Fault diagnosis method for pump and related product

Abstract

The invention discloses a pump fault diagnosis method and related products, and the method comprises the steps: obtaining the Mahalanobis distance of a target pump based on a dynamic voiceprint fingerprint baseline library and the working condition data of the target pump; the working condition data comprises voiceprint data and physical data; under the condition that the mahalanobis distance is larger than or equal to a first preset threshold value, a fault result of the target pump is obtained based on the working condition data and a fault model; the fault result comprises a plurality of fault types and a fault probability corresponding to each fault type; and performing rule verification on the fault result based on the working condition data according to the fault rule to obtain a final fault result. According to the method, the problem that a traditional method cannot adapt to performance gradual change caused by natural aging or working condition fluctuation in the whole life cycle due to dependence on a static baseline is effectively solved, and the accuracy, robustness and engineering reliability of a fault diagnosis result are remarkably enhanced through a'model + rule 'double-layer verification mechanism.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method for diagnosing faults in pumps and related products. Background Technology

[0002] In the field of predictive maintenance for industrial equipment, especially for critical pumps and machinery in process industries such as coal-to-natural gas and petrochemicals, early and accurate fault warnings and diagnoses are crucial. Currently, various equipment condition monitoring technologies exist in the industry, but all have certain limitations and cannot fully meet the stringent requirements for equipment health management in complex industrial scenarios.

[0003] Traditional vibration analysis techniques are currently the most widely used method for monitoring rotating equipment. However, this method has significant drawbacks: diagnostic models often rely on static baselines, making them unable to adapt to gradual performance changes caused by natural aging or fluctuations in operating conditions throughout the equipment's lifespan, easily leading to false alarms or missed alarms. In short, traditional vibration analysis techniques suffer from relatively low accuracy due to the aforementioned problems. Summary of the Invention

[0004] In view of the above problems, this application provides a method for diagnosing pump faults and related products, with the aim of improving the accuracy of pump fault diagnosis.

[0005] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides a method for diagnosing faults in pumps and motors, including: Based on the dynamic acoustic fingerprint baseline library and the operating condition data of the target pump, the Mahalanobis distance of the target pump is obtained; the operating condition data includes acoustic data and physical data; the physical data includes at least one of inlet and outlet pressure data, flow data, motor current data and medium temperature data; the baseline in the dynamic acoustic fingerprint baseline library is dynamically updated based on real-time operating condition data; When the Mahalanobis distance is greater than or equal to the first preset threshold, the failure results of the target pump are obtained based on the operating condition data and the failure model; the failure results include multiple failure types and the failure probability corresponding to each failure type; The fault results are validated based on the operating condition data according to the fault rules to obtain the final fault result; the final fault result includes the final fault type and the fault cause corresponding to the final fault type; multiple fault types include the final fault type.

[0006] Optionally, based on the dynamic voiceprint fingerprint baseline database and the operating data of the target pump, the Mahalanobis distance of the target pump is obtained, specifically including: The operating condition data is processed to obtain the time-spectrum diagram corresponding to the operating condition data; Feature extraction is performed on the time-spectrum graph to obtain a feature vector; the feature vector includes at least one of Mel frequency cepstral coefficients, spectral centroid, and impulse factor. Constructing multidimensional features based on time-spectral graphs and feature vectors; The Mahalanobis distance of the target pump was calculated based on a dynamic voiceprint and fingerprint baseline database and multidimensional features.

[0007] Optionally, the operating condition data is processed to obtain the corresponding time-frequency spectrum, specifically including: The operating condition data is processed by short-time Fourier transform or maximum overlap discrete wavelet transform to obtain the time spectrum corresponding to the operating condition data.

[0008] Optionally, based on operating condition data and a fault model, the fault results of the target pump are obtained, specifically including: The operating condition data is processed to obtain the time-spectrum diagram and feature vector corresponding to the operating condition data; The time-spectral graph is input into the first feature extraction channel in the fault model to obtain the fault features output by the first feature extraction channel; the first feature extraction channel includes a global average pooling layer and multiple convolutional layers; The feature vector is input into the second feature extraction channel in the fault model to obtain the enhanced feature vector output by the second feature extraction channel; the second feature extraction channel includes a normalization processing layer and a bidirectional long short-term memory layer. The fault features and enhanced feature vectors are input into the feature fusion classification channel in the fault model for fusion classification processing to obtain the fault result of the target pump; the feature fusion classification channel includes a feature fusion layer, a fully connected layer and a classifier; The failure results of the target pump are output through the output layer in the failure model.

[0009] Optionally, the method further includes: If the Mahalanobis distance is less than the first preset threshold and the operating data meets the update conditions, then the baselines in the dynamic voiceprint fingerprint baseline library are smoothly migrated and updated.

[0010] Optionally, the method further includes: If the Mahalanobis distance is greater than or equal to the first preset threshold and less than the second preset threshold, it indicates that the target pump is in the first fault stage, and a fault warning is issued. If the Mahalanobis distance is greater than or equal to the second preset threshold, the Mahalanobis distance is less than the third preset threshold, the fault probability corresponding to any fault type in the fault results is greater than the fourth preset threshold, and the fault time of any fault type in the fault results exceeds the preset time, it indicates that the target pump is in the second fault stage, and a fault alarm is issued. If the Mahalanobis distance is greater than or equal to the third preset threshold, the fault probability corresponding to any fault type in the fault results is greater than the fifth preset threshold, and the fault time of any fault type in the fault results exceeds the preset time, it indicates that the target pump is in the third fault stage, and a fault alarm is triggered; the first preset threshold is less than the second preset threshold, which is less than the third preset threshold; the fourth preset threshold is less than the fifth preset threshold.

[0011] Optionally, the fault model is trained based on transfer learning, and the fault model is updated based on federated learning.

[0012] Optionally, the fault results are validated based on the operating condition data according to the fault rules to obtain the final fault results, which specifically include: Based on the fault results, the voiceprint data weights and physical data weights are updated and adjusted to obtain the updated voiceprint data weights and updated physical data weights. The operating condition data is updated based on the updated voiceprint data weights and the updated physical data weights to obtain the updated operating condition data; the updated operating condition data includes the updated voiceprint data and the updated physical data. The fault results are validated based on the updated operating condition data according to the fault rules to obtain the final fault result.

[0013] Optionally, the method also includes: The final fault result is matched with the historical fault case database to determine the fault report; the fault report shall include at least the final fault type and the corresponding maintenance strategy.

