Flowmeter fault diagnosis system based on artificial intelligence

Abstract

The invention discloses a flowmeter fault diagnosis system based on artificial intelligence, and relates to the technical field of flow metering and artificial intelligence. The system comprises the following components: a data acquisition and preprocessing module, a feature analysis and scene recognition module, a multi-algorithm library and intelligent decision module, a diagnosis execution and result output module and a self-optimization learning module. According to the method, multiple advanced artificial intelligence diagnosis algorithms are integrated, and dynamic weight distribution, an adaptive adjustment strategy and a dynamic confidence fusion algorithm are combined, so that the accuracy of fault diagnosis is remarkably improved, and compared with a single algorithm of a CNN and GBDT fusion model, the fault diagnosis accuracy is improved by 15-20%; according to the random forest algorithm, under the self-adaptive adjustment strategy, the diagnosis speed in a low-dimensional fuzzy scene is increased by 30%, meanwhile, the accuracy is kept to be 94% or above, due to the improvements, the system can recognize flowmeter faults more rapidly and accurately, misjudgment and missed judgment are reduced, and the overall operation and maintenance efficiency is improved.

Description

Technical Field

[0001] This invention relates to the fields of flow metering and artificial intelligence technology, specifically to a flow meter fault diagnosis system based on artificial intelligence. Background Technology

[0002] With the rapid development of industrial automation and intelligent manufacturing, flow meters, as an indispensable key device in industrial production, directly affect production efficiency and product quality in terms of accuracy and stability. However, flow meters inevitably experience various malfunctions during long-term operation, such as sensor damage, signal interference, and data drift. These malfunctions not only lead to inaccurate measurement data but may also cause production accidents and result in serious economic losses.

[0003] Traditional flowmeter fault diagnosis technologies suffer from several drawbacks: First, they rely on human experience for fault diagnosis, which is highly subjective and susceptible to human error, leading to inconsistent and inaccurate results. Second, regular maintenance and inspections cannot monitor equipment status in real time, making it difficult to detect sudden faults and increasing the risk of unplanned downtime. Third, traditional methods struggle with complex operating conditions and large-scale data scenarios, failing to extract effective features from massive amounts of data for accurate diagnosis. Finally, traditional diagnostic systems lack self-learning and optimization capabilities, failing to dynamically adjust diagnostic strategies based on actual operating conditions, resulting in decreased diagnostic efficiency and increased maintenance costs over the long term. Specifically, traditional methods are significantly inadequate in terms of flowmeter vibration signal noise suppression, fault feature retention, and diagnostic speed and accuracy in low-dimensional fuzzy scenarios, making it difficult to meet the demands of modern industry for efficient and accurate fault diagnosis.

[0004] In view of the problems of low efficiency, limited accuracy, lack of real-time performance and self-learning ability in traditional flow meter fault diagnosis technology, it is particularly important to develop an artificial intelligence-based flow meter fault diagnosis system. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an artificial intelligence-based flow meter fault diagnosis system. By integrating multiple advanced artificial intelligence diagnostic algorithms and combining dynamic weight allocation, adaptive adjustment strategies, and dynamic confidence fusion algorithms, it significantly improves the accuracy and real-time performance of fault diagnosis. At the same time, the system's built-in self-optimizing learning module continuously optimizes the algorithm selection strategy through reinforcement learning mechanisms, thereby achieving a comprehensive performance improvement in the algorithm selection strategy and further reducing the reliance on human experience and long-term operation and maintenance costs.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an artificial intelligence-based flow meter fault diagnosis system, which includes the following components: a data acquisition and preprocessing module, a feature analysis and scene recognition module, a multi-algorithm library and intelligent decision-making module, a diagnosis execution and result output module, and a self-optimizing learning module; Data acquisition and preprocessing module: used to acquire multi-dimensional real-time operating data of the flow meter and perform preprocessing. The multi-dimensional real-time operating data includes flow rate, pressure, temperature, vibration acceleration, and acoustic signals. The acquisition frequency is not less than 1kHz. The preprocessing includes filtering, noise reduction, and normalization, and adopts a variable coefficient filtering algorithm. Feature analysis and scene recognition module: connected to the data acquisition and preprocessing module, used to extract three types of feature indicators: data dimension, noise level, and fault feature complexity, and to identify the operating scene. The fault feature complexity indicator is calculated by feature entropy value, and the total number of features N is determined to be 20 based on the feature distribution analysis of 2000+ flow meter fault and normal data. Multi-algorithm library and intelligent decision-making module: including a multi-algorithm library storing various artificial intelligence diagnostic algorithms, and an algorithm decision engine that dynamically selects or combines algorithms based on the feature indicators. The multi-algorithm library includes a CNN and GBDT fusion model, a random forest algorithm, and an LSTM algorithm based on an attention mechanism. The CNN and GBDT fusion model uses a dynamic weight allocation algorithm to calculate the fusion weights, the random forest algorithm uses an adaptive adjustment strategy to determine the number of decision trees and the feature selection method, and the multi-algorithm weighted combination uses a dynamic confidence fusion algorithm. Diagnosis execution and result output module: Connects the multi-algorithm library and intelligent decision module, used to perform fault diagnosis and output results, and can output fault root cause analysis. It associates the constructed fault knowledge graph with the diagnosis results. The entity association strength calculation of the knowledge graph adopts a multi-dimensional knowledge association reasoning algorithm. The self-optimizing learning module is connected to the multi-algorithm library, the intelligent decision-making module, and the diagnostic execution and result output module, respectively. It is used to optimize the algorithm selection strategy through reinforcement learning, which adopts a multi-objective reward function.

