Self-adaptive threshold predictive maintenance and diagnosis system for mechanical pump

Abstract

The invention relates to the technical field of mechanical pump fault diagnosis, and discloses a mechanical pump adaptive threshold predictive maintenance diagnosis system, which comprises a data acquisition module for acquiring state data of a mechanical pump in an operation process and preprocessing the state data to form standardized input for subsequent analysis; the adversarial network module comprises a generator and a discriminator and is used for generating an expansion sample and outputting an abnormal score; the self-adaptive threshold module is used for receiving the abnormal score and the operation condition information; a trend prediction module; a causal chain diagnosis module; and a result output module. An expansion sample is constructed under the standard input and working condition parameter conditions through an adversarial network generator, data difficult to obtain in practice is supplemented, distribution differences of the expansion sample and real-time data are compared through a discriminator, an abnormal score is output, combination of data enhancement and real-time discrimination is achieved, and the real-time discrimination accuracy is improved. Model training data distribution is expanded under the conditions of limited samples and complex working conditions, and the accuracy of model diagnosis is improved.

Description

Technical Field

[0001] This invention relates to the field of mechanical pump fault diagnosis technology, specifically to an adaptive threshold predictive maintenance diagnosis system for mechanical pumps. Background Technology

[0002] Mechanical pumps are common fluid transport devices widely used in petroleum, chemical, power, and water conservancy industries. Their operating status directly affects the safety and efficiency of production systems. With the improvement of industrial automation, the operation monitoring and maintenance diagnosis of mechanical pumps have gradually become a research focus. Predictive maintenance diagnosis systems for mechanical pumps analyze and evaluate the health status of mechanical pumps based on real-time acquisition of mechanical pump operating data, combined with signal processing and intelligent algorithms, so as to take measures before failure occurs. Current mechanical pump maintenance diagnosis is mostly based on building models with limited operating data and performing feature analysis and threshold determination on real-time input data.

[0003] However, with current technology, mechanical pumps have scarce fault samples in actual operation, especially in extreme conditions and early fault data, which leads to insufficient model training and low diagnostic accuracy. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an adaptive threshold predictive maintenance diagnostic system for mechanical pumps, which solves the problem that it is difficult to obtain data on extreme operating conditions and early faults, resulting in insufficient model training and low diagnostic accuracy.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a mechanical pump adaptive threshold predictive maintenance diagnostic system, comprising: The data acquisition module acquires the status data of the mechanical pump during operation and preprocesses the status data to form standardized input for subsequent analysis. The adversarial network module, including a generator and a discriminator, is used to generate extended samples and output anomaly scores; The adaptive threshold module receives anomaly scores and operating condition information, dynamically adjusts the fault judgment threshold, and judges the operating status based on the adjusted threshold. The trend prediction module constructs a health index based on standardized input and extended samples, and performs time-series prediction on the health index to output the health status trend and remaining life estimate for a future period of time. The causal chain diagnosis module, when the operating status is determined to be abnormal, combines health status trends and multi-source feature data to generate a causal chain from abnormal signals to potential causes and possible consequences. The results output module outputs health status trends, remaining lifespan estimates, and causal chain diagnostic results, and generates maintenance prompts or risk warnings.

[0006] The above technical solution utilizes an adversarial network generator to construct extended samples under standardized input and operating parameters to supplement fault data that is difficult to obtain in actual operation. By comparing the distribution differences between the extended samples and real-time input data through a discriminator, an anomaly score that can quantify the equipment status is output, thus realizing the combination of data augmentation and real-time discrimination. Therefore, even under conditions of limited fault samples and complex operating conditions, it can still provide a reliable basis for the dynamic adjustment of adaptive thresholds.

[0007] Preferably, the data acquisition includes: The operating status of the mechanical pump is monitored, and multi-source data including vibration signals, pressure signals, flow signals, temperature signals and current signals are collected. The collected multi-source data is parsed, time-aligned, and noise-filtered to form a unified input format; The processed data is normalized and features are extracted to output a standardized dataset.