[0014] A second aspect of this application provides a fault diagnosis device for a pump, comprising: The calculation module is used to obtain the Mahalanobis distance of the target pump based on the dynamic acoustic fingerprint baseline library and the operating condition data of the target pump; the operating condition data includes acoustic data and physical data; the physical data includes at least one of inlet and outlet pressure data, flow data, motor current data and medium temperature data; the baseline in the dynamic acoustic fingerprint baseline library is dynamically updated based on real-time operating condition data; The diagnostic module is used to obtain the fault results of the target pump based on operating condition data and fault model when the Mahalanobis distance is greater than or equal to a first preset threshold. The fault results include multiple fault types and the fault probability corresponding to each fault type. The verification module is used to verify the fault results based on the operating condition data according to the fault rules, and obtain the final fault result. The final fault result includes the final fault type and the fault cause corresponding to the final fault type. Multiple fault types include the final fault type.

[0015] Compared with the prior art, this application has the following beneficial effects: This application includes obtaining the Mahalanobis distance of the target pump based on a dynamic acoustic fingerprint baseline library and the operating condition data of the target pump; the operating condition data includes acoustic fingerprint data and physical data; the physical data includes at least one of inlet and outlet pressure data, flow data, motor current data, and medium temperature data; the baseline in the dynamic acoustic fingerprint baseline library is dynamically updated based on real-time operating condition data; when the Mahalanobis distance is greater than or equal to a first preset threshold, the failure result of the target pump is obtained based on the operating condition data and the failure model; the failure result includes multiple failure types and the failure probability corresponding to each failure type; the failure result is validated according to the failure rules based on the operating condition data to obtain the final failure result; the final failure result includes the final failure type and the failure cause corresponding to the final failure type; multiple failure types include the final failure type.

[0016] This application effectively overcomes the problem of traditional methods, which rely on static baselines and cannot adapt to the gradual performance changes caused by natural aging or fluctuations in operating conditions throughout the entire life cycle of equipment, by introducing a dynamically updated acoustic fingerprint baseline library based on real-time operating data. This dynamic update mechanism significantly improves the accuracy of the baseline itself, thereby enhancing the reliability of the Mahalanobis distance calculated based on the baseline. This enables more accurate determination of whether a target pump has a real fault, significantly reducing misdiagnosis and invalid diagnoses of healthy equipment. Furthermore, this application not only utilizes fault models for preliminary diagnosis but also introduces fault rules to logically verify and constrain the model output results. Through a "model + rule" dual-layer verification mechanism, the accuracy, robustness, and engineering credibility of the fault diagnosis results are significantly enhanced. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for diagnosing pump malfunctions provided in this application embodiment; Figure 2 Flowcharts of application embodiments provided in this application; Figure 3 A structural diagram of the fault model provided in the embodiments of this application; Figure 4 A flowchart of a fault alarm provided in an embodiment of this application; Figure 5 A schematic diagram illustrating the fault diagnosis of a pump provided in an embodiment of this application; Figure 6 This is a structural diagram of a pump fault diagnosis device provided in an embodiment of this application. Detailed Implementation

[0019] As mentioned earlier, existing technologies still have significant limitations in acoustic fault detection. One type of solution, while acquiring sound signals, extracting spectral features, dynamically weighting them, and combining them with machine learning classifiers for state recognition, suffers from a core flaw: it fails to construct a dynamic acoustic baseline model deeply coupled with multi-dimensional operating conditions such as the equipment's lifecycle stage, real-time process load, and operating environment. This makes it difficult to effectively distinguish between performance drift caused by normal aging or operating condition adjustments and genuine abnormal degradation. Furthermore, its diagnostic model architecture is relatively traditional, failing to fully leverage the advantages of deep learning in automatically extracting high-order abstract features, and lacking domain expert knowledge (such as physical mechanisms and fault rules), thus limiting the model's generalization ability and interpretability.

[0020] Another approach integrates acoustic detection modules into inspection robots to achieve mobile automated inspection. However, this method is essentially still a periodic and discrete data acquisition method, unable to provide truly online, continuous, and real-time status monitoring of critical pumps and motors; moreover, its technical focus is mostly on the mechanical structure, positioning, navigation, or marking functions of the robot itself, without building a predictive maintenance closed-loop system covering the entire chain of "intelligent perception - feature fusion - dynamic modeling - fault diagnosis".

[0021] In summary, current mainstream acoustic detection technologies either remain at the post-event diagnostic level or only address the data acquisition stage. They generally suffer from common problems such as delayed early warning, rigid baselines, weak model adaptability, a disconnect between human and machine knowledge, a lack of continuous evolution capabilities, and a lack of closed-loop operation and maintenance guidance. These issues make it difficult to meet the predictive maintenance requirements of modern process industries for critical rotating equipment, which demand high reliability, early warning, adaptability, and intelligence. Therefore, there is an urgent need for a new generation of intelligent acoustic diagnostic framework that deeply integrates dynamic modeling, multi-source knowledge collaboration, and continuous learning capabilities.

[0022] In view of the above problems, this application provides a method for generating fault diagnosis data for pumps and related products. The method includes: obtaining the Mahalanobis distance of the target pump based on a dynamic acoustic fingerprint baseline library and the operating condition data of the target pump; the operating condition data includes acoustic fingerprint data and physical data; the physical data includes at least one of inlet and outlet pressure data, flow data, motor current data, and medium temperature data; the baseline in the dynamic acoustic fingerprint baseline library is dynamically updated based on real-time operating condition data; when the Mahalanobis distance is greater than or equal to a first preset threshold, obtaining the fault result of the target pump based on the operating condition data; the fault result includes multiple fault types and the fault probability corresponding to each fault type; performing rule verification on the fault result based on the operating condition data according to fault rules to obtain the final fault result; the final fault result includes the final fault type and the fault cause corresponding to the final fault type; multiple fault types include the final fault type.

[0023] This application significantly improves the accuracy, timeliness, and interpretability of pump and motor fault diagnosis by integrating multi-source heterogeneous data with a dynamic modeling mechanism. First, the construction of a dynamic acoustic fingerprint baseline library breaks through the limitations of traditional static baselines. This baseline is not only based on acoustic fingerprint data but also integrates physical parameters to form a multi-dimensional health characterization that is strongly correlated with the current operating conditions of the equipment (such as load, medium state, and operating stage). Furthermore, through continuous updates of real-time operating condition data, it effectively tracks the gradual performance changes of the equipment caused by natural aging or process fluctuations, thereby ensuring that the baseline always reflects the "current normal" state of the equipment and avoids misjudging normal drift as abnormality.

[0024] Secondly, Mahalanobis distance is used as an anomaly metric, fully considering the correlation and dimensional differences between multidimensional indicators. Compared with simpler metrics such as Euclidean distance, it can more accurately characterize the degree to which the equipment status deviates from the healthy baseline. The system only triggers deep diagnosis when the Mahalanobis distance is ≥ a first preset threshold, effectively suppressing false alarms caused by noise interference and occasional fluctuations, achieving high-confidence early warning. Furthermore, the fault diagnosis stage in this application adopts a dual-drive mechanism of "model + rule." The fault model outputs multiple possible fault types and their probability distributions based on full-scale operating data, possessing strong pattern recognition capabilities. The fault rules logically verify the model results. For example, if the model determines "cavitation," but the inlet and outlet pressure difference does not reach the typical cavitation threshold, the result is corrected or downweighted. This mechanism not only improves the physical rationality and engineering credibility of the diagnostic results but also clearly outputs the final fault type and its root cause (such as "bearing outer ring wear—due to insufficient lubrication"), rather than just providing a fuzzy classification, providing maintenance personnel with actionable decision-making basis.

