A centrifugal pump intelligent detection method and system based on multi-source data

Abstract

The application relates to the technical field of industrial equipment health monitoring and fault diagnosis, and discloses a centrifugal pump intelligent detection method and system based on multi-source data, wherein the centrifugal pump intelligent detection method based on multi-source data comprises the following steps: collecting multi-source heterogeneous data of a centrifugal pump and preprocessing; extracting multi-scale features and constructing an optimal feature subset through hierarchical feature selection; constructing a multi-level dynamic Bayesian network to establish a probability inference framework; calculating feature credibility weights based on signal quality, feature stability and diagnostic correlation; performing multi-source evidence fusion and Bayesian state inference; and realizing confidence self-calibration and parameter self-adaptive adjustment through historical diagnosis result feedback; the application can effectively cope with complex and changeable environments in industrial fields, still maintains stable diagnosis performance under the condition that a part of sensors is invalid or data quality is uneven, improves the accuracy and reliability of fault diagnosis, and provides strong technical support for predictive maintenance of industrial equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment health monitoring and fault diagnosis, and more specifically, to a centrifugal pump intelligent detection method and system based on multi-source data. Background Art

[0002] With the rapid development of industrial automation and intelligent manufacturing, centrifugal pumps, as key equipment in industrial production, have attracted widespread attention for their health status monitoring and fault diagnosis technologies. Centrifugal pumps face complex and changing operating conditions during long-term operation, and various faults such as impeller damage, bearing wear, and seal leakage frequently occur. If these faults are not discovered and addressed promptly, they can lead to decreased equipment performance, energy waste, and even safety accidents and production line downtime, resulting in significant economic losses.

[0003] Traditional centrifugal pump health monitoring methods primarily rely on a single type of sensor data, such as analyzing only vibration or current signals. This approach has limitations in information acquisition and struggles to fully reflect the complex health status of the equipment. With advances in sensing technology, the collection of multi-source, heterogeneous data has become more convenient. The key challenge and challenge of current research is how to effectively integrate multi-source data such as vibration, acoustics, current, temperature, pressure, and flow to build a more comprehensive and reliable health assessment model.

[0004] In terms of data analysis, traditional feature engineering methods often use fixed feature extraction and fusion strategies, lacking the ability to adapt to different operating conditions and fault types, making them difficult to cope with the complex and changing environmental conditions of industrial sites. Furthermore, most existing fault diagnosis systems lack a quantitative assessment of the uncertainty of diagnostic results, failing to provide reliable confidence information for maintenance decisions. Therefore, developing a centrifugal pump health monitoring system that can fully utilize multi-source data, possess adaptive reasoning capabilities, and quantify diagnostic uncertainty is crucial for improving the reliability and maintenance efficiency of industrial equipment. Summary of the Invention

[0005] The present invention provides a centrifugal pump intelligent detection method and system based on multi-source data, which solves the technical problems of insufficient accuracy, poor robustness and weak adaptability in related technologies in terms of multi-source data fusion, adaptive reasoning ability, uncertainty quantification and end-to-end diagnostic chain.

[0006] The present invention provides a centrifugal pump intelligent detection method based on multi-source data, comprising the following steps:

[0007] Collect multi-source heterogeneous data of centrifugal pumps and perform preprocessing;

[0008] Extract multi-scale features from pre-processed multi-source data and construct the optimal feature subset through hierarchical feature selection;

[0009] Construct a multi-level dynamic Bayesian network based on the selected feature subset and establish a probabilistic reasoning framework from the observation layer to the state layer;

[0010] Calculate feature confidence weights based on signal quality, feature stability, and diagnostic relevance;

[0011] Multi-source evidence fusion and Bayesian state inference are performed using feature credibility weights to obtain the health status estimation of the centrifugal pump.

[0012] By comparing historical diagnostic results with actual status, confidence self-calibration and parameter adaptive adjustment are achieved to continuously optimize system performance.

[0013] In a preferred embodiment, the multi-source heterogeneous data includes at least two of vibration data, acoustic data, current data, temperature data, pressure data, and flow data.

[0014] In a preferred embodiment, multi-scale feature extraction includes:

[0015] Apply wavelet packet transform to vibration and acoustic signals to extract time-frequency features at different decomposition levels;

[0016] Calculate the power spectrum density and harmonic distortion of the current signal and extract the frequency domain features;

[0017] Extract statistical features and change rate features from temperature, pressure and flow data.