[0007] Furthermore, the multi-dimensional real-time operating data in the data acquisition and preprocessing module includes flow rate, pressure, temperature, vibration acceleration, and acoustic signals, with an acquisition frequency of no less than 1kHz. Preprocessing includes filtering, noise reduction, and normalization, employing a variable coefficient filtering algorithm, the formula of which is: ,in , The threshold value was determined based on the statistical values ​​of noise intensity from 500 sets of flowmeter data under different operating conditions. The value ranged from 0.1 to 0.3, with an optimal value of 0.2. This is the current sampled value. The previous sampled value, This is the current output value. The previous output value is used. This algorithm effectively suppresses noise while preserving fault characteristics by dynamically adjusting the filtering coefficients. According to tests, compared with traditional Kalman filtering, the noise suppression effect of flowmeter vibration signal is improved by 40%, and the fault characteristic preservation rate is improved by 25%.

[0008] Furthermore, the fault feature complexity index in the feature analysis and scene recognition module is calculated using feature entropy values, with the following formula: ,in , For the first The frequency of occurrence of each feature This represents the total number of features; the entropy value is calculated based on the feature distribution analysis of over 2000 flowmeter fault and normal data points. Taking 20, covering core characteristic dimensions such as flow rate, pressure, and vibration, this indicator can be used to clearly categorize the complexity of fault characteristics. ,Vague ,overlapping The three categories provide accurate basis for algorithm decision-making, enabling scene recognition accuracy to reach over 98%.

[0009] Furthermore, the fusion weight calculation of the CNN and GBDT fusion model in the multi-algorithm library and intelligent decision-making module adopts a dynamic weight allocation algorithm, the formula of which is: , ,in The weighting formula is determined based on algorithm fusion experiments of 100 sets of high-dimensional data scenarios, where the entropy value is the fault characteristic. hour, , It focuses on the high-dimensional feature extraction capabilities of CNNs; when When the value deviates from 1.0, the weights are tilted towards GBDT to enhance classification stability. According to tests, this fusion model improves the fault diagnosis accuracy by 15-20% compared to a single CNN or GBDT, and reaches more than 195% in high-dimensional slightly noisy scenarios.

[0010] Furthermore, the number of decision trees and feature selection method of the random forest algorithm in the multi-algorithm library and intelligent decision module adopt an adaptive adjustment strategy, as shown in the formula: , ,in For the number of decision trees, The fault characteristic entropy value, For data dimensions, The number of features selected for each tree. The current data noise level, converted from the signal-to-noise ratio, takes a value between 0 and 1. To achieve the maximum noise level; this strategy is optimized based on experimental data from 80 sets of low-dimensional blurred scenes, when , , hour, , This improves the diagnostic speed of random forests in low-dimensional fuzzy scenarios by 30%, while maintaining an accuracy of over 94%, significantly outperforming random forest algorithms with fixed parameters.