[0025] In summary, this application achieves a leap from "passive response" to "proactive prediction" and from "black box diagnosis" to "explainable attribution" through four major innovations: dynamic baseline adaptation, multi-source data fusion, accurate early warning based on Mahalanobis distance, and collaborative verification of models and rules. It effectively solves the core pain points of traditional methods, such as delayed early warning, high false alarm rate, inability to distinguish between normal drift and real faults, and lack of closed-loop guidance. It provides highly reliable and intelligent predictive maintenance capabilities for key pumps and machinery in the process industry.

[0026] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0027] Figure 1 A flowchart illustrating a pump fault diagnosis method provided in this application embodiment is shown below. Figure 1 As shown, a method for diagnosing pump malfunctions includes: S101: Based on the dynamic voiceprint fingerprint baseline library and the operating data of the target pump, the Mahalanobis distance of the target pump is obtained.

[0028] This application does not limit the operating condition data, but to ensure the diversity of the operating condition data, the operating condition data may include acoustic data and physical data. Physical data includes at least one of the following: inlet and outlet pressure data, flow rate data, motor current data, and medium temperature data. This application does not limit the method of collecting operating condition data, and it can be flexibly adjusted according to the environment of the target pump. If the target pump is in harsh conditions such as high temperature, high humidity, or strong corrosion, a high temperature resistant (>120°C), intrinsically safe explosion-proof acoustic sensor can be used, which can be directly magnetically or threadedly fixed to the nearest point of the pump casing (such as the pump group in the gasification workshop) or the bearing seat of the target pump for operating condition data collection; if the target pump is in a low temperature state, such as a liquid ammonia pump, C The pump can be equipped with a low-temperature acoustic sensor (operating temperature below -50°C) to ensure stable operation in cryogenic environments.

[0029] This application does not limit the specific frequency band for acquiring acoustic data, and can be flexibly adjusted according to the type of target pump. For example, for target pumps such as quenching water pumps and washing pumps containing solid media, the focus can be on acquiring and analyzing the acoustic energy in the mid-to-high frequency band (sensitive to the 4kHz-12kHz frequency band). The sensor should be fixed to the surface of the pump casing or bearing seat at a distance of ≤5mm, and the sensor should be sealed with epoxy resin AR2000. This frequency band is extremely sensitive to the erosion and wear of the impeller and casing. For high-pressure boiler feed water pumps and liquid C For pumps and the like, focus on monitoring the broadband noise in the low-frequency range (0.5kHz-2kHz), because cavitation will generate a lot of bubble bursting noise in this frequency range; for the bearings of all target pumps, monitor their characteristic frequencies (such as the Ball Pass Frequency Outer race (BPFO) and the Ball Pass Frequency Inner race (BPFI)) and the specific high-frequency narrow band (>8kHz) where their harmonics are located.

[0030] To ensure the validity of the collected operating data, a front-end integrated physical noise reduction hood (Helmholtz resonant cavity structure) and electronic filter module (bandpass filter range 0.5-12kHz) can be used to pre-remove low-frequency mechanical noise and high-frequency electromagnetic interference on site before collecting the operating data.

[0031] This application does not limit the dynamic acoustic fingerprint baseline library, but in order to ensure the effectiveness of the dynamic acoustic fingerprint baseline library, the baseline in the dynamic acoustic fingerprint baseline library is dynamically updated based on real-time operating condition data; the dynamic acoustic fingerprint baseline library is constructed based on operating condition data collected at different stages of the target pump's life cycle (new commissioning, stable operation, mid-term) and at different process load points (50%, 80%, 100% load), and the sampling frequency can be selected as 40kHz, which meets the requirements of the Nyquist sampling theorem.

[0032] S102: When the Mahalanobis distance is greater than or equal to the first preset threshold, the failure result of the target pump is obtained based on the operating condition data and the fault model.

[0033] This application does not limit the specific value of the first preset threshold, which can be flexibly configured according to the actual application scenario, equipment type, operating condition complexity, and operation and maintenance strategy. For example, in the initial deployment phase, the first preset threshold can be set to 1 as a general initial value, and then dynamically adjusted through historical data backtracking, online learning, or expert optimization to balance the warning sensitivity and false alarm rate. This threshold can also be set differently for different pumps or operating modes to achieve refined threshold management.

[0034] Meanwhile, this application does not impose rigid restrictions on the form of fault results, but fully leverages the powerful multi-class recognition and uncertainty quantification capabilities of the advanced fault models employed (such as deep neural networks, ensemble learning, or graph neural networks). The fault results can naturally be output as probability distributions encompassing multiple potential fault types. For example, it can not only determine "bearing failure" or "cavitation," but also further refine it into fine-grained diagnoses such as "bearing inner ring wear (probability 68%)," "impeller cavitation (probability 25%)," and "shaft misalignment (probability 7%)," along with confidence assessments. This probabilistic output retains the model's discriminative richness and provides a structured basis for subsequent rule verification, manual review, and risk ranking, significantly improving the robustness and decision support value of the diagnostic system.

[0035] S103: Verify the fault results based on the operating condition data according to the fault rules to obtain the final fault result.

[0036] This application does not limit the specific form of the fault rules, which can be flexibly constructed into a logical judgment system based on domain expert experience, physical mechanisms or historical operation and maintenance knowledge. The core purpose is to perform reasonableness verification, fine-grained correction and root cause inference on the preliminary results output by the fault model, thereby improving the engineering credibility and operability of the diagnostic conclusions.

[0037] For example, when the fault model is identified as "cavitation" and simultaneously meets the acoustic characteristic changes such as a significant shift of the spectral centroid to higher frequencies and a sharp increase in the time-domain impulse factor, the fault rules can be used to confirm the "cavitation" diagnosis and automatically increase its confidence level.

[0038] If the fault model is determined to be "bearing fault", but the overall sound pressure level energy does not show a significant increase, then the fault rules can be used to refine it into "early bearing fault". Conversely, if the energy is significantly enhanced, the fault rules can be used to classify it as "mid-term or late-term fault", thus achieving precise division of the fault stage.

[0039] For example, when the fault model diagnoses "cavitation" and the real-time value of the pump inlet pressure is lower than the normal process range, the fault rules can be used to further add root cause analysis and output clear root cause prompts such as "suspected inlet filter blockage or tank liquid level too low".