[0018] In a preferred embodiment, hierarchical feature selection includes:

[0019] Construct a feature correlation matrix based on the mutual information criterion to quantify the redundancy and complementarity between features;

[0020] Sort individual features based on information gain and select the top k high information gain features;

[0021] Calculate the mutual information between features and iteratively construct the optimal feature subset through the maximum correlation and minimum redundancy criterion.

[0022] In a preferred embodiment, constructing a multi-level dynamic Bayesian network includes:

[0023] Design a three-layer network structure, where the bottom-layer nodes represent sensor observations, the middle-layer nodes represent the health status of each component of the centrifugal pump, and the top-layer nodes represent the health index of the entire equipment;

[0024] Determine the transition probability of each state node in the time series based on the equipment degradation model and historical data;

[0025] Map the extracted features to each observation node and define the conditional probability distribution;

[0026] Connections representing time dependencies are introduced to connect adjacent time-step state nodes to form a complete dynamic Bayesian network structure.

[0027] In a preferred embodiment, the feature credibility weight calculation includes:

[0028] Evaluate the quality level of the original signal based on signal-to-noise ratio, entropy value and data integrity indicators;

[0029] The coefficient of variation of the feature is calculated by the sliding window method to quantify the stability of the feature in a short period of time;

[0030] Based on the historical diagnostic accuracy, calculate the diagnostic contribution of each feature to different fault types;

[0031] The feature credibility weight is calculated comprehensively based on the signal quality score, feature stability score, and diagnostic relevance score.

[0032] In a preferred embodiment, multi-source evidence fusion and Bayesian state reasoning include:

[0033] Convert the extracted features into observational evidence in the form of probabilities;

[0034] The evidence conflict coefficient is calculated based on DeShauver's evidence theory, and the conflict reconciliation mechanism is triggered when the conflict coefficient exceeds the preset threshold;

[0035] Perform weighted fusion of evidence based on feature credibility weights;

[0036] Posterior probability reasoning is performed based on a dynamic Bayesian network to calculate the probability distribution of the current state of the device.

[0037] In a preferred embodiment, confidence self-calibration and adaptive parameter adjustment include:

[0038] Compare the diagnostic results output by the system with the actual fault type confirmed by subsequent maintenance verification or experts;

[0039] Calculate diagnostic accuracy under different operating conditions and fault types to identify areas where system performance needs to be optimized;

[0040] Automatically adjust feature extraction parameters, balance coefficients in credibility weight calculation, and conditional probability distribution of Bayesian networks based on error analysis results;

[0041] Confidence intervals and uncertainty indicators of diagnostic results are calculated, and enhanced data collection or manual intervention is triggered when the uncertainty exceeds a threshold.

[0042] In a preferred embodiment, Bayesian state reasoning calculates the posterior probability distribution of the equipment state at the current moment by combining the state transition probability, the posterior probability of the state at the previous moment, the feature credibility weight, and the likelihood probability of the observation evidence, thereby realizing dynamic reasoning of the health state of the centrifugal pump; wherein the historical observation evidence sequence from time 1 to the current time t is taken into account, and the evidence contribution of all features is weighted and fused.

[0043] In a preferred embodiment, a centrifugal pump intelligent detection system based on multi-source data is used to perform a centrifugal pump intelligent detection method based on multi-source data, including:

[0044] Data acquisition and preprocessing module, used to collect and preprocess multi-source heterogeneous data of centrifugal pumps;

[0045] Feature extraction and selection module, which is used to extract multi-scale features from preprocessed multi-source data and construct the optimal feature subset through hierarchical feature selection;

[0046] Bayesian network building module, used to build multi-level dynamic Bayesian networks and establish a probabilistic reasoning framework from the observation layer to the state layer;

[0047] A feature credibility evaluation module for calculating feature credibility weights based on signal quality, feature stability, and diagnostic relevance;

[0048] The evidence fusion reasoning module is used to perform multi-source evidence fusion and Bayesian state reasoning to obtain the health status estimation of the centrifugal pump;

[0049] The adaptive parameter adjustment module is used to achieve confidence self-calibration and parameter adaptive adjustment through feedback from historical diagnostic results, thereby continuously optimizing system performance.

[0050] The beneficial effects of the present invention are:

[0051] By integrating multi-source heterogeneous data and applying dynamic Bayesian networks and evidence theory, a system was developed to achieve high-precision inference of the health status of centrifugal pumps. This system effectively copes with the complex and changing environmental conditions of industrial sites, significantly improving the accuracy and reliability of fault diagnosis, especially maintaining stable diagnostic performance in the event of partial sensor failure or uneven data quality.