[0011] Furthermore, the reinforcement learning in the self-optimizing learning module employs a multi-objective reward function, the formula of which is: ,in The diagnostic accuracy is represented by a value between 0 and 1. The diagnosis takes time. The maximum allowed time is set to 2 seconds. The penalty coefficient for misjudgment; when misjudgment occurs. ,otherwise The parameters of this function were determined through cost-benefit analysis of more than 1,000 diagnostic tasks. The weights of 0.7, 0.3, and 0.5 correspond to the importance of accuracy, time consumption, and misjudgment, respectively. This ensures that the system prioritizes accuracy during the optimization process while also taking into account time consumption and avoiding misjudgment. After 100,000 iterations of optimization, the overall performance of the algorithm selection strategy has been improved by 25%.

[0012] Furthermore, the multi-algorithm library in the intelligent decision-making module also includes an LSTM algorithm based on an attention mechanism, used to process the temporal features of traffic data. Its attention weight calculation employs a traffic feature attention algorithm, with the following formula: , ,in For the first Attention weight at any moment For the output of the LSTM hidden layer, For the first Flow characteristics at any given time , The weight matrix has dimensions of 256×128 and was trained based on 500 sets of traffic time-series fault data. The bias vector is used; this algorithm improves LSTM's attention to time-series features such as traffic mutations and trend anomalies by 40%, and achieves a diagnostic accuracy of 96% for traffic fluctuation faults, which is 18% higher than that of ordinary LSTM.

[0013] Furthermore, the multi-algorithm weighted combination strategy in the multi-algorithm library and intelligent decision-making module adopts a dynamic confidence fusion algorithm. The specific steps are as follows: first, obtain the fault category prediction confidence of each participating algorithm; then, calculate the dynamic weight based on the historical accuracy and time consumption of each algorithm; finally, sum the confidence of each algorithm according to the weights, and select the category corresponding to the maximum value as the final diagnosis result.

[0014] Furthermore, the diagnostic execution and result output module can also output root cause analysis of faults. By associating the constructed fault knowledge graph with the diagnostic results, the entity association strength calculation of the knowledge graph adopts a multi-dimensional knowledge association reasoning algorithm. The specific steps are as follows: first, extract the fault entities in the diagnostic results and the candidate entities in the knowledge graph; then, calculate the association scores of the three dimensions of semantic similarity, co-occurrence frequency, and expert annotation score respectively; finally, obtain the association strength between entities by weighted summation, thereby inferring the root cause of the fault and outputting maintenance suggestions.

[0015] Compared with existing technologies, this AI-based flow meter fault diagnosis system has the following advantages: I. This system significantly improves the accuracy of fault diagnosis by integrating multiple advanced artificial intelligence diagnostic algorithms and combining dynamic weight allocation, adaptive adjustment strategies, and dynamic confidence fusion algorithms. Compared with a single algorithm, the CNN and GBDT fusion model improves the fault diagnosis accuracy by 15-20%. Under the adaptive adjustment strategy, the Random Forest algorithm improves the diagnosis speed by 30% in low-dimensional fuzzy scenarios, while maintaining an accuracy of over 94%. These improvements enable the system to identify flow meter faults more quickly and accurately, reduce false positives and false negatives, and improve overall operation and maintenance efficiency.