[0040] Through this collaborative mechanism of "model initial judgment + rule refinement + root cause correlation," the final fault result generated by this application not only includes the definite final fault type (such as "cavitation"), but also simultaneously provides the corresponding specific fault cause or trigger (such as "insufficient inlet pressure"). The various candidate fault types output by the original fault model (such as "cavitation," "imbalance," and "cavitation-like noise") serve as intermediate results, naturally encompassing the final fault type (i.e., the final fault result includes the final fault type and its corresponding fault cause; multiple fault types include the final fault type), forming a complete reasoning chain from multiple hypotheses to a high-confidence conclusion. This design significantly enhances the system's robustness in complex operating conditions, diagnostic interpretability, and operational guidance value.

[0041] The above describes the main technical solution of this application. Further implementations of the main technical solution are now introduced. Details are as follows: Regarding the calculation of the Mahalanobis distance of the target pump based on the dynamic acoustic fingerprint baseline library and the operating data of the target pump in S101, this application provides an optional embodiment: The operating condition data is processed to obtain the corresponding time-frequency spectrum.

[0042] This application does not limit the method for determining the time spectrum. For example, the time spectrum corresponding to the working condition data can be obtained by performing short-time Fourier transform or maximum overlap discrete wavelet transform (db4 wavelet basis, 5-level decomposition) on the working condition data.

[0043] Feature extraction is performed on the time-spectrum graph to obtain the feature vector.

[0044] This application does not limit the feature vector, but the feature vector may include at least one of the following: Mel frequency cepstral coefficients (mainly extracting 13-20 dimension features, mimicking human hearing, and sensitive to periodic fault features of bearings, gears, etc.), spectral centroid (effectively characterizing the feature of high-frequency energy increase when cavitation occurs), and pulse factor (sensitive to impact faults such as loosening and rubbing).

[0045] The formula for calculating the spectral centroid (spectral roll-off point) is: ; Where P(k) is the energy value of the k-th frequency component.

[0046] The formula for calculating the impulse factor (peak-to-peak value of a time-domain signal) is: Pulse factor = Root mean square value / Average absolute value; Multidimensional features are constructed based on time-spectrum graphs and feature vectors.

[0047] The Mahalanobis distance of the target pump was calculated based on a dynamic voiceprint and fingerprint baseline database and multidimensional features.

[0048] This application does not limit the method for calculating Mahalanobis distance. For example, the Mahalanobis distance of the target pump can be calculated using the Mahalanobis distance calculation formula based on the dynamic voiceprint fingerprint baseline library and multi-dimensional features.

[0049] Formula for calculating Mahalanobis distance: ; Where x is a multidimensional feature, μ is the baseline mean in the dynamic voiceprint fingerprint baseline database, and Σ is the baseline covariance matrix in the dynamic voiceprint fingerprint baseline database.

[0050] This application also provides a specific application embodiment. Figure 2 Flowcharts of application embodiments provided in this application are shown below. Figure 2 As shown: Start: Acquisition of multi-condition acoustic data (condition data), specifically: sensor placement, housing 1, bearing housing 2 (i.e., sensors are placed on the housing and bearing housing of the target pump respectively); synchronous acquisition of condition data (such as pressure, flow rate, temperature, current, etc.); noise reduction processing of the acquired condition data (e.g., bandpass filtering 0.5-12kHz); signal processing of the noise-reduced condition data (e.g., time-frequency transformation (such as Short-Time Fourier Transform (STFT) and Maximum Overlap Discrete Wavelet Transform (MODWT))) and calculation of regular eigenvectors (e.g., Mel-Frequency coefficients). Cepstral Coefficients (MFCC), spectral centroid, impulse factor, etc. are used to generate a time-frequency spectrogram and feature vectors. Based on the time-frequency spectrogram and feature vectors, a multidimensional feature vector is constructed / compared to generate a feature vector (i.e., multidimensional features are generated based on the time-frequency spectrogram and feature vectors). The multidimensional features are then compared with historical baselines (located in a multidimensional labeled voiceprint and fingerprint baseline database) to calculate the Mahalanobis distance D. 2 If D 2 If D < threshold 1 (i.e., the first preset threshold) and the trend is stable, then the dynamic baseline is smoothly migrated to obtain an updated baseline library; if D 2 If the value is greater than or equal to the threshold 1 (i.e., the first preset threshold), then an alarm is triggered to output the feature deviation, and the AI ​​diagnosis is initiated (i.e., the fault result is determined using the fault model).

[0051] Regarding the failure results of the target pump obtained based on operating condition data and a failure model in S102, this application provides an optional embodiment: The operating condition data is processed to obtain the time-spectrum diagram and feature vector corresponding to the operating condition data.

[0052] The method for determining the time spectrum and eigenvector in this embodiment is similar to the method for determining the time spectrum and eigenvector in the above embodiments, and therefore will not be described in detail here.

[0053] Figure 3 A structural diagram of the fault model provided in the embodiments of this application, such as Figure 3 As shown, the time-spectrum graph (input dimension H×W×C) and the regular feature vector (dimensional N) are input together into the input layer of the fault model. The input layer processes them separately: the time-spectrum graph is sent to the first feature extraction channel (usually a convolutional neural network structure, used to capture local time-frequency patterns and spatial correlations in the acoustic signal), while the regular feature vector is sent to the second feature extraction channel (usually a fully connected or embedded layer, used to encode structured physical features such as process parameters and operating status). After extracting high-dimensional semantic features in parallel through the two channels, they can be fused in subsequent network layers (such as through splicing, weighting, or attention mechanisms) to achieve joint modeling driven by acoustic perception and domain knowledge, thereby improving the accuracy and interpretability of fault diagnosis.

[0054] The time-spectral graph is input into the first feature extraction channel in the fault model to obtain the fault features output by the first feature extraction channel. For example, the first feature extraction channel includes a global average pooling layer and multiple convolutional layers.

[0055] The time-spectral image is input into the first feature extraction channel of the fault model to extract its global, image-based fault semantic features (such as the "burr cloud" texture of cavitation in the time-spectral image, the periodic impact pattern corresponding to bearing failure, etc.). This channel adopts a typical deep convolutional structure, specifically including five sequentially stacked convolutional layers, each using a 3×3 convolutional kernel and ReLU activation function, with the number of channels increasing layer by layer: Layer 1: 32 feature maps, Layer 2: 64 feature maps, Layer 3: 128 feature maps, Layer 4: 256 feature maps, Layer 5: 512 feature maps.

[0056] Subsequently, the output feature map is spatially compressed by a Global Average Pooling (GAP) layer, which aggregates the spatial response of each channel into a single scalar, thereby generating a compact and highly discriminative one-dimensional fault feature vector.

