrated distribution or positive deviation from the concentrated distribution, the isolated forest is suitable for data with a positive long tail distribution, and the support vector machine is suitable for data with a negative deviation from the concentrated distribution or negative long tail distribution. This can give full play to the advantages of different models, better adapt to complex pipeline network operation data, improve the ability to diagnose various types of faults, and provide more reliable technical support for the safe and stable operation of the pipeline network. Attached Figure Description

[0046] The invention will now be further described with reference to the accompanying drawings.

[0047] Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1 As shown, the present invention is a fault diagnosis system for pipeline networks, comprising:

[0050] The data acquisition module is used to obtain pipeline operation data from SCADA systems, simulation environments, intermediate databases, API transmission, and manual inspection data entry.

[0051] The data preprocessing module is used to clean, normalize, and structure the collected data, including data usability testing, basic statistics calculation, and box plotting. For example, the cleaning process uses statistical analysis and machine learning methods, such as the Isolation Forest algorithm, to identify and handle outliers and missing values; normalization uses standardization or normalization functions to unify the data to the same scale range. Data usability testing checks data integrity to ensure its usability; this is done by setting data validity rules such as numerical range and data type. Basic statistics calculation calculates statistics such as mean, median, and standard deviation to quantify data characteristics; the mean reflects the central location of the data, the median is not sensitive to outliers, and the standard deviation measures the dispersion of the data. Mean: Calculates the average value of the data, reflecting its central location; Median: The middle value after arranging the data in ascending order, not sensitive to outliers; Standard deviation: Measures the dispersion of the data; the larger the standard deviation, the more dispersed the data; Range: The difference between the maximum and minimum values, reflecting the dispersion of the data. Box plotting: Visually displays data distribution and identifies outliers; box plots show the quartiles and outliers of the data, helping to quickly determine whether there are outliers deviating from the positive or negative direction; Quartiles: After arranging the data in order of size, divide it into four equal parts: Q1 (25th quartile), Q2 (50th quartile, i.e., median), and Q3 (75th quartile); IQR: Interquartile Range, Q3-Q1, reflects the dispersion of the data; Outliers: Data points that exceed the upper edge (Q3+1.5IQR) or lower edge (Q1-1.5IQR) may be outliers.

[0052] The data analysis module constructs a scoring model based on pipeline operation data, and then determines the data distribution type based on the scoring model combined with a dynamic threshold determination mechanism.

[0053] The model development module selects the optimal model based on data distribution characteristics, supports the integration of multiple models, generates and evaluates the results of multiple models, and outputs visualized model results using a hybrid modeling strategy. When selecting a model, residual models can be chosen if the data is centrally distributed or positively deviates from the central distribution; isolated forests are chosen if the data has a positive long-tail distribution; and support vector machines are chosen if the data has a negative deviation from the central distribution or a negative long-tail distribution. Regarding the integration of multiple models, stacked ensembles use the outputs of multiple base models as new features input to another model for training; hybrid models combine the advantages of different models, such as combining the learning ability of neural networks with the interpretability of decision trees. The selected model is trained using training data, and its accuracy, recall, and other metrics are evaluated using test data. The model is then optimized based on the evaluation results.

[0054] The real-time early warning module monitors the pipeline network's operational status in real time and issues early warning information based on diagnostic results, displaying the information visually. When the predicted fault value exceeds the predicted fault range, the early warning level is red, triggering audible and visual alarms and SMS notifications; for severe faults, it automatically shuts down the system. When the predicted fault value falls within the predicted fault range, the early warning level is yellow, issuing corresponding alerts. When the predicted fault value is below the predicted fault range, the early warning level is blue, providing general alerts. Simultaneously, the pipeline network's operational status and early warning information are displayed visually through an interface, which can be designed using dashboards, map markers, or other methods to intuitively present the real-time condition of the pipeline network.

[0055] This invention collects data through multiple channels to ensure the comprehensiveness and diversity of the data, providing rich information for accurate diagnosis; data preprocessing improves data quality, making subsequent analysis more reliable; constructing a scoring model and determining the data distribution type helps to accurately assess the pipeline network's operating status; the model development module selects and integrates the optimal model, improving the accuracy and adaptability of fault diagnosis; the real-time early warning module enables real-time monitoring and visual early warning, promptly detecting faults and notifying relevant personnel, reducing the risk of accidents.