[0008] Preferably, the adversarial network includes: The generator constructs extended samples based on standardized inputs to simulate the state data of mechanical pumps under different operating conditions. The discriminator identifies the distribution differences between the real-time input data and the normal operating mode, outputs the corresponding anomaly score, and provides the anomaly score to the subsequent adaptive threshold module.

[0009] Preferably, the adversarial network implementation includes: A generative adversarial network based on operating parameters is used, in which the generator generates extended samples using the operating parameters as input, and the discriminator updates the discriminant model based on the comparison between the extended samples and real-time data.

[0010] Preferably, the adaptive threshold includes: Receive anomaly scores and operating condition parameters; Thresholds are generated based on anomaly scores and operating condition parameters; The system compares real-time input data using a threshold and outputs the result of the running status determination.

[0011] Preferably, the threshold is a dynamic threshold, and the threshold is generated in the following ways: The threshold is adjusted based on the anomaly score; The threshold is adjusted by weighting based on the operating condition parameters; After the threshold is updated, it is applied to determine the running status.

[0012] Preferably, the trend prediction includes: A health index is constructed based on standardized input and expanded samples; Using health indices as time-series inputs, a predictive model is built. The predictive model outputs the future health status trend and remaining life estimate for a certain period of time.

[0013] Preferably, the prediction model is a time series model based on a recurrent neural network, and the health index is input into the prediction model according to a preset time window, and the prediction results within the corresponding time range are output.

[0014] Preferably, the causal chain diagnosis includes: Receives operational status information provided by the adaptive threshold module and the trend prediction module; Based on a pre-defined causal relationship model, abnormal signals are associated with their corresponding potential causes; Generate a causal chain from the abnormal signal to the possible consequences in a causal order.

[0015] Preferably, the causal relationship model is constructed based on anomaly scores, health status trends, and multi-source feature data; The causal relationship model uses a causal graph structure. The nodes of the causal graph include signal feature nodes, operating condition nodes, and health status nodes. The direction of the edges is determined by the statistical dependency relationship between the nodes and the preset process logic. When the causal chain diagnosis module receives an abnormal signal, it takes the relevant signal feature nodes as observation input, determines the potential cause and generates a causal chain through causal path reasoning.

[0016] This invention provides an adaptive threshold predictive maintenance diagnostic system for mechanical pumps. It offers the following advantages: 1. This invention constructs extended samples under standardized input and operating parameters using an adversarial network generator to supplement data that is difficult to obtain in practice. It then compares the distribution differences between the extended samples and real-time data using a discriminator to output anomaly scores. This achieves a combination of data augmentation and real-time discrimination, thereby expanding the distribution of model training data under limited samples and complex operating conditions and improving the accuracy of model diagnosis.

[0017] 2. This invention combines anomaly scoring correction and weighted adjustment of operating parameters on the basis of the basic threshold, so that the threshold is updated as the operating status changes. Thus, the comparison between the operating data and the threshold can accurately reflect the equipment status under different operating conditions, avoiding the problem that fixed threshold judgment is not suitable for complex operating conditions.

[0018] 3. This invention fuses standardized inputs with extended samples to construct a health index, which is then input into a recurrent neural network for time-series prediction, outputting future health trends. When the prediction result is lower than a preset threshold, the remaining lifespan is estimated. By combining real data with virtual samples, the accuracy of health status prediction and lifespan estimation is improved.

[0019] 4. This invention establishes a causal relationship between signal characteristics, operating condition information and health status based on causal graphs. When an abnormal signal is received, conditional probability reasoning is performed to identify potential causes and generate a causal chain from the abnormal signal to the potential cause and then to the possible consequences, thereby improving the interpretability and causal tracking ability of the diagnosis. Attached Figure Description

[0020] Figure 1 This is an architecture diagram of the mechanical pump adaptive threshold predictive maintenance diagnostic system of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be clearly and completely described 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.

