A pump set health monitoring and predictive maintenance system

Abstract

This invention discloses a pump set health monitoring and predictive maintenance system, relating to the field of digital twin technology. The system includes: an operating parameter acquisition module for acquiring a pump set operating parameter matrix; a feature acquisition module for acquiring the frequency domain energy distribution characteristics of the fluid end pressure time series; a sparsification processing module for sparsifying the hydraulic transient tensor; a coupling weight generation module for generating a coupling weight masking matrix; a masking and denoising module for masking and denoising the mechanical degradation feature matrix based on the coupling weight masking matrix to generate a purified degradation feature matrix; a health index output module for inputting the purified degradation feature matrix into a pre-trained health state evolution model and outputting the corresponding pump set health index; and a maintenance signal generation module for generating a maintenance guidance signal with the location coordinates of the target component when the pump set health index is less than or equal to a preset alarm threshold. This invention improves the accuracy of identifying early latent hydraulically induced faults in pump sets.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, specifically to a pump set health monitoring and predictive maintenance system. Background Technology

[0002] Pump sets, as core power equipment in fluid transportation, industrial water circulation, chemical media transfer, and water supply and drainage systems, are widely used in key fields such as power, chemical, municipal, water conservancy, and intelligent manufacturing. Pump sets operate in a complex environment of continuous variable conditions, high load, and media impact for a long time, making them prone to progressive failures such as cavitation, bearing wear, impeller fatigue, pipeline resonance, and seal failure. If the degradation trend cannot be detected in advance and preventive maintenance is not carried out in time, it can easily lead to problems such as reduced pump output, interruption of media transportation, equipment shutdown, or even shutdown of the entire production line, which seriously affects the continuity of industrial production and the safety of equipment operation.

[0003] Currently, most pump health monitoring and maintenance solutions in the industry adopt a post-fault alarm mode based on single parameter threshold comparison and independent time-domain signal analysis. They rely primarily on individual mechanical vibration signals or fluid pressure signals for status determination, failing to consider the hydraulic-mechanical coupling time-delay degradation characteristics commonly found in pump failures. This results in poor accuracy in identifying early latent faults, significant false negatives and missed positives, and an inability to meet predictive maintenance needs. Specifically, during actual pump operation, hydraulic transient excitation and mechanical structural response do not occur synchronously, exhibiting significant dynamic time-delay characteristics. Early equipment degradation is often initially induced by fluid-side anomalies, then delayed and transmitted to the mechanical structure. For example, when a sudden change in flow rate or fluctuation in inlet negative pressure occurs during pump operation, latent anomalies such as hydraulic transient impacts and micro-cavitation will first appear, causing high-frequency subtle disturbances in the fluid pressure sequence. However, the vibration amplitude of the pump at this stage... When conventional monitoring parameters such as temperature do not change significantly, a single mechanical monitoring method cannot capture such early hydraulic anomalies. By the time parameters such as mechanical vibration and equipment temperature rise significantly exceed the standard, problems such as impeller erosion and wear, bearing micro-deformation, and pipeline resonance damage have already caused irreversible structural degradation, and the fault has developed to the middle and late stages, completely missing the best opportunity for early warning and preventive maintenance. In addition, a few existing pump set monitoring solutions attempt to introduce digital twin technology for equipment status replication and simulation monitoring, but traditional digital twin models are mostly static structural mappings and single state data loading, which can only achieve simple visualization of equipment appearance and conventional operating parameters. They have not established a dynamic coupling relationship between hydraulic transient characteristics and mechanical degradation characteristics, and cannot adapt to the degradation mechanism of pump set fluid-structure time delay response. They are not capable of identifying early hidden faults and are difficult to achieve high-precision health assessment and predictive early warning. Summary of the Invention