[0052] The adaptive weight calculation mechanism enables the system to dynamically adjust the weights of different features based on signal quality, feature stability, and diagnostic relevance, effectively addressing the difficulty of traditional fixed-weight methods in adapting to changing operating conditions. Through hierarchical feature selection and multi-scale feature extraction, the system extracts the most diagnostically valuable feature subsets from a high-dimensional feature space, significantly reducing computational complexity while maintaining high diagnostic sensitivity.

[0053] A multi-level dynamic Bayesian network establishes an end-to-end reasoning framework from the observation layer to the state layer. This framework not only outputs health state estimates for centrifugal pumps but also provides confidence assessments of diagnostic results, providing more comprehensive information support for maintenance decisions. The system's evidence fusion mechanism effectively handles conflicting information in multi-source data, improving the system's robustness in the face of unknown failure modes.

[0054] The adaptive parameter adjustment module automatically optimizes system parameters by comparing historical diagnostic results with actual status, enabling the system to continuously improve diagnostic performance over time. This self-learning capability makes the system particularly suitable for new or upgraded equipment that lacks a complete fault sample set. It can gradually build more accurate fault models from limited historical data, providing strong technical support for predictive maintenance of industrial equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of a centrifugal pump intelligent detection method based on multi-source data of the present invention;

[0056] Figure 2 is a bar chart comparing the quality scores of different sensor data of the present invention;

[0057] Figure 3 is the health index degradation curve of different fault types of the centrifugal pump of the present invention;

[0058] Figure 4 It is a radar chart comparing the performance of the present invention's method with the traditional single-source method;

[0059] Figure 5 is a bar chart comparing the accuracy of different fault diagnosis methods of the present invention in complex scenarios;

[0060] Figure 6 is a pie chart analyzing the sources of computational complexity reduction of the present invention;

[0061] Figure 7 It is a box plot of the diagnosis accuracy distribution of different fault types of the present invention. DETAILED DESCRIPTION

[0062] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0063] At least one embodiment of the present invention discloses a centrifugal pump intelligent detection method based on multi-source data, such as Figure 1 As shown, the following steps are included:

[0064] Step 1: Collect multi-source heterogeneous data of centrifugal pumps and perform preprocessing;

[0065] By placing various types of sensors at key locations on the centrifugal pump, we collect heterogeneous data from multiple sources, including vibration, acoustics, current, temperature, pressure, and flow. The collected raw data requires systematic preprocessing to improve the accuracy and reliability of subsequent analysis. This process involves the following four sub-steps:

[0066] First, the vibration and acoustic signals are processed, mainly using filtering and resampling methods to effectively remove high-frequency noise and electromagnetic interference, ensure the purity and representativeness of the signal, and lay the foundation for subsequent feature extraction.

[0067] Next, the current signal undergoes amplitude normalization and baseline drift correction. Normalization makes current signals from different acquisition batches or different devices comparable. Baseline drift correction eliminates signal baseline shifts caused by changes in the device's operating environment or sensor aging, ensuring that the current signal reflects actual operating conditions.

[0068] Slow-varying signals such as temperature, pressure, and flow are smoothed, and combined with outlier detection methods to remove data points with sudden or abnormal fluctuations. This effectively suppresses the impact of occasional interference on data quality and improves the stability and representativeness of these signals.

[0069] Finally, given the varying sampling frequencies of different sensor types, all signals must be time-synchronized. By unifying the time base and aligning the various signals, we ensure temporal consistency across multiple data sources, providing a foundation for subsequent data fusion and joint analysis.

[0070] Furthermore, in some implementations, preprocessing may also include a data completion step. To address data missing within a short period of time, algorithms such as interpolation can be used to complete the data, ensuring data continuity and integrity, further improving the robustness of the system and the effectiveness of the analysis.

[0071] like Figure 2 The figure shows the data quality scores of various sensor types under different operating environments, comparing the differences in data quality under standard conditions with those under harsh environments (high temperature, high humidity, and strong electromagnetic interference). This chart validates the necessity of the feature credibility assessment algorithm proposed in this patent, demonstrating significant differences in data quality between different sensor types in different environments. In particular, vibration and acoustic sensor quality scores significantly decrease in environments with strong electromagnetic interference, while temperature sensors maintain high reliability across all environments.

[0072] Step 2: extract multi-scale features from the preprocessed multi-source data and construct the optimal feature subset through hierarchical feature selection;

[0073] A multi-scale analysis method is used to extract a rich feature set from the preprocessed multi-source data, and a hierarchical feature selection mechanism is used to construct the optimal feature subset to reduce the dimension of the feature space.

[0074] Wavelet packet transform is applied to vibration and acoustic signals to extract time-frequency features at different decomposition levels. Specifically, Daubechies wavelet basis functions are used to decompose vibration signals into 3 to 5 layers to extract time-frequency features such as energy entropy, singular value decomposition coefficients, and frequency band energy ratios.