[0016] Second, this system, through its built-in self-optimizing learning module, employs a reinforcement learning mechanism and a multi-objective reward function to continuously optimize the algorithm selection strategy. This allows the system to maintain diagnostic accuracy while balancing diagnostic time and false positive penalties, achieving a 25% improvement in the overall performance of the algorithm selection strategy. Furthermore, the system can construct a fault knowledge graph and output fault root cause analysis and maintenance suggestions through multi-dimensional knowledge association reasoning algorithms, further reducing reliance on human experience and decreasing the time and cost of on-site troubleshooting. This intelligent decision-making and self-optimization capability enables the system to continuously adapt to the flow meter fault diagnosis needs under different operating conditions, reducing long-term operation and maintenance costs.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a flowchart illustrating the operation of an AI-based flow meter fault diagnosis system. Figure 2 This is an overall flowchart of an AI-based flow meter fault diagnosis system. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1 Data Acquisition and Preprocessing Stage: After the data acquisition and preprocessing module is started, it first acquires five types of multi-dimensional real-time operating data from the flow meter: flow rate, pressure, temperature, vibration acceleration, and acoustic signals. Flow rate data reflects the stability of the crude oil delivery rate; pressure data reflects changes in pipeline delivery resistance; temperature data helps determine if there are flow problems caused by abnormal crude oil viscosity; and vibration acceleration and acoustic signals capture the operating status of the internal mechanical components of the flow meter. The acquisition frequency is strictly set to 1kHz to meet the parameter requirement of not less than 1kHz, ensuring high-frequency capture of instantaneous fault signals. Subsequently, the acquired raw data undergoes filtering, noise reduction, and normalization preprocessing. A variable coefficient filtering algorithm is used in this process, with the formula: ,in , The threshold was determined based on the statistical values ​​of noise intensity from 500 sets of flow meter data under different operating conditions. This is the current sampled value. The previous sampled value, This is the current output value. Using the previous output value as an example, the algorithm can dynamically adjust the filtering coefficient according to data fluctuations, effectively filtering out irrelevant noise such as vibration and electromagnetic interference in the surrounding environment of the pipeline, while retaining the true fault characteristic signals, avoiding noise from masking key data, providing a high-quality and highly reliable data foundation for subsequent feature analysis and diagnosis, and reducing diagnostic errors caused by data quality issues.

[0022] Feature Analysis and Scene Recognition Stage: The feature analysis and scene recognition module connects to the data acquisition and preprocessing module, receiving the preprocessed clean data. This module first extracts three types of feature indicators: data dimension, noise level, and fault feature complexity. The data dimension indicator clarifies the completeness and coverage of the current analysis data, ensuring no key monitoring dimensions are missing. The noise level indicator assesses the degree of residual interference in the preprocessed data, determining whether it meets the input requirements of subsequent diagnostic algorithms. The fault feature complexity indicator is calculated using feature entropy values, with the following formula: ,in , For the first The frequency of occurrence of each feature The total number of features is determined by N = 20, based on the feature distribution analysis of over 2000 flowmeter fault and normal data. This indicator can quantify the richness and complexity of fault features in the current data, providing a core basis for subsequent algorithm selection. At the same time, the module combines the actual working conditions of oil pipeline transportation to accurately identify the operating scenario of the flowmeter, such as determining whether it is currently in a normal high-pressure transportation scenario, a low-load maintenance scenario, or a crude oil batch change transition scenario. The normal operating parameter range and fault mode of the flowmeter differ under different scenarios. The scenario identification results can help the subsequent algorithm to better fit the actual working conditions, improve the diagnostic targeting, and avoid fault omissions or misjudgments caused by scenario misjudgments.

[0023] Multi-algorithm selection and decision-making stage: The multi-algorithm library and intelligent decision-making module play a core decision-making role. The library includes a CNN and GBDT fusion model, a random forest algorithm, and an attention-based LSTM algorithm, capable of addressing different types of fault diagnosis needs in oil pipeline flow meters. The algorithm decision engine dynamically selects or combines algorithms based on the extracted feature indicators and scene recognition results: when the fault features are complex and the data contains high-dimensional features, the CNN and GBDT fusion model is preferentially selected. This model uses a dynamic weight allocation algorithm to calculate the fusion weights, as shown in the formula: , ,in The fault characteristic entropy value; when hour, , It focuses on the high-dimensional feature extraction capabilities of CNNs; when When the value deviates from 1.0, the weights are tilted towards GBDT to enhance classification stability. This not only allows CNN to efficiently extract high-dimensional spatial features from the data, but also enhances classification stability through GBDT, avoiding diagnostic bias of a single algorithm under complex features.