[0057] like Figure 3As shown, the input time-spectrum image flows sequentially through convolutional layer 1 → convolutional layer 2 → convolutional layer 3 → convolutional layer 4 → convolutional layer 5 → global average pooling (GAP) layer. The final output is the high-level fault features extracted by the first feature extraction channel. This design not only effectively captures multi-scale local patterns in acoustic signals but also preserves global semantic information through GAP, while significantly reducing the number of parameters and the risk of overfitting. This provides a high-quality, highly abstract acoustic representation for subsequent fusion with regular features and fault classification.

[0058] The feature vector is input into the second feature extraction channel in the fault model to obtain the enhanced feature vector output by the second feature extraction channel. For example, the second feature extraction channel includes a normalization layer and a bidirectional long short-term memory layer (e.g., a hidden layer with a dimension of 128, and the bidirectional outputs are concatenated and then subjected to Dropout (rate 0.5) to prevent overfitting).

[0059] like Figure 3 As shown, the feature vector passes through the normalization processing layer and the bidirectional long short-term memory layer in the second feature extraction channel in sequence to obtain the enhanced feature vector output by the second feature extraction channel.

[0060] The fault features and enhanced feature vectors are input into the feature fusion classification channel in the fault model for fusion classification processing to obtain the fault result of the target pump. The feature fusion classification channel may include a feature fusion layer (such as a cascade / attention mechanism), a fully connected layer (FC), and a classifier (such as a Softmax classifier).

[0061] like Figure 3 As shown, a feature fusion classification channel was designed to effectively integrate and classify fault features and enhanced feature vectors. This channel first combines fault features and enhanced feature vectors from the first feature extraction channel through a feature fusion layer. The feature fusion method can be a simple concatenation or a more complex attention mechanism to highlight more important feature dimensions during the fusion process.

[0062] The fused features are then fed into one or more fully connected layers (FC) to further transform and abstract the feature space, thereby capturing deeper fault mode information. Finally, the features processed by the fully connected layers are fed into a classifier, such as the commonly used Softmax classifier, to accurately classify the fused features and output the probability distribution corresponding to different fault types of the target pump.

[0063] The failure results of the target pump are output through the output layer in the failure model.

[0064] This embodiment can be summarized as follows: fault features + enhanced feature vectors → feature fusion layer (cascaded / attention mechanism) → fully connected layer (FC) → Softmax classifier → outputting the fault probability distribution of the target pump (i.e., the fault result). This method not only fully utilizes the global fault features in the original time-spectrum graph, but also combines additional enhanced feature vectors to improve the accuracy and reliability of the final fault diagnosis. Furthermore, using Softmax as the classifier helps to transform the model output into an easily understandable probabilistic form, directly reflecting the confidence level of each fault type.

[0065] For cases where the Mahalanobis distance is less than a first preset threshold, this application provides an optional embodiment: If the Mahalanobis distance is less than the first preset threshold and the operating data meets the update conditions, then the baselines in the dynamic voiceprint fingerprint baseline library are smoothly migrated and updated.

[0066] If the Mahalanobis distance is less than the first preset threshold and a slow, consistent drift (rather than a sudden jump) is detected in the set of discrete points of the collected operating data, and the equipment operating parameters have not deteriorated, a message will be displayed that "the baseline may need to be updated". After confirmation by the engineer, the baseline will be smoothly migrated to the new position to avoid false alarms about the normal aging of the equipment performance.

[0067] In response to fault alarm situations, this application provides an optional embodiment: If the Mahalanobis distance is greater than or equal to the first preset threshold and less than the second preset threshold, it indicates that the target pump is in the first fault stage, and a fault warning is issued.

[0068] If the feature deviation (Mahanobis distance) is greater than the threshold of 1 and the model anomaly probability is between 0.7 and 0.85, the system will issue an "early warning" to alert the user.

[0069] If the Mahalanobis distance is greater than or equal to the second preset threshold, the Mahalanobis distance is less than the third preset threshold, the fault probability corresponding to any fault type in the fault results is greater than the fourth preset threshold, and the fault time of any fault type in the fault results exceeds the preset time, it indicates that the target pump is in the second fault stage, and a fault alarm is issued.

[0070] If the feature deviation is greater than the threshold of 2, the model anomaly probability is greater than 0.85, and the abnormal signal persists for more than 3 consecutive acquisition cycles (30 seconds), the system will issue an "alarm".

[0071] If the Mahalanobis distance is greater than or equal to the third preset threshold, the fault probability corresponding to any fault type in the fault results is greater than the fifth preset threshold, and the fault time of any fault type in the fault results exceeds the preset time, indicating that the target pump is in the third fault stage, a fault alarm will be triggered. This application does not limit the values ​​of the first, second, third, fourth, and fifth preset thresholds, and can set them according to the actual situation, but it must be ensured that the first preset threshold is less than the second preset threshold, which is less than the third preset threshold, and the fourth preset threshold is less than the fifth preset threshold.

[0072] An "emergency alarm" is issued when the feature deviation is greater than the threshold of 3, the model anomaly probability is greater than 0.95, the abnormal signal lasts for more than 3 cycles, and the model's classification confidence for a specific fault is greater than 90%.

[0073] This application also provides a specific application embodiment. Figure 4 The flowchart for the fault alarm provided in the embodiments of this application is as follows: Figure 4 As shown: Real-time acoustic feature extraction and AI model diagnosis (i.e., real-time acquisition of operating condition data and real-time diagnosis of operating condition data using a fault model); Condition 1: Does the warning condition meet (i.e., feature deviation (Mahathano distance) > threshold 1 and model anomaly probability between 0.7 and 0.85)? If the warning condition is not met, continue monitoring; if the warning condition is met, notify the engineer to pay attention; Condition 2: Does the condition continue to deteriorate and meet the alarm condition (i.e., feature deviation > threshold 2, model anomaly probability > 0.85, and abnormal signal lasting for more than 3 consecutive acquisition cycles (30 seconds))? If the alarm condition is not met, maintain the status (e.g., maintain and notify the engineer to pay attention); if the alarm condition is met, notify the maintenance team; Condition 3: Does the emergency alarm condition reach (i.e., feature deviation > threshold 3, model anomaly probability > 0.95, abnormal signal lasting for more than 3 cycles, and model classification confidence for a specific fault > 90%)? If the emergency alarm condition is reached, it is recommended to stop the machine for inspection and trigger intensive monitoring; if the emergency alarm condition is not reached, maintain the status (e.g., maintain and notify the maintenance team).

[0074] This embodiment ensures the rationality of fault stage division and the progressiveness of response strategies. This multi-dimensional, multi-threshold collaborative judgment mechanism effectively avoids misjudgment based on a single indicator, significantly improving the timeliness of early warnings, the accuracy of alarms, and the reliability of emergency responses.

[0075] In the early stages of the project, facing the challenge of scarce field failure samples, this application proposes a model construction and evolution strategy that integrates multi-level transfer learning and federated learning, effectively solving the "cold start" problem while taking into account both data privacy and the need for continuous model optimization.