[0075] For typical centrifugal pump faults such as impeller damage, the energy distribution changes in the frequency band from 1000Hz to 3000Hz are particularly obvious, so more detailed feature extraction is performed in this frequency band.

[0076] The power spectrum density and harmonic distortion of the current signal are calculated to extract the frequency domain features. In practical applications, the motor current signal is subjected to fast Fourier transform to calculate the power spectrum density, with special attention paid to characteristic frequency points such as the rotor frequency and its multiples, the power supply frequency and the sidebands.

[0077] For unbalance faults, it is usually manifested as a significant increase in the amplitude at the speed frequency; while for bearing faults, obvious peaks appear at the characteristic frequency and its sidebands.

[0078] Extract statistical features and rate-of-change characteristics from temperature, pressure, and flow data. For slowly varying signals like temperature, calculate statistical features such as mean, standard deviation, kurtosis, and skewness, as well as rate-of-change and trend characteristics within a sliding window. For example, seal failures are often accompanied by abnormal temperature fluctuations, while the relative rate of change of pressure and flow can effectively indicate pipeline blockages or leaks.

[0079] Construct a feature correlation matrix based on the mutual information criterion to quantify the redundancy and complementarity between features;

[0080] A hierarchical feature selection algorithm was applied to select feature subsets with high diagnostic value and low redundancy based on the information gain criterion and redundancy analysis.

[0081] The algorithm adopts a two-stage selection strategy: first, the individual features are ranked based on information gain, and the top High-information-gain features are generated. Mutual information between features is then calculated, and the optimal feature subset is iteratively constructed using the maximum relevance minimum redundancy (mRMR) criterion. In a large chemical plant application, this algorithm compressed the original 964-dimensional feature space to 48 dimensions while maintaining diagnostic accuracy.

[0082] Optionally, in some implementations, principal component analysis (PCA) or autoencoders can be applied for feature dimensionality reduction, further reducing the number of features and improving computational efficiency. Furthermore, for different fault types, dedicated feature subsets can be constructed to improve detection sensitivity for specific fault types.

[0083] like Figure 3 The figure shows the degradation trend of the health index of a centrifugal pump over time under healthy conditions and different fault types (bearing fault, impeller fault, and seal fault). This chart verifies the ability of the dynamic Bayesian network model in this patent to accurately capture the degradation patterns of different fault types. It shows that the health index decreases fastest for bearing faults, followed by impeller faults, while seal faults degrade more slowly. These degradation curves provide the basis for the prior probability model for Bayesian state inference.

[0084] Step 3: construct a multi-level dynamic Bayesian network based on the selected feature subset to establish a probabilistic reasoning framework from the observation layer to the state layer;

[0085] Based on the physical structure and failure mechanism of the centrifugal pump, a multi-level dynamic Bayesian network is constructed, and a probabilistic reasoning framework is established from the observation layer to the state layer.

[0086] Design a network structure: the bottom-level nodes represent sensor observations, the middle-level nodes represent the health status of each centrifugal pump component (such as the impeller, bearings, seals, etc.), and the top-level nodes represent the health index of the entire equipment;

[0087] The network consists of three layers: the first layer is the observation layer, which contains feature nodes from multiple sensors; the second layer is the component health status layer, which represents the health status of key components of the centrifugal pump such as impellers, bearings, seals, mechanical shafts and motors. The status of each component is divided into four discrete states: "healthy", "slightly degraded", "severely degraded" and "faulty"; the third layer is the overall health index of the equipment, which is represented by a continuous value of 0-100.

[0088] The connections between nodes are determined based on the centrifugal pump structure and fault propagation paths. For example, bearing failure usually affects the health of the mechanical shaft, which in turn affects the impeller performance.

[0089] Defining state transition probabilities: Based on the equipment degradation model and historical data, the transition probability of each state node in the time series is determined. For centrifugal pump bearings, the state transition probability matrix is ​​determined based on the bearing degradation model and equipment operating conditions. For example, under normal operating conditions, the probability of transitioning from a "healthy" state to a "slightly degraded" state is approximately 0.05 / 1000 hours, while under high-load conditions, this probability may increase to 0.15 / 1000 hours. These transition probabilities are estimated using a combination of historical fault data and degradation models.

[0090] Establish an association between observational evidence and status: Map the features extracted in step 2 to each observation node and define a conditional probability distribution. For different fault types, define conditional probability distributions between specific features and status. For example, for bearing faults, the characteristic frequency amplitude in the vibration signal is highly correlated with the bearing health status.