[0004] To address the aforementioned technical problems, a pump set health monitoring and predictive maintenance system is provided, which solves the problems described above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A pump set health monitoring and predictive maintenance system, comprising: The operating parameter acquisition module acquires the pump set operating parameter matrix, which includes the fluid end pressure time series and the mechanical end vibration time series. The feature acquisition module acquires the frequency domain energy distribution features of the fluid end pressure time series and generates a hydraulic transient tensor; it also acquires the time domain envelope features of the mechanical end vibration time series and generates a mechanical degradation feature matrix. The sparsification module performs sparsification on the hydraulic transient tensor, setting elements with energy variance no higher than the preset noise floor to zero in order to reduce the data size of subsequent processing. The coupling weight generation module constructs a time-delay cross-correlation analysis window, which slides along the time axis of the sparsified hydraulic transient tensor to determine the cross-correlation sequence between the hydraulic transient tensor and the mechanical degradation feature matrix under different time lag factors. The time lag factor corresponding to the peak with the largest absolute value in the cross-correlation sequence is selected as the optimal fluid-structure interaction (FSI) time lag. Based on the optimal FSI time lag, the hydraulic transient tensor is shifted along the time axis to complete the calibration. The calibrated hydraulic transient tensor is then input into a nonlinear activation function for mapping to obtain the hydraulic modulation intensity tensor. The hydraulic modulation intensity tensor and the mechanical degradation feature matrix are then subjected to Hadamard fusion processing to generate the coupling weight masking matrix. The masking and denoising module performs masking and denoising on the mechanically degraded feature matrix based on the coupled weight masking matrix to generate a purified degraded feature matrix; The health index output module inputs the purified degradation feature matrix into the pre-trained health state evolution model and outputs the corresponding pump group health index. The maintenance signal generation module generates a maintenance guidance signal with the location coordinates of the target component when the pump group health index is less than or equal to the preset alarm threshold.

[0006] Furthermore, a hydraulic transient tensor is generated; the temporal envelope features of the mechanical end vibration time series are obtained, and a mechanical degradation feature matrix is ​​generated, including: Perform discrete wavelet transform on the fluid end pressure time series to obtain multiple pressure subsequences in different frequency bands; determine the local energy variance of each pressure subsequence; concatenate the local energy variances into a one-dimensional feature vector according to the preset frequency from low to high, and then convert the one-dimensional feature vector into a two-dimensional hydraulic transient tensor. Obtain the root mean square value and peak factor of the mechanical end vibration time series within a preset sliding time window; map the root mean square value and peak factor to a preset geometric topology grid to generate a mechanical degradation feature matrix with spatial distribution attributes.

[0007] Furthermore, the health index output module inputs the purified degradation feature matrix into the pre-trained health state evolution model and outputs the corresponding pump group health index, including: The purified degradation feature matrix is ​​input into the feature extraction layer of the healthy state evolution model to obtain the hidden layer state vector; The hidden layer state vector is input to the fully connected layer for nonlinear mapping, and the output value reflecting the degree of equipment degradation is used as the pump group health index.

[0008] Furthermore, before obtaining the pump unit operating parameter matrix, a pre-training process for the health state evolution model is also included, including: Obtain the historical sample matrix; the historical sample matrix contains historical hydraulic transient tensors and historical mechanical degradation feature matrices labeled with actual health status. The historical hydraulic transient tensor and the historical mechanical degradation feature matrix are input into the initial network model for forward propagation, and the predicted health status index is output. Determine the basic mean square error between the predicted health status index and the actual health status label; obtain the high-frequency perturbation coefficients in the historical hydraulic transient tensor that indicate fluid cavitation; fuse the high-frequency perturbation coefficients with the basic mean square error to generate a penalty loss function; The gradient value is determined based on the penalty loss function. The weight parameters of the initial network model are updated according to the gradient descent direction. When the penalty loss function converges to the preset interval, the parameter update is stopped and the model parameters are saved to obtain the health state evolution model.