[0024] When the data noise level is low but rapid diagnosis is required, the Random Forest algorithm is chosen, and the formula is: , ,in For the number of decision trees, The fault characteristic entropy value, For data dimensions, The number of features selected for each tree. Current data noise level To maximize noise levels, the algorithm employs an adaptive adjustment strategy to determine the number of decision trees and the feature selection method. This optimizes the model structure based on the current data dimensionality and noise level, improving computational efficiency while maintaining diagnostic accuracy. When it's crucial to monitor the temporal changes in traffic data, an attention-based LSTM algorithm is activated, as shown in the formula: , ,in For the first Attention weight at any moment For the output of the LSTM hidden layer, For the first Flow characteristics at any given time , This is the weight matrix. As a bias vector, its attention weight calculation adopts the flow feature attention algorithm, which can automatically focus on the key time segment for fault diagnosis, improve the diagnostic accuracy of time series data. If a single algorithm cannot cover complex fault scenarios, the module adopts a dynamic confidence fusion algorithm, which combines the historical accuracy and real-time consumption of each algorithm to calculate dynamic weights, and performs weighted fusion of the prediction results of multiple algorithms to further reduce diagnostic errors and ensure decision reliability.

[0025] Diagnosis Execution and Result Output Phase: The diagnosis execution and result output module connects to multiple algorithm libraries and the intelligent decision-making module. It executes the fault diagnosis process according to the selected algorithm or algorithm combination. First, it calls the corresponding algorithm model to perform feature matching and fault classification on the preprocessed data. Then, it confirms the consistency of the diagnosis results through an internal verification mechanism. After the diagnosis is completed, the module not only outputs clear fault diagnosis results but also conducts root cause analysis by associating the constructed fault knowledge graph with the diagnosis results. The entity association strength calculation of the knowledge graph adopts a multi-dimensional knowledge association reasoning algorithm. First, it extracts the fault entities in the diagnosis results and candidate entities in the knowledge graph. Then, it calculates the association score from three dimensions: semantic similarity, co-occurrence frequency, and expert annotation score. Finally, it determines the association strength between entities through weighted summation, accurately locating the root cause of the fault. Simultaneously, the module outputs targeted maintenance suggestions based on the root cause and presents the diagnosis results and root cause analysis in the form of visual reports. This facilitates pipeline maintenance personnel to quickly understand the fault situation, shortens fault investigation and maintenance time, and reduces the impact on oil transportation.

[0026] Self-optimization learning phase: The self-optimization learning module connects to multiple algorithm libraries and the intelligent decision-making module, as well as the diagnostic execution and result output module, to continuously iterate and optimize the system's diagnostic performance. This module uses reinforcement learning to optimize algorithm selection strategies and employs a multi-objective reward function during reinforcement learning—using diagnostic accuracy (A) as the core reward indicator—to ensure that the optimization direction always prioritizes improving diagnostic accuracy. Simultaneously, it introduces a negative correlation factor for diagnostic time (T) to avoid excessively increasing computation time in pursuit of high accuracy, ensuring the system's response speed in real-time pipeline monitoring scenarios. Furthermore, it sets a misjudgment penalty coefficient (P) to penalize errors or omissions through a penalty mechanism. To encourage the algorithm selection strategy to avoid similar errors, the module periodically collects historical diagnostic data and calculates the reward value for each algorithm selection based on a multi-objective reward function. It continuously adjusts the weights and logic of algorithm selection. For example, if the LSTM algorithm based on the attention mechanism repeatedly obtains high reward values ​​in a certain scenario, the probability of prioritizing this algorithm in that scenario is increased. If a combination of algorithms frequently misjudges under a specific fault type, its usage frequency is reduced and the fusion weights are optimized. Through continuous self-optimization learning, the system can gradually adapt to changes in operating conditions and fault mode updates in the oil pipeline transportation scenario, maintain high diagnostic performance in the long term, and provide continuous and reliable technical support for the stable operation of pipeline flow meters.

[0027] Example 2 Data Acquisition and Preprocessing Stage: The data acquisition and preprocessing module begins to work, acquiring real-time multi-dimensional operating data of flow, pressure, temperature, vibration acceleration, and acoustic signals from the flow meter in the chemical production process. The acquisition frequency is set to 1.2kHz to meet the parameter requirement of not less than 1kHz. Subsequently, the acquired data is subjected to filtering, noise reduction, and normalization preprocessing operations. The preprocessing process adopts a variable coefficient filtering algorithm to remove noise data caused by various interference factors in the chemical production environment, ensuring the accuracy and reliability of the data.