[0022] Please see the appendix Figure 1 This invention provides an adaptive threshold predictive maintenance diagnostic system for mechanical pumps, comprising: The data acquisition module acquires the status data of the mechanical pump during operation and preprocesses the status data to form standardized input for subsequent analysis. Further data collection includes: The operating status of the mechanical pump is monitored, and multi-source data including vibration signals, pressure signals, flow signals, temperature signals and current signals are collected. The collected multi-source data is parsed, time-aligned, and noise-filtered to form a unified input format; The processed data is normalized and features are extracted to output a standardized dataset.

[0023] Specifically, during the operation of the mechanical pump, vibration sensors, pressure transmitters, flow meters, temperature sensors, and current sensors are deployed to monitor the key parts and operating conditions of the mechanical pump in real time, collect multi-source status data, and ensure the consistency of data from each channel through a unified timestamp, covering the operating status of the mechanical pump, avoiding information loss, and providing a complete data foundation for subsequent analysis. The collected multi-source operational status data is parsed to unify the storage format and time stamp of data from different sensors. Time alignment is achieved through interpolation or resampling. At the same time, conventional signal processing methods are used to filter noise from the raw data to reduce the impact of sampling errors and environmental interference on data quality, providing a clean and comparable data foundation for subsequent processing. After format parsing and cleaning, the data is normalized to eliminate dimensional and order-of-magnitude differences between different physical quantities. Typical features reflecting the operating status of mechanical pumps, such as mean, variance, frequency band energy, or peak factor, are extracted from the data to output a standardized dataset, forming a unified input.

[0024] The adversarial network module, including a generator and a discriminator, is used to generate extended samples and output anomaly scores; Furthermore, adversarial networks include: The generator constructs extended samples based on standardized inputs to simulate the state data of mechanical pumps under different operating conditions. The discriminator identifies the distribution differences between real-time input data and normal operating mode, outputs the corresponding anomaly score, and provides the anomaly score to the subsequent adaptive threshold module. The implementation of adversarial networks includes: Generative Adversarial Networks (GANs) based on operating parameters generate extended samples using the operating parameters as input, and the discriminator updates the discrimination model based on the comparison between the extended samples and real-time data.

[0025] Specifically, in the adversarial network module, the generator receives standardized input from the data acquisition module. At the same time, combined with operating condition parameters As a conditional vector, through a nonlinear mapping function Generate extended samples The process is represented as ; in, This represents the input features after normalization and feature extraction. Indicates operating parameters (such as pressure, speed, and medium temperature). These are the generator parameters. The generator simulates state data that may occur under different working conditions but are scarce in actual samples, thereby expanding the coverage of training data, ensuring that the model can learn and adapt to diverse working conditions, and overcoming the problem of insufficient fault samples. The discriminator receives real-time input data. Expanded samples of the generator output Through mapping function Calculate the discriminant value: ; in, This indicates real-time input data. To expand the sample, For discriminator parameters, Indicates an abnormal score, if A value close to 0 indicates a significant difference in data distribution compared to normal operating conditions; if If the score is close to 1, it means that the data is close to the normal distribution. The discriminator outputs the anomaly score as a numerical signal and passes it to the adaptive threshold module for dynamic threshold adjustment. By comparing the distribution differences, the complex data stream is compressed into a single indicator, ensuring efficient use of data in the thresholding process. To ensure that the expanded samples output by the generator reflect the characteristics of actual working conditions, the adversarial network adopts a conditional generative adversarial network based on working condition parameters, and its training objective function is: ; in, This represents the actual data distribution from standardized inputs and operating parameters. This represents the input noise distribution of the generator. Indicates the generation of samples, The discriminant function is represented so that the generator gradually approximates the real data distribution. The discriminator continuously optimizes its comparison ability to ensure that the extended samples output by the generator are consistent with the real working condition data in terms of feature distribution, thereby improving the reliability of subsequent anomaly scoring.