[0009] Furthermore, the masking and denoising module performs masking and denoising on the mechanically degraded feature matrix based on the coupled weight masking matrix to generate a purified degraded feature matrix, including: Obtain the feature elements in the coupling weight masking matrix whose values ​​are not lower than a preset masking threshold; Construct an adjacency graph of the mechanical degradation feature matrix and determine the local curvature of each feature element and its neighboring feature elements; For feature elements whose values ​​are not lower than a preset masking threshold, a Laplacian operator is constructed using feature elements whose local curvature in their neighborhood is not higher than a preset smoothing threshold. Based on the Laplacian operator, manifold interpolation is performed on feature elements whose values ​​are not lower than a preset masking threshold to generate repair feature values. The purified degradation feature matrix is ​​obtained by replacing the corresponding element values ​​in the mechanical degradation feature matrix with the repair feature values.

[0010] Furthermore, after generating the maintenance guidance signaling with the target component's location coordinates, it also includes: Obtain coordinate request commands sent by external augmented reality terminals; The maintenance guidance signaling will be parsed into a three-dimensional spatial mapping matrix; The three-dimensional spatial mapping matrix is ​​sent to an external augmented reality terminal, and the bounding box of the three-dimensional geometric model of the faulty component is overlaid and rendered on the display interface of the external augmented reality terminal.

[0011] Furthermore, when the pump group health index is less than or equal to a preset alarm threshold, the maintenance signal generation module generates a maintenance guidance signal with the location coordinates of the target component, including: Real-time acquisition of the pump set's current operating parameters, which include at least the spindle speed and outlet flow rate; The operating condition parameters are input into the pre-constructed operating condition-threshold mapping surface to determine the current adaptive alarm threshold; the operating condition-threshold mapping surface records the baseline fluctuation range of the pump group health index under different operating conditions. Determine the slope of the deviation between the pump unit health index and the adaptive alarm threshold; When the deviation slope remains negative for a duration not less than the preset hysteresis time window, the pump unit is determined to have entered the fault evolution stage, and a maintenance guidance signal with the location coordinates of the target component is generated.

[0012] Furthermore, the operating parameter acquisition module acquires the pump set operating parameter matrix, including: Simultaneously acquire simulated pressure signals at the fluid end and simulated vibration signals at the mechanical end of the pump unit using a preset sampling frequency; The pressure analog signal and the vibration analog signal are converted from analog to digital respectively to obtain the original digital signal sequence; Outlier removal is performed on the original digital signal sequence, replacing outliers whose values ​​exceed the preset sensor range with the center value of the signal at adjacent time points. The signal after removing outliers is subjected to wavelet threshold denoising, and the denoised pressure signal and vibration signal are aligned and spliced ​​according to the timestamp to construct the pump group operation parameter matrix.

[0013] Furthermore, the feature extraction layer of the health state evolution model adopts a graph convolutional network structure; The purified degradation feature matrix is ​​input into the feature extraction layer of the healthy state evolution model, including: The purified degenerate feature matrix is ​​treated as graph-structured data, where each element of the matrix is ​​a graph node. The features of neighboring nodes are aggregated using graph convolution kernels, the feature representation of the current node is updated, and the hidden layer state vector is obtained.