[0056] The model development module employs convolutional neural networks, recurrent neural networks and their variants, and gradient boosting tree models for model selection, and uses stacked ensemble and hybrid models for model integration. Specifically, stacked ensemble involves: first, dividing historical pipeline operation data into training and testing sets; training different base models such as CNN, RNN, and gradient boosting tree models on the training set to obtain their respective prediction results; then, using the prediction results of these base models as new features to construct a meta-model such as a logistic regression model or a multilayer perceptron; on the testing set, the base model first makes predictions, and then the prediction results are input into the meta-model for secondary prediction, combining the advantages of multiple base models to improve the overall fault diagnosis accuracy.

[0057] Hybrid model: Divide the historical data into multiple non-overlapping subsets and train different models such as CNN, RNN, and gradient boosting tree models on these subsets; dynamically select the appropriate model or model combination for fault diagnosis based on the characteristics of the data or the current operating environment; for example, when the data fluctuates greatly, prioritize the use of models that are sensitive to changes; when the data is relatively stable, select models with better stability, so as to obtain better diagnostic results under different operating conditions.

[0058] In this invention, convolutional neural networks excel at processing spatial features of data, recurrent neural networks and their variants have good processing capabilities for time series data, and gradient boosting tree models perform well in handling complex nonlinear relationships. Multiple model selection and integration methods can fully leverage the advantages of different models, better adapt to the complexity of pipeline network operation data, and improve the accuracy and reliability of fault diagnosis.

[0059] The pipeline operation data includes pressure, flow rate, temperature, vibration frequency, fluid density, corrosion rate, valve opening, pump speed, acoustic signal, and gas concentration parameters.

[0060] The process of constructing the scoring model is as follows:

[0061] The average values ​​of various parameters were calculated based on historical pipeline operation data. and standard deviation σ i This allows us to obtain the normal range of each parameter. i represents the i-th parameter;

[0062] Calculate each parameter x i degree of deviation d i ,when At that time, d i =0, when hour, when hour, The overall score S is obtained based on the weight allocation, and its expression is: θ i The weight coefficients corresponding to each parameter are determined using the analytic hierarchy process (AHP), where n represents the total number of parameters. After constructing the scoring model, cross-validation is used to validate the model. The specific process is as follows: historical pipeline operation data is divided into multiple subsets. Each time, one subset is selected as the test set, and the remaining subsets are used as the training set. The scoring model is trained and evaluated on the test set. The model's performance is evaluated by calculating its precision and recall on multiple test sets. If the model's performance does not meet expectations, the reasons are analyzed and optimizations are made. Optimization measures include adjusting parameter weights and adding or deleting some parameters. After multiple validations and optimizations, the scoring model can stably and accurately reflect the degree of abnormality in the pipeline network's operating status.

[0063] By combining a dynamic threshold determination mechanism, the data distribution type and anomaly level are determined, including concentrated distribution, positive deviation from concentration, positive long tail, negative deviation from concentration, and negative long tail.

[0064] Based on the comprehensive score and the fault prediction value, the fault probability F is output; F = a × P + (1-a) × S. a is a weighting coefficient, obtained based on historical data analysis.

[0065] This invention considers multiple pipeline operating parameters to comprehensively reflect the pipeline network's operating status; it determines the normal range by calculating the mean and standard deviation to accurately measure the degree of parameter deviation, and combines weights to obtain a comprehensive score, making the score more scientific; it determines the data distribution type and anomaly level, enabling a more detailed assessment of the pipeline network's operating status, and outputs the failure probability to provide a quantitative basis for maintenance decisions.

[0066] The dynamic threshold determination mechanism includes:

[0067] Obtain the standard deviation threshold, interquartile range threshold, and data centrality threshold from historical pipeline operation data;

[0068] Based on the real-time range R of the data i =max(x i )-min(x i ) and historical data average range The threshold ratio is dynamically adjusted using the formula: Adjusted threshold = Initial threshold × [1 + k × (Real-time data range / Historical data average range)], which calculates the adjusted standard deviation threshold, interquartile range threshold, and data concentration threshold, where k is the adjustment coefficient, determined through experiments or experience.