[0076] A two-stage transfer learning mechanism is adopted: Cross-domain pre-training: When local fault data is lacking, deep models (e.g., ResNet50) pre-trained on publicly available rotating machinery acoustic datasets (e.g., PU, ​​CWRU, etc.) or general image datasets (e.g., ImageNet) are used as the initialization basis, leveraging their learned general acoustic or visual feature representation capabilities.

[0077] Cross-workshop migration and fine-tuning: Furthermore, in workshops with relatively sufficient data accumulation (such as gasification workshops), a domain-adaptive basic model is trained using a large number of normal and typical fault samples. Then, the model is migrated to new workshops with sparse data (such as purification and methanation workshops). Only a small amount of local labeled data is needed for targeted fine-tuning, which can quickly adapt to new pump types and new operating conditions, significantly shorten the model deployment cycle, and solve the modeling dilemma of "starting from scratch" in new scenarios.

[0078] On the other hand, to achieve multi-workshop collaborative evolution and safeguard data sovereignty, this application introduces a privacy-preserving federated learning framework: Edge-cloud collaborative architecture: The initial model is deployed independently on the edge server in each workshop. All raw voiceprint data and physical data do not leave the local area. Model training is completed only on the edge side.

[0079] Security parameter aggregation: Each edge node uploads its local fault model updates (such as gradients or weight differences) to the central server after perturbation and encryption using a differential privacy mechanism (e.g., adding Gaussian noise with a standard deviation of σ=0.1). The cloud uses federated averaging (FedAvg) and dynamically allocates aggregation weights based on the amount of effective data in each workshop to ensure that workshops with larger data contributions have a greater impact on the global model.

[0080] Efficient Iteration Mechanism: Set dynamic aggregation trigger conditions, and upload and update only when the accuracy of the local fault model on the validation set improves by ≥5%, avoiding invalid communication, significantly reducing bandwidth consumption and computational overhead, and improving federated training efficiency.

[0081] Through the aforementioned technical path of "pre-training → cross-domain migration → federated co-evolution", this application not only solves the problem of early data scarcity in industrial sites, but also builds a scalable, evolvable, and privacy-protected intelligent diagnostic ecosystem: the model is automatically optimized as data from each workshop continues to accumulate, and there is no need to centralize original sensitive data, truly realizing "the model moves while the data remains still, and knowledge is shared while privacy is protected", providing sustainable technical support for the large-scale implementation of predictive maintenance systems in process industries.

[0082] Regarding S103, which verifies the fault results based on the operating condition data according to the fault rules to obtain the final fault result, this application provides an optional embodiment: Based on the fault results, the voiceprint data weights and physical data weights are updated and adjusted to obtain the updated voiceprint data weights and updated physical data weights.

[0083] For example, in cavitation diagnosis, acoustic fingerprint data has a weight of 0.6, while physical data (inlet pressure, flow rate) has a weight of 0.4. In bearing fault diagnosis, acoustic fingerprint data has a weight of 0.7, while physical data (vibration, temperature) has a weight of 0.3.

[0084] Weights can be calculated using a dynamic exponential function: ; The greater the deviation of the parameter value from the optimal range, the greater its weight index, reflecting its contribution to fault diagnosis.

[0085] The operating condition data is updated based on the updated voiceprint data weights and the updated physical data weights, resulting in updated operating condition data. The updated operating condition data includes both updated voiceprint data and updated physical data.

[0086] For example, when a cavitation warning is triggered (model probability 0.7-0.85), the sound wave frequency gradually increases from 20kHz to 28kHz. When a bearing fault alarm is triggered (model probability > 0.85), the power increases from 200W to 400W. PID parameter settings: proportional coefficient Kp = 0.3, integral coefficient Ki = 0.05, derivative coefficient Kd = 0.1, to ensure smooth frequency adjustment.

[0087] The fault results are validated based on the updated operating condition data according to the fault rules to obtain the final fault result.

[0088] This application also provides an optional embodiment for fault repair strategies: The final fault result is matched with the historical fault case database to determine the fault report; the fault report shall include at least the final fault type and the corresponding maintenance strategy.

[0089] This application does not limit the specific composition of the historical failure case library, which can be flexibly constructed into a structured knowledge base, such as containing multi-dimensional information such as maintenance strategies, handling procedures, spare parts lists, working hours records and root cause analysis corresponding to various known failure modes of the target pump.

[0090] After diagnosis, the current final fault result (including fault type, feature vector, operating condition, etc.) is intelligently matched with entries in the historical fault case database. Specifically, a weighted cosine similarity algorithm is used to calculate the semantic similarity between the current fault representation and each historical case, where the weights can be dynamically adjusted according to the importance of features (such as acoustic anomaly intensity, deviation of key physical parameters, etc.) to improve matching accuracy.

[0091] The matching results are further combined with expert scores of historical repair effectiveness (using a 0-10 scale to quantify the effectiveness of the treatment) to form a comprehensive decision-making basis: If the highest similarity is ≥0.8 and the corresponding case has an expert score of ≥7, it is considered a high-confidence reuse scenario. The historical maintenance strategy will be automatically recommended, and complete maintenance records, operation steps and root cause analysis will be output.

[0092] If the similarity is less than 0.8 or the score is low, it will be marked as "requires manual review". After the operation and maintenance experts intervene and make corrections, the new handling plan and verification results will be fed back into the case library to achieve continuous accumulation and optimization of knowledge.

[0093] Through this mechanism, this application not only achieves a closed-loop connection from "diagnosis" to "disposal", but also constructs a self-learning, evolvable, human-machine collaborative intelligent operation and maintenance knowledge system, which significantly improves fault response efficiency and maintenance decision-making quality, while reducing reliance on individual experience.

[0094] Figure 5 This is a schematic diagram of the pump fault diagnosis provided in the embodiments of this application, as shown below. Figure 5 As shown, the input includes real-time voiceprint features, operating parameters, and preliminary AI diagnosis (i.e., acquiring operating data). This operating data is then input into the fault model, which analyzes the fault probability distribution and preliminarily determines the fault type and probability (e.g., cavitation 0.82, bearing 0.76). The fault results are validated using the expert system's embedded rule base (e.g., fault rules) (e.g., Rule 1: IF cavitation AND spectral centroid shifts upward THEN, confirming cavitation diagnosis; Rule 2: IF bearing fault AND energy not increased THEN, determining early fault), determining the diagnosis / severity / root cause (i.e., the final fault result). The final fault result is then matched with a historical fault case library (e.g., organized hierarchically by equipment type (e.g., quenching water pump) and fault category (cavitation, bearing wear), with each layer containing fault feature parameters, handling measures, effect evaluation, etc.) using similarity (e.g., weighted cosine similarity) to obtain historical cases / maintenance record backup information / remaining lifespan (i.e., maintenance strategy). Based on the maintenance strategy, an intelligent maintenance work order is generated and transmitted to the maintenance team.