[0014] Furthermore, after overlaying and rendering the 3D geometric model bounding box of the faulty component onto the display interface of the external augmented reality terminal, it also includes: In response to user interaction on an external augmented reality terminal, the historical maintenance records and recommended spare parts list corresponding to the location coordinates of the target component are analyzed. Obtain the current camera pose parameters and field of view of the external augmented reality terminal; Based on camera pose parameters, the text and graphic data of historical maintenance records and recommended spare parts lists are converted into three-dimensional spatial anchor points, and the spatial occlusion relationship between the three-dimensional spatial anchor points and the bounding box of the three-dimensional geometric model of the faulty component is determined. Based on spatial occlusion and field of view, the historical maintenance records and recommended spare parts list are adaptively rendered in the display interface of the external augmented reality terminal in a manner that conforms to the laws of real visual perspective.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention simultaneously collects time-series data of pressure at the fluid end and vibration at the mechanical end of the pump unit, and constructs hydraulic transient tensors and mechanical degradation feature matrices respectively. It can effectively identify hidden defects that traditional solutions cannot detect, such as early cavitation, micro-wear, and potential mechanical degradation caused by hydraulic excitation, thus solving the problem of missed early fault detection and improving the early fault perception capability of the pump unit. Through tensor sparsification and manifold masking denoising, it improves the anti-interference capability in complex industrial scenarios. This invention performs sparsification and zeroing of the hydraulic transient tensor to reduce noise, eliminate invalid noise data, and reduce the computational scale. At the same time, based on the fluid-structure interaction weighted masking matrix combined with neighborhood curvature and Laplace interpolation, it achieves accurate denoising of mechanical features. While retaining weak and effective degradation features, it filters out interference noise caused by industrial clutter and operating condition fluctuations, solving the problems of feature submersion and excessive denoising in current filtering algorithms, and ensuring the authenticity and effectiveness of degradation features. 2. By introducing time-delay cross-correlation analysis and dynamic time-series calibration, this invention solves the problem of time-series misalignment in fluid-structure interaction response. It calculates the optimal coupling time delay of fluid-structure characteristics through a sliding time-delay window, adaptively completing the hydraulic characteristic time-series translation calibration. This matches the true response time-series relationship between hydraulic excitation and mechanical degradation, avoiding the feature distortion problem caused by traditional fixed time-series alignment and improving the accuracy and reliability of multi-source feature fusion. Furthermore, it constructs an adaptive threshold early warning mechanism based on operating conditions. Unlike fixed alarm thresholds, it builds an operating condition-threshold mapping surface based on real-time operating parameters such as rotational speed and flow rate, dynamically adapting to the health baseline fluctuation range under different operating conditions. Simultaneously, it combines the deviation slope and hysteresis time window to determine the fault evolution trend, effectively distinguishing between normal operating condition fluctuations and true fault degradation trends, thus improving the operating condition adaptability and engineering stability of the monitoring system. 3. This invention employs a graph convolutional health evolution model with penalized loss to improve the accuracy of health index assessment. During model training, the invention integrates cavitation high-frequency perturbation coefficients to construct a dedicated penalized loss function, specifically constraining assessment biases caused by fluid anomalies. Simultaneously, relying on graph convolutional networks to aggregate feature neighborhood association information, it accurately mines the progressive degradation patterns of equipment, outputs a high-precision pump group health index, and quantifies the degree of equipment degradation, achieving a technological upgrade from fault alarm to quantitative health assessment. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of the present invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Example 1: Refer to Figure 1 As shown, a pump set health monitoring and predictive maintenance system includes: The system comprises a parameter acquisition module, a feature acquisition module, a sparsity processing module, a coupling weight generation module, a mask denoising module, a health index output module, and a maintenance signaling generation module. These modules work together to complete multi-source time-series data acquisition for the pump group, flow-mechanical coupling feature modeling, intelligent denoising, quantitative health assessment, and visualized predictive maintenance guidance. The specific implementation steps are as follows: Step 101: The operating parameter acquisition module acquires the pump set operating parameter matrix. Specifically, the system synchronously acquires the pump set fluid end pressure time-series signal and mechanical end vibration time-series signal through the front-end sensor acquisition unit. After signal preprocessing, outlier removal, wavelet denoising, and time-series alignment, a standardized two-dimensional operating parameter matrix is ​​constructed. The matrix dimension is N×2, where N is the total number of sampling time-series points. The row dimension strictly corresponds to the unified sampling time. The first column corresponds to the pressure time-series data, and the second column corresponds to the vibration time-series data, achieving one-to-one matching