[0091] By statistically analyzing the characteristic distribution under different bearing conditions in historical data, a conditional probability table is constructed to represent the probability of observing a specific characteristic value range under a given bearing condition.

[0092] Introducing time dependency: Constructing dynamic links connecting adjacent time step state nodes to form a complete dynamic Bayesian network structure. In the dynamic Bayesian network, the time dimension representing the time dependency is introduced to connect and The state node at each moment captures the temporal evolution of the device state. This structure allows the algorithm to reason about the current state based on historical state information, effectively suppressing misjudgments caused by transient interference.

[0093] It should be understood that the complexity of the network structure can be adjusted according to actual application requirements. In scenarios with limited computing resources, the network structure can be simplified, for example, by reducing the number of nodes in the state layer or lowering the granularity of state discretization. In scenarios requiring high-precision diagnosis, a more detailed network structure can be constructed, with additional state nodes or more complex conditional dependencies.

[0094] like Figure 4 The chart below compares the performance of the proposed method with traditional single-source methods across six key dimensions: diagnostic accuracy, interference immunity, false alarm elimination, adaptability to sensor faults, adaptability to unknown fault types, and real-time performance. This chart comprehensively demonstrates the comprehensive performance advantages of the proposed method, particularly in terms of adaptability to sensor faults and unknown fault types, which show significant improvements over traditional methods.

[0095] Step 4: Calculate feature confidence weights based on signal quality, feature stability, and diagnostic relevance;

[0096] According to one embodiment of the present application, by constructing a feature credibility evaluation algorithm, the credibility weight of each feature is dynamically calculated according to signal quality, feature stability and diagnostic relevance.

[0097] Signal quality assessment: Evaluate the quality level of the original signal based on signal-to-noise ratio, entropy value and data integrity indicators;

[0098] Signal quality assessment uses a multi-index comprehensive scoring mechanism, including:

[0099] Signal-to-noise ratio calculation: The vibration signal is denoised using wavelet de-noising and the signal-to-noise ratio is calculated. Signals above 20dB are rated as high quality.

[0100] Information entropy evaluation: Calculate the Shannon entropy value of the signal to quantify the information richness;

[0101] Data integrity check: Detects data missing rate and outlier ratio. Data sources with a completeness rate below 95% receive a lower quality score.

[0102] In practice, when the ambient temperature is higher than 60°C, the quality of vibration sensor data decreases significantly, and the system automatically reduces its weight.

[0103] Feature stability analysis: The feature volatility is calculated by the sliding window method to quantify the stability of the feature in a short period of time. This module uses a sliding window of fixed length (such as 10 minutes) to calculate the coefficient of variation of the feature within the window ( ), which is the ratio of the standard deviation to the mean.

[0104] For health devices, key features Usually remain within a stable range; while a sudden increase This may indicate a sensor failure or data quality issue. The system establishes a stability baseline for each feature based on historical data and dynamically evaluates the stability level of the current feature.

[0105] Diagnostic relevance assessment: Based on historical diagnostic accuracy, calculate the diagnostic contribution of each feature to different fault types;

[0106] Adaptive Weight Calculation: The feature credibility weights are calculated by integrating the three dimensions mentioned above. First, three scoring metrics are obtained: signal quality score, feature stability score, and diagnostic relevance score. Then, corresponding balance coefficients are set for different operating conditions and fault types. These balance coefficients reflect the relative importance of the three scoring metrics in the current context.

[0107] Afterwards, each scoring indicator is multiplied by its corresponding balance coefficient to obtain three weighted scores.

[0108] Finally, the three weighted scores are summed to obtain the comprehensive feature credibility weight, which represents the reliability and importance of a specific feature in the current context.

[0109] The function is obtained by normalizing the original signal quality indicators (SNR, entropy value and data integrity, etc.) and weighted summing them, and its value range is 0 to 1;

[0110] The function calculates the stability score of the feature within the time window, which is achieved through nonlinear mapping of the coefficient of variation;

[0111] The function calculates the diagnostic contribution of the features to each fault type based on the confusion matrix and weights them according to the fault type currently being focused on.

[0112] In practical applications, the adaptive weight algorithm has a balance coefficient 、 、 The initial values ​​of are set to 0.4, 0.3, and 0.3 respectively, and then dynamically adjusted through the feedback learning mechanism in step 6. For example, in a high-noise environment, the system will automatically increase the balance coefficient value of the signal quality score. To give more attention to signal quality; when facing new failure modes, the balance coefficient value of the diagnostic relevance score will be increased To enhance the impact of diagnostic relevance.