[0028] Feature analysis and scene recognition stage: The feature analysis and scene recognition module receives the output data from the data acquisition and preprocessing module and extracts three types of feature indicators: data dimension, noise level, and fault feature complexity. The fault feature complexity indicator is calculated using feature entropy value. Based on the feature distribution analysis of 2000+ flow meter fault and normal data, the total number of features N is determined to be 20. At the same time, this module identifies the operating scenario of the flow meter in the chemical production process, such as determining whether it is in a normal production load scenario, a low-load production scenario, or a high-load production scenario.

[0029] Multi-algorithm selection and decision-making stage: The multi-algorithm library and intelligent decision-making module are activated. The multi-algorithm library includes a CNN and GBDT fusion model, a random forest algorithm, and an attention-based LSTM algorithm. The algorithm decision engine dynamically selects or combines algorithms based on the extracted feature indicators. For the complex time-series changes that may exist in the flow data in the chemical production process, the attention-based LSTM algorithm is given priority, and its attention weight calculation adopts the flow feature attention algorithm. If the data features are more complex, the CNN and GBDT fusion model can be selected. This model calculates the fusion weight through a dynamic weight allocation algorithm. The random forest algorithm determines the number of decision trees and the feature selection method according to an adaptive adjustment strategy. If a single algorithm cannot meet the diagnostic needs, a dynamic confidence fusion algorithm is used to perform a weighted combination of multiple algorithms.

[0030] Diagnosis Execution and Result Output Stage: The diagnosis execution and result output module is connected to the multi-algorithm library and intelligent decision-making module to execute fault diagnosis tasks and output results. In addition to basic fault diagnosis results, it can also output root cause analysis of faults. This is achieved by associating the constructed fault knowledge graph with the diagnosis results. The entity association strength calculation of the knowledge graph adopts a multi-dimensional knowledge association reasoning algorithm. Based on the root cause of the fault derived by reasoning, it provides targeted maintenance suggestions to help chemical production plants quickly solve flow meter faults and reduce the impact on the production process.

[0031] Self-optimization learning phase: The self-optimization learning module establishes connections with the multi-algorithm library and intelligent decision-making module, as well as the diagnosis execution and result output module. It uses reinforcement learning to optimize the algorithm selection strategy. Reinforcement learning adopts a multi-objective reward function, comprehensively considering factors such as diagnostic accuracy, diagnostic time, and misjudgment penalty coefficient, and continuously optimizes the algorithm selection logic. This enables the system to continuously improve the accuracy and efficiency of fault diagnosis in the specific scenario of chemical production process, and better ensure the stable operation of chemical production.

[0032] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An artificial intelligence-based flowmeter fault diagnosis system, characterized by, The system comprises the following components: a data acquisition and preprocessing module, a feature analysis and scene recognition module, a multi-algorithm library and intelligent decision module, a diagnosis execution and result output module, and a self-optimizing learning module. The data acquisition and preprocessing module is used to acquire multi-dimensional real-time running data of the flowmeter and perform preprocessing, and adopts a variable coefficient filtering algorithm. The feature analysis and scene recognition module is connected to the data acquisition and preprocessing module and is used to extract three types of feature indexes, i.e., data dimension, noise level, and fault feature complexity, and recognize the running scene, the fault feature complexity index is calculated by a feature entropy value, and the feature total number N is determined to be 20 based on feature distribution analysis of 2000+ flowmeter fault and normal data. The multi-algorithm library and intelligent decision module comprises a multi-algorithm library storing multiple artificial intelligence diagnosis algorithms, and an algorithm decision engine dynamically selecting or combining algorithms according to the feature indexes, the multi-algorithm library contains a CNN and GBDT fusion model, a random forest algorithm, and an LSTM algorithm based on an attention mechanism, the CNN and GBDT fusion model adopts a dynamic weight distribution algorithm to calculate a fusion weight, the random forest algorithm adopts an adaptive adjustment strategy to determine the number of decision trees and the feature selection method, and the multi-algorithm weighted combination adopts a dynamic confidence fusion algorithm. The diagnosis execution and result output module is connected to the multi-algorithm library and intelligent decision module and is used to execute fault diagnosis and output results, and can output fault root cause analysis, the entity correlation strength calculation of the knowledge graph is performed by a multi-dimensional knowledge correlation reasoning algorithm. The self-optimizing learning module is connected to the multi-algorithm library and intelligent decision module and the diagnosis execution and result output module, and is used to select strategies by a reinforcement learning optimization algorithm, the reinforcement learning adopts a multi-objective reward function.