of the fluid and mechanical dual-source data with no phase deviation, providing a standardized raw data base for subsequent multi-source coupling feature analysis. For example, the acquired pressure time-series signal can be the dynamic pressure fluctuation data at the inlet and outlet of the centrifugal pump, and the vibration time-series signal can be the high-frequency vibration data of the pump set bearing housing and pump body frame. It can be understood that sensor acquisition data that can reflect the fluid working conditions and mechanical operating status during the pump set operation can be used as the data source for constructing the above operating parameter matrix. Of course, it is not limited to only pressure and vibration data. Auxiliary parameters such as temperature and speed can be added according to monitoring needs. Step 102: The feature acquisition module generates hydraulic transient tensors and mechanical degradation feature matrices respectively. Specifically, for the fluid end pressure time series, a multi-band discrete wavelet frequency domain decomposition method is used to separate the steady-state low-frequency components and the disturbance high-frequency components, extracting energy fluctuation features band by band to accurately capture hydraulic anomalies such as pump cavitation, turbulence, and pressure pulsation. A two-dimensional hydraulic transient tensor with spatiotemporal distribution characteristics is generated through dimensionality upscaling and reconstruction. For the mechanical end vibration time series, a fixed-duration sliding window time domain statistical method is used to extract vibration energy and impact peak degradation features window by window, transforming a... The time-domain features are mapped to a pre-defined pump component topology mesh space to generate a two-dimensional mechanical degradation feature matrix that can accurately characterize bearing wear, impeller fatigue, and pipeline resonance. For example, the hydraulic transient tensor can be used to characterize transient hydraulic anomalies such as low-flow cavitation and high-load turbulent impact in the pump unit, while the mechanical degradation feature matrix can characterize mechanical fault features such as rolling bearing pitting wear, impeller fatigue under long-term alternating loads, and inlet / outlet pipeline resonance vibration. It can be understood that all hydraulic anomalies caused by fluid disturbances and vibration anomalies caused by mechanical component degradation can be structurally modeled through this step. Step 103: The sparsity processing module performs sparsity denoising processing on the hydraulic transient tensor. Specifically, it iterates through all elements within the hydraulic transient tensor and uses a local temporal window energy variance algorithm to calculate the feature energy point by point: taking a single element as the center, it extracts consecutive fixed sampling points before and after it to form a local calculation window. First, it calculates the arithmetic mean of all elements within the window, then calculates the squared residuals between each element and the mean point by point. After accumulating all squared residuals, it divides by the total number of samples in the window to obtain the local energy variance at that location. The calculated local energy variance is compared with the system's preset noise floor threshold, and weak responses with energy variances not higher than the noise floor threshold are identified. Elements are uniformly set to zero, retaining only valid transient features with energy variance higher than the threshold. This processing method can accurately eliminate invalid features caused by sensor noise and minor disturbances in operating conditions, significantly reducing the data computation scale of subsequent coupled calculations and model inference, and fully retaining the valid feature components corresponding to real hydraulic transient anomalies. For example, sensor noise may include weak random fluctuations caused by electromagnetic interference from sensor circuits, and minor disturbances in operating conditions may include invalid signals such as small and stable fluctuations in pipeline pressure and slight baseline jitter during steady-state operation of pump units. It can be understood that all interference features with weak energy fluctuations and no fault characterization significance can be eliminated by setting them to zero in this step. Step 104: The coupling weight generation module constructs the fluid-structure time-delay coupling relationship and generates the coupling weight masking matrix. Specifically, the system presets a time-delay cross-correlation analysis window of fixed duration, and continuously slides the window along the time axis of the sparsed hydraulic transient tensor to perform fluid-structure time-series cross-correlation calculations window by window. Specifically: For each sliding window, a global time lag factor set is preset, and all lag factors are traversed to calculate the Pearson cross-correlation coefficient between the hydraulic transient characteristic sequence and the mechanical degradation characteristic sequence under different lag durations. The cross-correlation coefficient is calculated by taking the mean of the two sets of sequences, calculating the sum of the products of the sequence deviations, and then dividing by the product of the standard deviations of the two sets of sequences to obtain the correlation coefficient in the interval [-1,1]. The larger the absolute value of the coefficient, the stronger the temporal coupling correlation. The peak values ​​of the correlation coefficients corresponding to all lag factors are traversed, and the lag duration corresponding to the peak with the largest absolute value is selected as the optimal fluid-structure interaction time delay to accurately match the actual response delay of the hydraulic excitation to the mechanical structure. Based on the optimal fluid-structure interaction time delay, the hydraulic transient tensor is calibrated by shifting the entire time axis to completely eliminate the problem of timing misalignment in the fluid-machine response. The calibrated hydraulic