[0113] In addition, according to another embodiment of the present application, a spatial correlation score based on the physical location of the sensor can be introduced to give a higher correlation to sensor data with similar physical locations, thereby further improving the rationality of weight calculation.

[0114] like Figure 5 The figure shows a comparison of the diagnostic accuracy of the proposed method, a traditional expert system method, and a traditional single-feature method in various complex scenarios (normal operation, varying numbers of sensor failures, high-noise environments, and variable load conditions). This chart demonstrates the robustness of the proposed method under partial sensor failure and complex operating conditions. Even in the extreme case of simultaneous failure of three sensors, the proposed method maintains a diagnostic accuracy of over 85%.

[0115] Step 5: Multi-source evidence fusion and Bayesian state reasoning are performed using feature credibility weights to obtain the health state estimation of the centrifugal pump;

[0116] Based on the feature credibility weights calculated in step 4, multi-source evidence fusion and Bayesian state reasoning are performed to obtain the current health status estimate of the centrifugal pump.

[0117] Evidence generation: Convert the extracted features into observational evidence in the form of probability, indicating the degree of support for the device status;

[0118] The evidence generation module adopts two methods:

[0119] For known failure modes, the characteristic values ​​are mapped to probabilistic evidence supporting a specific failure hypothesis based on a statistical model;

[0120] For unknown fault modes, an anomaly detection method based on distance metrics is used to calculate the degree of deviation of the feature from the normal baseline and convert it into the strength of evidence for the anomaly. For example, for the frequency amplitude of the bearing fault feature, the system establishes a probability mapping function based on historical data, mapping the observed value of 0.15g to 76% support for the bearing fault hypothesis.

[0121] Evidence conflict detection: Identify contradictory evidence from different data sources and quantify the degree of evidence conflict using De-Schauffeur evidence theory; the algorithm calculates the evidence conflict coefficient based on De-Schauffeur evidence theory .

[0122] Specifically, we first determine all possible sets of evidence events, and then check the event sets supported by different evidence sources one by one. and When the intersection is empty, it means there is a conflict between the two evidence sources.

[0123] For all such conflicting situations, the values ​​of their respective basic credibility allocation functions are multiplied together, and then the sum of the products of all conflicting situations is summed to obtain the overall evidence conflict coefficient. The basic credibility allocation function represents the support for a specific hypothesis set and is a probability mass function obtained by mapping features to evidence.

[0124] When the conflict coefficient exceeds a preset threshold (typically 0.7), the system detects conflicting evidence and triggers a conflict reconciliation mechanism. In one factory application, when vibration data indicated a bearing fault while temperature data showed normal, the system used this mechanism to identify a loose vibration sensor, avoiding false alarms.

[0125] Weighted evidence fusion: Evidence is weighted and fused according to feature credibility weights to reduce the impact of low-quality evidence;

[0126] Bayesian state reasoning: Performs posterior probability reasoning based on a dynamic Bayesian network to calculate the probability distribution of the device's current state.

[0127] The Bayesian state inference process first considers the temporal evolution of device states, namely, the dependency between the current device state and the state at the previous moment. For each possible state at the previous moment, the system calculates the probability of transitioning to each possible state at the current moment.

[0128] Then, the system combines the posterior probability distribution of the state at the previous moment to obtain the prior distribution of the state at the current moment.

[0129] Next, the system processes the observational evidence at the current moment. Each observational evidence corresponds to a feature, and the system adjusts the influence of the evidence based on the feature's credibility weight.

[0130] Specifically, for each feature, the system calculates the probability of observing the current feature value assuming a specific device state. This probability is then multiplied by the feature's credibility weight to obtain a weighted conditional probability. The weighted conditional probabilities for all features are multiplied together to obtain the combined observation likelihood probability. Finally, the system combines the prior probability with the observation likelihood probability and applies the Bayesian formula to calculate the posterior probability distribution for the device state at the current moment.

[0131] This process integrates (sums in the discrete case) all possible states at the previous moment to obtain a complete posterior probability distribution as the estimated result of the health state of the device at the current moment.

[0132] Bayesian state inference uses the JunctionTree algorithm to achieve precise reasoning, handle the ring structure in the network, and ensure the mathematical rigor of the reasoning results.

[0133] For complex network structures, when the exact inference calculation complexity is too high, the system automatically switches to an approximate inference algorithm based on particle filtering, obtaining a near-optimal state estimate while ensuring real-time performance. In actual applications, this inference module can complete a complete inference calculation within 200ms, meeting real-time monitoring requirements.