2. The artificial intelligence based flow meter fault diagnosis system of claim 1, wherein, The multi-dimensional real-time running data in the data acquisition and preprocessing module includes flow, pressure, temperature, vibration acceleration, and acoustic signal, the acquisition frequency is not less than 1 kHz, the preprocessing includes filtering, denoising, and normalization, a variable coefficient filtering algorithm is adopted, and the formula is: Wherein , is a threshold value determined based on noise intensity statistical values of 500 groups of flowmeter data in different working conditions, is a current sampling value, is a previous sampling value, is a current output value, is a previous output value.

3. The artificial intelligence based flow meter fault diagnosis system of claim 1, wherein, The feature analysis and scene recognition module calculates the fault feature complexity index through feature entropy value, and the formula is: , wherein , is the frequency of the first feature, is the total number of features.

4. The artificial intelligence based flow meter fault diagnosis system of claim 1, wherein, The dynamic weight distribution algorithm is used for calculating the fusion weight of the CNN and GBDT fusion model in the intelligent decision module of the multi-algorithm library, and the formula is: , , wherein is the fault feature entropy value; When time, , , focusing on the high-dimensional feature extraction ability of CNN; When The weight is tilted to GBDT to enhance the classification stability when deviating from 1.

0.

5. The artificial intelligence based flow meter fault diagnosis system of claim 1, wherein, The adaptive adjustment strategy is adopted for the number of decision trees and the feature selection mode of the random forest algorithm in the multi-algorithm library and the intelligent decision module, and the formula is: , , wherein is the number of decision trees, is the fault feature entropy value, is the data dimension, is the number of selected features for each tree, is the current data noise level is the maximum noise level.

6. The artificial intelligence based flow meter fault diagnosis system of claim 1, wherein, The reinforcement learning in the self-optimizing learning module adopts a multi-objective reward function, and a formula is as follows: , wherein is a diagnostic accuracy rate, is a diagnostic time consumption, is a maximum allowed time consumption, is a misjudgment penalty coefficient, when a misjudgment occurs , otherwise .

7. The artificial intelligence based flow meter fault diagnosis system of claim 1, wherein, The multi-algorithm library and intelligent decision-making module also include an LSTM algorithm based on an attention mechanism, used to process the temporal features of traffic data. Its attention weight calculation employs a traffic feature attention algorithm, with the following formula: , ,in For the first Attention weight at any moment For the output of the LSTM hidden layer, For the first Flow characteristics at any given time , This is the weight matrix. This is the bias vector.

8. The artificial intelligence based flow meter fault diagnosis system of claim 1, wherein, The multi-algorithm weighted combination strategy in the multi-algorithm library and intelligent decision module adopts a dynamic confidence fusion algorithm, the specific steps are as follows: first, the fault category prediction confidence of each participating algorithm is obtained, then the dynamic weights are calculated according to the historical accuracy and time consumption of each algorithm, and finally the confidence of each algorithm is weighted and summed according to the weights, and the category corresponding to the maximum value is selected as the final diagnosis result.

9. The artificial intelligence based flow meter fault diagnosis system of claim 1, wherein, The diagnosis execution and result output module can also output fault root cause analysis, the entity correlation strength calculation of the knowledge graph is performed by a multi-dimensional knowledge correlation reasoning algorithm, the specific steps are as follows: first, the fault entities in the diagnosis result and the candidate entities in the knowledge graph are extracted, then the correlation scores in three dimensions, i.e., semantic similarity, co-occurrence frequency, and expert annotation score, are calculated, and finally the correlation strength between entities is obtained by weighted summation, so as to infer the root cause of the fault and output maintenance suggestions.

Images