transient tensor is then input into the Sigmoid nonlinear activation function to complete the feature intensity normalization mapping, compressing all elements of the tensor to the 0-1 interval, resulting in a hydraulic modulation intensity tensor that can quantitatively characterize the severity of hydraulic disturbances. Finally, the hydraulic modulation intensity tensor and the mechanical degradation feature matrix are fused by performing element-wise Hadamard product fusion calculation, i.e., elements at the same row and column positions are multiplied pairwise, and different positions are calculated independently, generating a coupling weight shielding matrix that combines fluid excitation characteristics and mechanical degradation characteristics, realizing deep coupling correlation of fluid-machine features. For example, the hydraulic impact excitation generated by pump cavitation usually lags by 100ms to 300ms, causing abnormal bearing vibration. This step can accurately match this delay and complete the timing calibration. It can be understood that all scenarios of lag response from hydraulic excitation to mechanical structure can be coupled and modeled through this step, although it is not limited to cavitation excitation scenarios. Step 105: The masking and noise reduction module completes mechanical feature purification based on the coupled weight masking matrix. Specifically, it uses the spatial distribution weight values ​​of the coupled weight masking matrix as the discrimination criterion, and combines a preset masking threshold to divide the noise interference area and the effective degradation area. For the high-weight interference area, an eight-neighborhood topology is constructed, the curvature of the local surface is calculated to identify abrupt noise points, and the interference area features are smoothed and corrected by the manifold neighborhood weighted interpolation algorithm. The false feature interference caused by industrial clutter and hydraulic pseudo-response is accurately removed, and the true mechanical degradation features of the equipment are completely preserved. Finally, a purified degradation feature matrix with a high signal-to-noise ratio is generated. For example, industrial clutter may include electromagnetic vibration interference generated by the start-up and shutdown of frequency converters and motors, and hydraulic pseudo-response may include false vibration peaks caused by steady-state water flow disturbance. It can be understood that all false vibration features caused by the degradation of non-mechanical components can be removed through this step. Of course, this is not limited to the above two types of interference scenarios. Step 106: The health index output module quantifies the pump group's health status. Specifically, the purified high-confidence degradation feature matrix is ​​input into the trained health status evolution model. Through multi-layer graph convolutional neighborhood aggregation, deep feature mining, and high-dimensional feature nonlinear mapping, the model finally outputs a pump group health index continuously distributed in the range of 0 to 1. The closer the value is to 0, the higher the degree of equipment degradation and the greater the risk of failure. The closer the value is to 1, the better the equipment's operating status and the better the equipment's performance. This achieves continuous and refined quantitative evaluation of the equipment status. For example, a health index of 0.9 to 1.0 corresponds to a brand-new, non-degrading, and stable, efficient operating state of the pump group; 0.6 to 0.9 corresponds to a normal degradation state with slight wear and no risk of failure; 0.3 to 0.6 corresponds to a moderate degradation state requiring regular inspection; and 0 to 0.3 corresponds to a severe degradation state, a high-risk state that is about to fail and shut down. It can be understood that the continuous values ​​in the range of 0 to 1 can accurately cover all health states of the pump group throughout its entire life cycle. Of course, this is not limited to using only the above range division method and can be adaptively fine-tuned according to the scenario. Step 107: The maintenance signaling generation module outputs intelligent maintenance guidance signals. Specifically, the system collects pump unit operating parameters in real time, dynamically calculates adaptive alarm thresholds, and continuously compares the real-time pump unit health index with the alarm thresholds adapted to the current operating conditions. When the health index is less than or equal to the adaptive alarm threshold and meets the timing hysteresis judgment condition, the system determines that the equipment has entered the continuous fault evolution stage. Combining the preset component three-dimensional coordinate mapping relationship, it generates refined maintenance guidance signals carrying the three-dimensional position coordinates of the faulty component, fault type, and degradation level, realizing a technical upgrade from a single over-limit alarm to precise positioning and hierarchical operation and maintenance. For example, if impeller cavitation degradation is detected, causing the health index to continuously fall below the adaptive threshold, a maintenance signal carrying the impeller's three-dimensional coordinates, cavitation fault type, and moderate degradation level can be generated. If bearing wear exceeds the standard, an operation and maintenance signal corresponding to the bearing position, wear fault type, and severe degradation level can be generated. It can be understood that the degradation faults of all key components of the pump unit can be used to generate refined maintenance signals. Through the collaborative work of the above modules, early latent degradation features are accurately extracted by time-delay coupling modeling and adaptive masking noise reduction. The health status is quantified by relying on graph convolutional intelligent models. Combined with the working condition adaptive early warning mechanism, predictive maintenance instructions that can be implemented are output, thereby improving the monitoring accuracy of pump groups and the level of intelligent operation and maintenance.