[0134] Alternatively, if computing power permits, the Markov Chain Monte Carlo (MCMC) method can be used to perform more accurate posterior probability inference, which is particularly suitable for cases with complex network structures or continuous state spaces.

[0135] like Figure 6 The figure shows the main sources and contribution ratios of the 72% reduction in computational complexity achieved by our method compared to traditional methods. This chart illustrates that feature space compression is the most important factor in the reduction in computational complexity, followed by dynamic weight adjustment, evidence conflict detection, and verification feedback, which also contribute significantly.

[0136] Step 6: By comparing historical diagnostic results with actual status, confidence self-calibration and parameter adaptive adjustment are achieved to continuously optimize system performance;

[0137] Through the feedback of historical diagnostic results, adaptive adjustment of system parameters is achieved to improve the accuracy and reliability of diagnosis.

[0138] Diagnostic Result Verification: The system compares the diagnostic results output by the system with the actual fault type verified by subsequent maintenance or confirmed by experts. The system maintains a diagnostic result verification database, recording each diagnostic result and its final verification result. For example, if the system diagnoses a "bearing inner race fault (87% confidence)" and subsequent maintenance confirms a "combined bearing inner and outer race fault," the result is recorded as "partially correct" and used for subsequent model adjustments.

[0139] Error Analysis and Calibration: This module calculates diagnostic accuracy under different operating conditions and fault types, identifying areas of weak system performance. The error analysis module constructs a diagnostic performance matrix with operating conditions (such as low load, high load, and variable load) plotted on the horizontal axis and fault types (such as bearing failure, impeller failure, and seal failure) plotted on the vertical axis. The matrix elements represent the diagnostic accuracy under these conditions. This matrix allows the system to identify weak areas of performance, such as seal fault detection under variable load conditions, which has an accuracy rate of only 62%, requiring significant optimization.

[0140] Adaptive parameter adjustment: Based on error analysis results, key parameters such as feature extraction parameters, the balance coefficient in credibility weight calculations, and the conditional probability distribution of the Bayesian network are automatically adjusted. This parameter adjustment utilizes a gradient-based optimization algorithm to fine-tune relevant parameters in areas of weak performance to improve diagnostic accuracy. For example, in the case of "seal fault detection under variable load conditions," the system analyzed and found that temperature and pressure characteristics significantly contribute to the diagnosis of this fault type. Therefore, the system automatically increases the gamma coefficient of these characteristics in the weight calculation to strengthen their influence.

[0141] Uncertainty quantification: Calculate confidence intervals and uncertainty indicators for diagnostic results, and trigger enhanced data collection or manual intervention when the uncertainty exceeds a threshold.

[0142] The uncertainty quantification module uses the Monte Carlo method to randomly perturb the network parameters, generate multiple sets of diagnosis results, calculate their variance and entropy values, and quantify the uncertainty of the diagnosis.

[0143] When the uncertainty index exceeds 0.7, the system automatically triggers enhanced monitoring mode, increasing the sensor sampling frequency and optionally sending an alert to maintenance personnel, recommending manual inspection. This mechanism effectively prevents incorrect decisions under high uncertainty conditions and, in one application at a petrochemical company, reduced the false alarm rate by approximately 40%.

[0144] It should be noted that adaptive parameter adjustment can employ a variety of optimization algorithms. In addition to gradient-based methods, other methods, such as evolutionary algorithms and Bayesian optimization, can also be used, depending on the specific application scenario and computing resources. Furthermore, the frequency of parameter adjustment can be set based on actual needs, such as periodic adjustments or adjustments triggered when diagnostic performance degrades.

[0145] like Figure 7The figure shows a statistical comparison of the diagnostic accuracy of the proposed method and traditional single-source methods for three typical fault types: bearing faults, impeller faults, and seal faults. This chart visually demonstrates that the proposed method has a higher mean and more concentrated distribution of diagnostic accuracy for each fault type, demonstrating its superior accuracy and stability. The proposed method demonstrates significant advantages, especially for fault types such as seal faults, which are difficult to accurately diagnose using traditional methods.