[0026] The adaptive threshold module receives anomaly scores and operating condition information, dynamically adjusts the fault judgment threshold, and judges the operating status based on the adjusted threshold. Furthermore, the adaptive threshold includes: Receive anomaly scores and operating condition parameters; Thresholds are generated based on anomaly scores and operating condition parameters; The system compares real-time input data using thresholds and outputs the running status determination result. The threshold is a dynamic threshold, and its generation methods include: The threshold is adjusted based on the anomaly score; The threshold is adjusted by weighting based on the operating condition parameters; After the threshold is updated, it is applied to determine the running status.

[0027] Specifically, the adaptive threshold module first receives anomaly scores from the discriminator output. Simultaneously, obtain the operating parameters of the mechanical pump. (Such as rotational speed, inlet pressure, medium temperature, etc.), forming the current state input pair By simultaneously considering anomaly scores and operating condition information, false alarms and missed alarms caused by relying on a single indicator are avoided, ensuring that threshold adjustments are real-time and targeted. In obtaining Then, the adaptive threshold module generates a dynamic threshold based on the anomaly score and operating parameters. Specifically: ; in, Based on the threshold, The historical average anomaly score, This is the correction factor for abnormal scoring. A weighted vector of operating condition parameters. As the current operating condition parameter vector, the threshold will be corrected as the abnormal score fluctuates, and automatically adjusted according to the changes in operating conditions to achieve dynamic updating of the threshold. This can both track changes in equipment status and avoid the lack of adaptability caused by fixed thresholds. When the new threshold After generation, the status indicators corresponding to the input data will be updated in real time (e.g., the latest anomaly score output by the discriminator). Or extracted health index ) is compared with a threshold, if or If the current operating status is abnormal, it is determined to be abnormal; otherwise, it is determined to be normal. The determination result is output as a numerical signal and transmitted to the causal chain diagnosis module for subsequent fault interpretation and cause analysis. This ensures that the determination of the operating status no longer depends on a fixed threshold, but is dynamically adjusted according to the real-time operating status and working conditions of the equipment, thereby improving the sensitivity and accuracy of the diagnosis.

[0028] The trend prediction module constructs a health index based on standardized input and extended samples, and performs time-series prediction on the health index to output the health status trend and remaining life estimate for a future period of time. Furthermore, trend forecasting includes: A health index is constructed based on standardized input and expanded samples; Using health indices as time-series inputs, a predictive model is built. The prediction model is a time series model based on a recurrent neural network. The health index is input into the prediction model according to a preset time window, and the prediction results within the corresponding time range are output. The predictive model outputs the future health status trend and remaining life estimate for a certain period of time.

[0029] Specifically, the trend prediction module receives standardized input from the data acquisition module and extended samples generated by the adversarial network module, and performs feature fusion on the two to construct a health index, which is achieved through a weighted combination: ; in, Represents standardized input features. Indicates an expanded sample. To integrate the weight vector, its health index can comprehensively reflect the characteristics of the actual collected data and the expanded sample in a single indicator, providing input for subsequent time series prediction; The health index sequence is input into a time-series prediction model based on a recurrent neural network in chronological order. The recurrent structure stores historical information and learns the dynamic pattern of health status changes over time, thereby enabling prediction of future status. To ensure the continuity of prediction, the health index sequence is divided into fixed-length time windows as input to the prediction model. The model generates predicted health index values ​​for several future moments at the output, forming a health status trend sequence to reflect the operating trend over a period of time in the future. After obtaining the predicted sequence, based on the preset health threshold Perform a remaining lifetime estimate; if the first condition that is met appears in the prediction results... The moment The remaining lifespan is then determined as follows: ; in, It represents the time step from the current moment until the health index falls below the threshold, outputting the future health status trend and quantifying the remaining life of the mechanical pump, providing a basis for subsequent causal chain diagnosis and maintenance decisions.