[0095] For example: early fatigue damage of the bearing inner ring (confidence level 92%), characterized by: significantly increased energy of 162 Hz and its 2-4 harmonic components, and a 35% increase in pulse factor; historical matching: 3 similar cases, with an average remaining life of about 21 days; maintenance recommendation: include it in the planned maintenance list for next week and prepare spare parts SKF-6312 in advance.

[0096] This application utilizes high-sensitivity acoustic signature analysis to detect subtle acoustic anomalies in the early stages of cavitation, such as the attachment of micron-sized particles and the collapse of microbubbles. These anomalies include specific resonant peak frequency shifts, an increase in the proportion of high-frequency energy, and the initial appearance of a time-spectrum "burr cloud" structure. This enables early warning of degradation processes such as scaling and wear, allowing for the detection of potential problems several hours to days earlier than traditional vibration analysis methods.

[0097] It supports 24 / 7 online continuous monitoring without downtime or manual intervention, significantly reducing unplanned downtime and maintenance costs. Furthermore, based on the equipment's current operating status and historical trends, it can dynamically adjust the sampling frequency, feature extraction granularity, and diagnostic thresholds to achieve an adaptive intelligent monitoring strategy.

[0098] Crucially, this application constructs a dynamic voiceprint fingerprint baseline library deeply integrated with the entire lifecycle of the equipment and its real-time operating conditions. This effectively distinguishes between performance drift caused by natural aging or process fluctuations and actual fault degradation, fundamentally preventing the misjudgment of normal conditions as abnormal ones. Experimental verification shows that this dynamic baseline mechanism can reduce the false alarm rate by more than 40%.

[0099] In terms of data efficiency, by using transfer learning (such as pre-training on ImageNet or public mechanoacoustic datasets) and cross-workshop knowledge transfer (such as transferring from a data-rich gasification workshop to a purification or methanation workshop), high-precision fault identification can be achieved with only 10% of local labeled data, significantly reducing the costs of data collection, labeling, and model cold start.

[0100] Furthermore, relying on the federated learning update mechanism, each workshop, without leaving the domain with the original operational data, uploads model parameter updates (such as gradients under differential privacy protection) in encrypted form, and aggregates them in the cloud to achieve global model collaborative evolution. Empirical evidence in reference

[29] shows that this mechanism can improve the overall accuracy of the model by more than 25%, while protecting the enterprise's data sovereignty and commercial privacy.

[0101] Ultimately, this application adopts a three-level fusion judgment mechanism (Mahalanx distance early warning + depth model classification + expert rule verification), which not only outputs whether it is abnormal, but also provides interpretable and actionable diagnostic conclusions.

[0102] Figure 6 A structural diagram of a pump fault diagnosis device provided in an embodiment of this application is shown below. Figure 6 As shown, based on the pump fault diagnosis method provided in the preceding embodiments, this application also provides a pump fault diagnosis device comprising: The calculation module is used to obtain the Mahalanobis distance of the target pump based on the dynamic acoustic fingerprint baseline library and the operating condition data of the target pump. The operating condition data includes acoustic data and physical data. The physical data includes at least one of inlet and outlet pressure data, flow data, motor current data, and medium temperature data. The baseline in the dynamic acoustic fingerprint baseline library is dynamically updated based on real-time operating condition data.

[0103] The diagnostic module is used to obtain the fault results of the target pump based on the operating condition data and fault model when the Mahalanobis distance is greater than or equal to a first preset threshold. The fault results include multiple fault types and the fault probability corresponding to each fault type.

[0104] The verification module is used to verify the fault results based on the operating condition data according to the fault rules, and obtain the final fault result. The final fault result includes the final fault type and the fault cause corresponding to the final fault type. Multiple fault types include the final fault type.

[0105] As an optional embodiment, the computing module specifically includes: The first preprocessing unit is used to process the operating condition data to obtain the time spectrum diagram corresponding to the operating condition data. The extraction unit is used to extract features from the time-spectrum graph to obtain a feature vector; the feature vector includes at least one of Mel frequency cepstral coefficients, spectral centroid, and impulse factor.

[0106] The building unit is used to construct multidimensional features based on time-spectrum maps and feature vectors.

[0107] The calculation unit is used to calculate the Mahalanobis distance of the target pump based on the dynamic acoustic fingerprint baseline library and multidimensional features.

[0108] As an optional embodiment, the preprocessing unit is specifically used for: The operating condition data is processed by short-time Fourier transform or maximum overlap discrete wavelet transform to obtain the time spectrum corresponding to the operating condition data.

[0109] As an optional embodiment, the diagnostic module is specifically used for: The second preprocessing unit is used to process the operating condition data to obtain the time-spectrum diagram and feature vector corresponding to the operating condition data.

[0110] The first channel unit is used to input the time-spectrum map into the first feature extraction channel in the fault model to obtain the fault features output by the first feature extraction channel; the first feature extraction channel includes a global average pooling layer and multiple convolutional layers.

[0111] The second channel unit is used to input the feature vector into the second feature extraction channel in the fault model to obtain the enhanced feature vector output by the second feature extraction channel; the second feature extraction channel includes a normalization processing layer and a bidirectional long short-term memory layer.

[0112] The third channel unit is used to input the fault features and enhanced feature vectors into the feature fusion classification channel in the fault model for fusion classification processing to obtain the fault result of the target pump; the feature fusion classification channel includes a feature fusion layer, a fully connected layer and a classifier.

[0113] The output unit is used to output the fault results of the target pump through the output layer in the fault model.

[0114] As an optional embodiment, the apparatus further includes: The baseline update module is used to smoothly migrate and update the baselines in the dynamic voiceprint fingerprint baseline library if the Mahalanobis distance is less than a first preset threshold and the working condition data meets the update conditions.

[0115] As an optional embodiment, the apparatus further includes: The early warning module is used to issue a fault warning if the Mahalanobis distance is greater than or equal to a first preset threshold and less than a second preset threshold, indicating that the target pump is in the first fault stage.

[0116] The alarm module is used to issue a fault alarm if the Mahalanobis distance is greater than or equal to the second preset threshold, the Mahalanobis distance is less than the third preset threshold, the fault probability corresponding to any fault type in the fault results is greater than the fourth preset threshold, and the fault time of any fault type in the fault results exceeds the preset time, indicating that the target pump is in the second fault stage.

[0117] The alarm module is used to trigger a fault alarm if the Mahalanobis distance is greater than or equal to the third preset threshold, the fault probability corresponding to any fault type in the fault results is greater than the fifth preset threshold, and the fault time of any fault type in the fault results exceeds a preset time, indicating that the target pump is in the third fault stage; the first preset threshold is less than the second preset threshold, which is less than the third preset threshold; the fourth preset threshold is less than the fifth preset threshold.

[0118] As an optional embodiment, the fault model is trained based on transfer learning, and the fault model is updated based on federated learning.