[0069] The final data distribution type is determined based on the adjusted standard deviation threshold, interquartile range threshold, and data centrality threshold; then, the comprehensive score output by the scoring model is compared with the preset scoring interval.

[0070] When the overall score is lower than the score range and the data distribution is concentrated, the anomaly level is low; when the overall score is higher than the score range and the data distribution is positively or negatively deviating from the concentration, the anomaly level is high; otherwise, the anomaly level is medium.

[0071] In this invention, the dynamic threshold determination mechanism can dynamically adjust the threshold based on changes in real-time and historical data, more flexibly adapting to the dynamic changes in pipeline network operation. It accurately determines the data distribution type and anomaly level, improving the accuracy and timeliness of fault diagnosis and providing an accurate basis for taking appropriate maintenance measures.

[0072] The logic for determining the data distribution type is as follows:

[0073] When the absolute value of the difference between the mean and the median is less than or equal to the median, and the standard deviation and interquartile range are both lower than the corresponding standard deviation threshold and interquartile range threshold, the data type is determined to be a concentrated distribution.

[0074] When the data shows positive skewness and the proportion of data exceeding the upper bound is higher than the concentration threshold, the data type is determined to be positively skewed from concentration.

[0075] When the data is positively skewed, but the proportion of data exceeding the upper bound is lower than the concentration threshold, the data type is determined to be positive long-tailed.

[0076] When the data shows negative skewness and the proportion of data below the lower bound is higher than the concentration threshold, the data type is determined to be negatively skewed from concentration.

[0077] When data exhibits negative skewness, but the proportion of data below the lower bound is below the concentration threshold, the data type is determined to be negatively long-tailed. In this invention, the explicit logic for determining data distribution types provides a unified standard for data analysis, making the process of determining data distribution types more scientific and accurate, and contributing to further improving the reliability of fault diagnosis.

[0078] In a further embodiment, the operation of the real-time early warning module includes:

[0079] Based on the cumulative changes in historical pipeline operation data, the comprehensive score output by the scoring model, and the anomaly level, the fault prediction value is obtained through the fault early warning analysis model. Then, the fault early warning level is divided according to the comparison between the fault prediction value and the fault prediction interval. The fault early warning levels are blue, yellow, and red from low to high.

[0080] Differentiated early warning measures are implemented for different fault levels, including audible and visual alarms, SMS notifications, and automatic shutdown.

[0081] In this invention, fault prediction values ​​are obtained by combining multiple pieces of information, fault warning levels are divided, and differentiated warning measures are taken. This enables timely and accurate communication of fault information to maintenance personnel, allowing them to take corresponding measures based on different fault levels and effectively reduce accident losses.

[0082] In a further embodiment, the expression for the fault early warning analysis model is:

[0083]

[0084] In the formula, σ(S) represents the transformation result of the Sigmoid function on the comprehensive score S. The comprehensive score S comprehensively considers the deviation of multiple pipeline operating parameters, such as pressure and flow rate, from the normal range, fully reflecting the degree of abnormality in the pipeline network's operating status. When multiple parameters deviate from the normal range, the comprehensive score will increase accordingly, indicating that the pipeline network's abnormality is relatively serious. σ(S) maps the comprehensive score S to the (0,1) interval, performing a nonlinear transformation on the comprehensive score to highlight the different impacts of different score intervals on fault prediction values. For example, when the comprehensive score is low, its value is close to 0 after transformation by the Sigmoid function; when the comprehensive score is high, the transformed value is close to 1, making the model more sensitive to comprehensive scores of different degrees of abnormality. ρ1, ρ2, and ρ3 represent exponential coefficients, and C represents the cumulative change in historical pipeline operating data, reflecting the changing trend of pipeline operating parameters over time. For example, for pressure parameters, if the pressure continues to rise over a period of time, the value of C will increase accordingly, indicating that the pressure change trend has a significant impact on fault prediction. ta and tb represent the start time and end time, respectively, and x... i (t) represents the value of the i-th pipeline operating parameter at time t, ψ i Γ represents the weighting factor corresponding to the change in the i-th pipeline operating parameter, L represents the anomaly level, with 1 for low level, 2 for medium level, and 3 for high level, intuitively reflecting the severity of pipeline network anomalies and providing an important reference for fault prediction; a low anomaly level indicates that the pipeline network operating status is slightly abnormal, but may not affect normal operation; a high anomaly level indicates that there is a serious anomaly in the pipeline network, and the probability of a fault is high. Γ represents the intermediate impact index, ξ j ω represents the comprehensive score characteristic value obtained from historical data statistics. j represents the weight factor of the j-th Sigmoid function variant, m represents the number of Sigmoid function variants, and P represents the fault prediction value;