[0030] The causal chain diagnosis module, when the operating status is determined to be abnormal, combines health status trends and multi-source feature data to generate a causal chain from abnormal signals to potential causes and possible consequences. Furthermore, causal chain diagnosis includes: Receives operational status information provided by the adaptive threshold module and the trend prediction module; Based on a pre-defined causal relationship model, abnormal signals are associated with their corresponding potential causes; Among them, the causal relationship model is constructed based on abnormal scores, health status trends and multi-source feature data; The causal relationship model uses a causal graph structure. The nodes of the causal graph include signal feature nodes, operating condition nodes, and health status nodes. The direction of the edges is determined by the statistical dependencies between nodes and the preset process logic. When the causal chain diagnosis module receives an abnormal signal, it takes the relevant signal feature nodes as observation input, determines the potential cause and generates a causal chain through causal path reasoning; Generate a causal chain from the abnormal signal to the possible consequences in a causal order.

[0031] Specifically, when the operating status is determined to be abnormal, the causal chain diagnosis module first receives the abnormal score and judgment result from the adaptive threshold module, as well as the health status trend data from the trend prediction module. Combined with the multi-source feature data provided in the data acquisition stage, it forms the operating status information input, which constitutes the observation basis for the subsequent causal relationship model. This makes the diagnosis not only dependent on a single threshold result, but also combined with multi-dimensional time series and operating condition data. Based on the above input, the causal chain diagnosis module calls the preset causal relationship model. This model is constructed from anomaly scores, health status trends, and multi-source feature data, and is represented by a causal graph structure. The nodes of the causal graph are divided into signal feature nodes, operating condition nodes, and health status nodes. Different types of nodes are connected by directed edges. The direction of the edges is determined by the statistical dependency between nodes and the preset process logic, thereby linking low-level signal features with high-level health status to form a hierarchical causal description. Upon receiving an abnormal signal, the causal chain diagnostic module uses the corresponding signal feature node as the observation input and performs causal path inference on the causal graph. Its calculation can be expressed as: ; in, Indicate the potential reasons for the candidate. This represents the observed anomalous signals and related characteristic data. The prior probability of the cause. The conditional probability of producing this observation given the cause; After completing the causal reasoning, the causal chain diagnostic module associates the identified potential causes with the functional events preset by the process logic, and generates a causal chain from the abnormal signal to the potential cause and then to the possible consequences in a causal order. At the same time, this causal chain not only includes the direct signal that triggers the abnormality and its cause, but also points to the potential operational consequences, such as performance degradation, component wear or failure risk. Its hierarchical causal chain provides an interpretable basis for subsequent maintenance and decision-making, so that the diagnostic results have both predictive and causal attribution capabilities.

[0032] The results output module outputs health status trends, remaining lifespan estimates, and causal chain diagnostic results, and generates maintenance prompts or risk warnings.

[0033] Specifically, after receiving the health status trend and remaining life estimate output by the trend prediction module, and the causal chain results output by the causal chain diagnosis module, the result output module compares and comprehensively analyzes these results with the preset judgment rules. When the health status trend shows a continuous decline or the remaining life is less than the preset threshold, the module generates a maintenance prompt according to the threshold rules. When the causal chain diagnosis result points to a critical component or predicts a potential cause that may lead to serious consequences, a risk warning is issued, thereby realizing preventive maintenance and risk management.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A mechanical pump adaptive threshold predictive maintenance diagnostic system, characterized in that, include: The data acquisition module acquires the status data of the mechanical pump during operation and preprocesses the status data to form standardized input for subsequent analysis. The adversarial network module, including a generator and a discriminator, is used to generate extended samples and output anomaly scores; The adaptive threshold module receives anomaly scores and operating condition information, dynamically adjusts the fault judgment threshold, and judges the operating status based on the adjusted threshold. The trend prediction module constructs a health index based on standardized input and extended samples, and performs time-series prediction on the health index to output the health status trend and remaining life estimate for a future period of time. The causal chain diagnosis module, when the operating status is determined to be abnormal, combines health status trends and multi-source feature data to generate a causal chain from abnormal signals to potential causes and possible consequences. The results output module outputs health status trends, remaining lifespan estimates, and causal chain diagnostic results, and generates maintenance prompts or risk warnings.