[0119] As an optional embodiment, the verification module is specifically used for: Based on the fault results, the weights of the voiceprint data and physical data are updated and adjusted to obtain the updated voiceprint data weights and updated physical data weights; based on the updated voiceprint data weights and updated physical data weights, the operating condition data is updated to obtain the updated operating condition data; the updated operating condition data includes the updated voiceprint data and updated physical data; according to the fault rules, the fault results are validated based on the updated operating condition data to obtain the final fault results.

[0120] As an optional embodiment, the apparatus further includes: The matching module is used to match the final fault result with the historical fault case library to determine the fault report; the fault report includes at least the final fault type and the corresponding maintenance strategy.

[0121] This application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for diagnosing pump malfunctions.

[0122] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for diagnosing pump malfunctions.

[0123] This application provides a computer program product, including a computer program that, when executed by a processor, implements a method for diagnosing pump malfunctions.

[0124] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and equipment embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and equipment embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0125] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for diagnosing faults in a pump, characterized in that, The method includes: Based on the dynamic acoustic fingerprint baseline library and the operating condition data of the target pump, the Mahalanobis distance of the target pump is obtained; the operating condition data includes acoustic data and physical data; the physical data includes at least one of inlet and outlet pressure data, flow data, motor current data, and medium temperature data; the baseline in the dynamic acoustic fingerprint baseline library is dynamically updated based on real-time operating condition data; When the Mahalanobis distance is greater than or equal to a first preset threshold, the failure result of the target pump is obtained based on the operating condition data and the failure model; the failure result includes multiple failure types and the failure probability corresponding to each failure type; The fault results are validated based on the operating condition data according to the fault rules to obtain the final fault result; the final fault result includes the final fault type and the fault cause corresponding to the final fault type; the multiple fault types include the final fault type.

2. The pump fault diagnosis method according to claim 1, characterized in that, The Mahalanobis distance of the target pump is obtained based on the dynamic voiceprint fingerprint baseline database and the operating condition data of the target pump, specifically including: The operating condition data is processed to obtain the time-spectrum diagram corresponding to the operating condition data; Feature extraction is performed on the time-spectrum graph to obtain a feature vector; the feature vector includes at least one of Mel frequency cepstral coefficients, spectral centroid, and impulse factor. Construct multidimensional features based on the time-spectrum diagram and the feature vector; The Mahalanobis distance of the target pump is calculated based on the dynamic voiceprint fingerprint baseline library and the multidimensional features.

3. The pump fault diagnosis method according to claim 2, characterized in that, The step of processing the operating condition data to obtain the time-spectrum diagram corresponding to the operating condition data specifically includes: The operating condition data is processed by short-time Fourier transform or maximum overlap discrete wavelet transform to obtain the time-spectrum diagram corresponding to the operating condition data.

4. The method for diagnosing pump faults according to claim 1, characterized in that, The process of obtaining the fault result of the target pump based on the operating condition data and fault model specifically includes: The operating condition data is processed to obtain the time-spectrum diagram and feature vector corresponding to the operating condition data; The time-spectrum image is input into the first feature extraction channel in the fault model to obtain the fault features output by the first feature extraction channel; the first feature extraction channel includes a global average pooling layer and multiple convolutional layers; The feature vector is input into the second feature extraction channel in the fault model to obtain the enhanced feature vector output by the second feature extraction channel; the second feature extraction channel includes a normalization processing layer and a bidirectional long short-term memory layer; The fault features and the enhanced feature vectors are input into the feature fusion classification channel in the fault model for fusion classification processing to obtain the fault result of the target pump; the feature fusion classification channel includes a feature fusion layer, a fully connected layer, and a classifier; The fault results of the target pump are output through the output layer in the fault model.

5. The method for diagnosing pump faults according to claim 1, characterized in that, The method further includes: If the Mahalanobis distance is less than the first preset threshold and the operating data meets the update conditions, then the baselines in the dynamic voiceprint fingerprint baseline library are smoothly migrated and updated.

6. The method for diagnosing pump faults according to claim 1, characterized in that, The method further includes: If the Mahalanobis distance is greater than or equal to the first preset threshold and less than the second preset threshold, it indicates that the target pump is in the first fault stage, and a fault warning is issued. If the Mahalanobis distance is greater than or equal to the second preset threshold, the Mahalanobis distance is less than the third preset threshold, the fault probability corresponding to any fault type in the fault results is greater than the fourth preset threshold, and the fault time of any fault type in the fault results exceeds the preset time, it indicates that the target pump is in the second fault stage, and a fault alarm is issued. If the Mahalanobis distance is greater than or equal to the third preset threshold, the fault probability corresponding to any fault type in the fault results is greater than the fifth preset threshold, and the fault time of any fault type in the fault results exceeds the preset time, indicating that the target pump is in the third fault stage, then a fault alarm is triggered; the first preset threshold is less than the second preset threshold, which is less than the third preset threshold; the fourth preset threshold is less than the fifth preset threshold.

7. The method for diagnosing pump faults according to claim 1, characterized in that, The fault model is trained based on transfer learning, and its parameters are updated based on federated learning.

8. The method for diagnosing pump faults according to claim 1, characterized in that, The step of performing rule verification on the fault result based on the operating condition data according to the fault rules to obtain the final fault result specifically includes: Based on the fault results, the voiceprint data weights and physical data weights are updated and adjusted to obtain the updated voiceprint data weights and updated physical data weights. The operating condition data is updated based on the updated voiceprint data weights and the updated physical data weights to obtain updated operating condition data; the updated operating condition data includes updated voiceprint data and updated physical data. The fault result is verified according to the fault rules based on the updated operating condition data to obtain the final fault result.

9. The method for diagnosing pump faults according to claim 1, characterized in that, The method further includes: The final fault result is matched with the historical fault case database to determine the fault report; the fault report includes at least the final fault type and the maintenance strategy corresponding to the final fault type.

10. A fault diagnosis device for a pump, characterized in that, The fault diagnosis device for the pump includes: The calculation module is used to obtain the Mahalanobis distance of the target pump based on the dynamic acoustic fingerprint baseline library and the operating condition data of the target pump; the operating condition data includes acoustic data and physical data; the physical data includes at least one of inlet and outlet pressure data, flow data, motor current data, and medium temperature data; the baseline in the dynamic acoustic fingerprint baseline library is dynamically updated based on real-time operating condition data; The diagnostic module is used to obtain the fault result of the target pump based on the operating condition data and the fault model when the Mahalanobis distance is greater than or equal to a first preset threshold; the fault result includes multiple fault types and the fault probability corresponding to each fault type; The verification module is used to perform rule verification on the fault result based on the operating condition data according to the fault rules, and obtain the final fault result; the final fault result includes the final fault type and the fault cause corresponding to the final fault type; the multiple fault types include the final fault type.

Images