[0085] When fault prediction P exceeds the fault prediction range, the fault warning level is red; when P falls within the fault prediction range, the fault warning level is yellow; when P is below the fault prediction range, the fault warning level is blue. By multiplying multiple Sigmoid function variants with the comprehensive score as the variable, the complex influence of the comprehensive score on the fault prediction value under different value ranges and trends is captured. After processing the anomaly level, an exponential calculation is performed to highlight the importance of the anomaly level in fault prediction. Multiplying the nonlinear transformation result of the comprehensive score with the logarithmic transformation result of the cumulative change and performing an exponential operation demonstrates the combined effect of the comprehensive score and the cumulative change.

[0086] In this invention, the fault early warning analysis model comprehensively considers multiple factors such as overall score, cumulative changes in historical data, and anomaly level. It calculates fault prediction values ​​more accurately using the aforementioned expression, and classifies fault early warning levels based on these prediction values, providing a more scientific basis for real-time early warning and further improving the accuracy and reliability of fault early warning.

[0087] In a further embodiment, the model development module includes the following models:

[0088] The residual model is suitable for data that are centrally distributed or positively deviate from the central distribution.

[0089] Isolation forests are suitable for data with a positive long-tail distribution.

[0090] Support vector machines are suitable for data that are negatively skewed from the central distribution or have a negative long-tailed distribution.

[0091] In this invention, selecting appropriate models for different data distribution types can fully leverage the advantages of each model, improve the model's ability to identify different fault modes, and thus enhance the diagnostic effect of the entire fault diagnosis system.

[0092] It should be noted that the calculation formulas and all parameters involved in the calculations in this invention have been dimensionless beforehand. The process of dimensionless processing is well known in the industry and will not be described here.

[0093] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A fault diagnosis system for a pipeline network, characterized in that, include: The data acquisition module is used to obtain pipeline operation data from SCADA systems, simulation environments, intermediate databases, API transmission, and manual inspection data entry. The data preprocessing module is used to clean, normalize, and structure the collected data, including data availability testing, basic statistics calculation, and box plot drawing. The data analysis module constructs a scoring model based on pipeline operation data, and then determines the data distribution type based on the scoring model combined with a dynamic threshold determination mechanism. The model development module selects the optimal model based on data distribution characteristics, supports the integration of multiple models, generates and evaluates the results of multiple models, and outputs visualized model results using a hybrid modeling strategy. The real-time early warning module is used to monitor the pipeline network operation status in real time and issue early warning information based on the diagnostic results, which are then displayed visually. The model development module uses convolutional neural networks, recurrent neural networks and their variants, and gradient boosting tree models for model selection, and employs stacked ensemble and hybrid models for model integration. The pipeline operation data includes pressure, flow rate, temperature, vibration frequency, fluid density, corrosion rate, valve opening, pump speed, acoustic signal, and gas concentration parameters. The process of constructing the scoring model is as follows: The average values ​​of various parameters were calculated based on historical pipeline operation data. and standard deviation σ i This allows us to obtain the normal range of each parameter. i represents the i-th parameter; Calculate each parameter x i degree of deviation d i ,when At that time, d i =0, when hour, when hour, The overall score S is obtained based on the weight allocation, and its expression is: θ i This represents the weight coefficients corresponding to each parameter, and n represents the total number of parameter terms. By combining a dynamic threshold determination mechanism, the data distribution type and anomaly level are determined, including concentrated distribution, positive deviation from concentration, positive long tail, negative deviation from concentration, and negative long tail. Output the failure probability based on the comprehensive score and the failure prediction value; The dynamic threshold determination mechanism includes: Obtain the standard deviation threshold, interquartile range threshold, and data centrality threshold from historical pipeline operation data; Based on the real-time range R of the data i =max(x i )-min(x i ) and historical data average range The threshold ratio is dynamically adjusted using the formula: Adjusted threshold = Initial threshold × [1 + k × (Real-time data range / Historical data average range)], which calculates the adjusted standard deviation threshold, interquartile range threshold, and data concentration threshold, where k is the adjustment coefficient, determined through experiments or experience. The final data distribution type is determined based on the adjusted standard deviation threshold, interquartile range threshold, and data centrality threshold; then, the comprehensive score output by the scoring model is compared with the preset scoring interval. When the overall score is lower than the score range and the data distribution is concentrated, the anomaly level is low; when the overall score is higher than the score range and the data distribution is positively or negatively deviating from the concentration, the anomaly level is high; otherwise, the anomaly level is medium.