2. The adaptive threshold predictive maintenance diagnostic system for mechanical pumps according to claim 1, characterized in that, The data collection includes: The operating status of the mechanical pump is monitored, and multi-source data including vibration signals, pressure signals, flow signals, temperature signals and current signals are collected. The collected multi-source data is parsed, time-aligned, and noise-filtered to form a unified input format; The processed data is normalized and features are extracted to output a standardized dataset.

3. The mechanical pump adaptive threshold predictive maintenance diagnostic system according to claim 1, characterized in that, The adversarial network includes: The generator constructs extended samples based on standardized inputs to simulate the state data of mechanical pumps under different operating conditions. The discriminator identifies the distribution differences between the real-time input data and the normal operating mode, outputs the corresponding anomaly score, and provides the anomaly score to the subsequent adaptive threshold module.

4. The mechanical pump adaptive threshold predictive maintenance diagnostic system according to claim 3, characterized in that, The adversarial network implementation includes: A generative adversarial network based on operating parameters is used, in which the generator generates extended samples using the operating parameters as input, and the discriminator updates the discriminant model based on the comparison between the extended samples and real-time data.

5. The adaptive threshold predictive maintenance diagnostic system for mechanical pumps according to claim 1, characterized in that, The adaptive threshold includes: Receive anomaly scores and operating condition parameters; Thresholds are generated based on anomaly scores and operating condition parameters; The system compares real-time input data using a threshold and outputs the result of the running status determination.

6. The adaptive threshold predictive maintenance diagnostic system for mechanical pumps according to claim 5, characterized in that, The threshold is a dynamic threshold, and the threshold is generated in the following ways: Thresholds are adjusted based on anomaly scores; The threshold is adjusted by weighting based on the operating condition parameters; After the threshold is updated, it is applied to determine the running status.

7. The adaptive threshold predictive maintenance diagnostic system for mechanical pumps according to claim 1, characterized in that, The trend prediction includes: A health index is constructed based on standardized input and expanded samples; Using health indices as time-series inputs, a predictive model is built. The predictive model outputs the future health status trend and remaining life estimate for a certain period of time.

8. The adaptive threshold predictive maintenance diagnostic system for mechanical pumps according to claim 7, characterized in that, The prediction model is a time series model based on a recurrent neural network. The health index is input into the prediction model according to a preset time window, and the prediction results within the corresponding time range are output.

9. The adaptive threshold predictive maintenance diagnostic system for mechanical pumps according to claim 1, characterized in that, The causal chain diagnosis includes: Receives operational status information provided by the adaptive threshold module and the trend prediction module; Based on a pre-defined causal relationship model, abnormal signals are associated with their corresponding potential causes; Generate a causal chain from the abnormal signal to the possible consequences in a causal order.

10. The adaptive threshold predictive maintenance diagnostic system for mechanical pumps according to claim 9, characterized in that, The causal relationship model is constructed based on anomaly scores, health status trends, and multi-source feature data; The causal relationship model uses a causal graph structure. The nodes of the causal graph include signal feature nodes, operating condition nodes, and health status nodes. The direction of the edges is determined by the statistical dependency relationship between the nodes and the preset process logic. When the causal chain diagnosis module receives an abnormal signal, it takes the relevant signal feature nodes as observation input, determines the potential cause and generates a causal chain through causal path reasoning.

Images