[0146] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A centrifugal pump intelligent detection method based on multi-source data, characterized in that: The following steps are involved: Collect multi-source heterogeneous data of centrifugal pumps and perform preprocessing; Extract multi-scale features from pre-processed multi-source data and construct the optimal feature subset through hierarchical feature selection; Multi-scale feature extraction includes: Apply wavelet packet transform to vibration and acoustic signals to extract time-frequency features at different decomposition levels; Calculate the power spectrum density and harmonic distortion of the current signal and extract the frequency domain features; Extract statistical features and change rate features from temperature, pressure and flow data; Hierarchical feature selection includes: Construct a feature correlation matrix based on the mutual information criterion to quantify the redundancy and complementarity between features; Sort individual features based on information gain and select the top k high information gain features; Calculate the mutual information between features and iteratively construct the optimal feature subset through the maximum correlation and minimum redundancy criterion; Construct a multi-level dynamic Bayesian network based on the selected feature subset and establish a probabilistic reasoning framework from the observation layer to the state layer; Building a multi-level dynamic Bayesian network includes: Design a three-layer network structure, where the bottom-layer nodes represent sensor observations, the middle-layer nodes represent the health status of each component of the centrifugal pump, and the top-layer nodes represent the health index of the entire equipment; Determine the transition probability of each state node in the time series based on the equipment degradation model and historical data; Map the extracted features to each observation node and define the conditional probability distribution; Introducing connections that represent time dependencies, connecting adjacent time-step state nodes to form a complete dynamic Bayesian network structure; Calculate feature confidence weights based on signal quality, feature stability, and diagnostic relevance; The feature credibility weight calculation includes: Evaluate the quality level of the original signal based on signal-to-noise ratio, entropy value and data integrity indicators; The coefficient of variation of the feature is calculated by the sliding window method to quantify the stability of the feature in a short period of time; Based on the historical diagnostic accuracy, calculate the diagnostic contribution of each feature to different fault types; The feature credibility weight is calculated comprehensively based on the signal quality score, feature stability score, and diagnostic relevance score; Multi-source evidence fusion and Bayesian state inference are performed using feature credibility weights to obtain the health status estimation of the centrifugal pump. By comparing historical diagnostic results with actual status, confidence self-calibration and parameter adaptive adjustment are achieved to continuously optimize system performance.

2. The intelligent detection method for centrifugal pumps based on multi-source data according to claim 1, characterized in that: The multi-source heterogeneous data includes at least two of vibration data, acoustic data, current data, temperature data, pressure data, and flow data.

3. The intelligent detection method for centrifugal pumps based on multi-source data according to claim 1, characterized in that: Multi-source evidence fusion and Bayesian state inference include: Convert the extracted features into observational evidence in the form of probabilities; The evidence conflict coefficient is calculated based on DeShauver's evidence theory, and the conflict reconciliation mechanism is triggered when the conflict coefficient exceeds the preset threshold; Perform weighted fusion of evidence based on feature credibility weights; Posterior probability reasoning is performed based on a dynamic Bayesian network to calculate the probability distribution of the current state of the device.

4. The intelligent detection method for centrifugal pumps based on multi-source data according to claim 1, characterized in that: Confidence self-calibration and adaptive parameter adjustment include: Compare the diagnostic results output by the system with the actual fault type confirmed by subsequent maintenance verification or experts; Calculate diagnostic accuracy under different operating conditions and fault types to identify areas where system performance needs to be optimized; Automatically adjust feature extraction parameters, balance coefficients in credibility weight calculation, and conditional probability distribution of Bayesian networks based on error analysis results; Confidence intervals and uncertainty indicators of diagnostic results are calculated, and enhanced data collection or manual intervention is triggered when the uncertainty exceeds a threshold.

5. The intelligent detection method for centrifugal pumps based on multi-source data according to claim 1, characterized in that: Bayesian state reasoning calculates the posterior probability distribution of the equipment state at the current moment by combining the state transition probability, the posterior probability of the state at the previous moment, the feature credibility weight, and the likelihood probability of the observation evidence, thereby realizing dynamic reasoning of the health state of the centrifugal pump. This method considers the historical observation evidence sequence from time 1 to the current time t, and performs a weighted fusion of the evidence contributions of all features.

6. A centrifugal pump intelligent detection system based on multi-source data, used to execute the centrifugal pump intelligent detection method based on multi-source data according to any one of claims 1 to 5, characterized in that: include: Data acquisition and preprocessing module, used to collect and preprocess multi-source heterogeneous data of centrifugal pumps; Feature extraction and selection module, which is used to extract multi-scale features from preprocessed multi-source data and construct the optimal feature subset through hierarchical feature selection; Bayesian network building module, used to build multi-level dynamic Bayesian networks and establish a probabilistic reasoning framework from the observation layer to the state layer; A feature credibility evaluation module for calculating feature credibility weights based on signal quality, feature stability, and diagnostic relevance; The evidence fusion reasoning module is used to perform multi-source evidence fusion and Bayesian state reasoning to obtain the health status estimation of the centrifugal pump; The adaptive parameter adjustment module is used to achieve confidence self-calibration and parameter adaptive adjustment through feedback from historical diagnostic results, thereby continuously optimizing system performance.

Images