2. The pipeline fault diagnosis system according to claim 1, characterized in that, The logic for determining the data distribution type is as follows: When the absolute value of the difference between the mean and the median is less than or equal to the median, and the standard deviation and interquartile range are both lower than the corresponding standard deviation threshold and interquartile range threshold, the data type is determined to be a concentrated distribution. When the data shows positive skewness and the proportion of data exceeding the upper bound is higher than the concentration threshold, the data type is determined to be positively skewed from concentration. When the data is positively skewed, but the proportion of data exceeding the upper bound is lower than the concentration threshold, the data type is determined to be positive long-tailed. When the data shows negative skewness and the proportion of data below the lower bound is higher than the concentration threshold, the data type is determined to be negatively skewed from concentration. When the data shows negative skewness, but the proportion of data below the lower bound is lower than the concentration threshold, the data type is determined to be negative long-tailed.

3. The pipeline fault diagnosis system according to claim 1, characterized in that, The operation process of the real-time early warning module includes: Based on the cumulative changes in historical pipeline operation data, the comprehensive score output by the scoring model, and the anomaly level, the fault prediction value is obtained through the fault early warning analysis model. Then, the fault early warning level is divided according to the comparison between the fault prediction value and the fault prediction interval. The fault early warning levels are blue, yellow, and red from low to high. Differentiated early warning measures are implemented for different fault levels, including audible and visual alarms, SMS notifications, and automatic shutdown.

4. The pipeline fault diagnosis system according to claim 3, characterized in that, The expression for the fault early warning analysis model is: In the formula, σ(S) represents the transformation result of the Sigmoid function on the comprehensive score S. ρ1, ρ2, and ρ3 represent exponential coefficients, C represents the cumulative change in historical pipeline operation data, ta and tb represent the start and end times respectively, and x i (t) represents the value of the i-th pipeline operating parameter at time t, ψ i Let L represent the weighting factor corresponding to the change in the i-th pipeline operating parameter, L represent the anomaly level (1 for low level, 2 for medium level, and 3 for high level), Γ represent the intermediate impact index, and ξ represent the intermediate impact index. j ω represents the comprehensive score characteristic value obtained from historical data statistics. j represents the weight factor of the j-th Sigmoid function variant, m represents the number of Sigmoid function variants, and P represents the fault prediction value; When P exceeds the fault prediction range, the fault warning level is red; when P falls within the fault prediction range, the fault warning level is yellow; when P is below the fault prediction range, the fault warning level is blue.

5. The pipeline fault diagnosis system according to claim 1, characterized in that, The models in the model development module include: The residual model is suitable for data that are centrally distributed or positively deviate from the central distribution. Isolation forests are suitable for data with a positive long-tail distribution. Support vector machines are suitable for data that are negatively skewed from the central distribution or have a negative long-tailed distribution.

Images