[0019] In some embodiments, the feature generation process of the feature acquisition module specifically includes the following steps: Step 201: Construction of the hydraulic transient tensor at the fluid end; Specifically, a preset number of db-series discrete wavelet transforms are selected to perform multi-scale decomposition on the fluid end pressure time series, separating the low-frequency steady-state pressure component from the high-frequency transient disturbance component, and obtaining multiple pressure subsequences in different frequency bands. The low-frequency component corresponds to steady-state conditions and fault-free information, while the high-frequency component corresponds to transient anomalies such as cavitation and pressure pulsation. Local window energy variance calculation is performed on each high-frequency pressure subsequence of each frequency band. The calculation logic is as follows: a fixed-duration local window is extracted, the mean of the signal within the window is calculated, the sum of squared residuals is calculated point by point and the mean is calculated to obtain the energy variance. The magnitude of the energy variance directly and quantitatively reflects the corresponding frequency. The intensity of hydraulic pulsation and cavitation disturbance is determined. Following a fixed frequency sorting rule from low to high frequency, the local energy variances corresponding to all frequency bands are sequentially concatenated into a one-dimensional feature vector. This one-dimensional feature vector is then reconstructed by increasing its dimensionality, transforming it into a two-dimensional spatiotemporal distribution matrix to obtain a hydraulic transient tensor that can fully characterize the spatiotemporal variation of the pump unit's hydraulic transients. For example, the low-frequency pressure component corresponds to the steady-state water delivery condition of the pump unit with constant speed and constant flow, while the high-frequency disturbance component corresponds to the cavitation impact and turbulent pulsation conditions caused by pipeline throttling and inlet negative pressure. It can be understood that all hydraulic condition changes that can cause transient pressure fluctuations can have their features extracted through this step. Step 202: Construction of the mechanical end degradation feature matrix; Specifically, the system presets a fixed-duration sliding time window of 100ms to 500ms to perform window-by-window traversal processing on the mechanical end vibration time series; within each sliding window, two types of core time-domain features are calculated: the first type is the root mean square value of vibration, calculated by summing the squares of all vibration amplitudes within the window, dividing by the number of sampling points, and taking the square root, used to characterize the overall vibration energy and uniform wear degree of the equipment; the second type is the peak factor, calculated as the ratio of the absolute value of the vibration peak value to the root mean square value within the window, used to quantitatively characterize the characteristics of local impact failures; a two-dimensional matrix adapted for multi-component monitoring of pump sets is pre-constructed. A three-dimensional geometric topological mesh is created, with each mesh node corresponding to a monitoring area of ​​key components in the pump unit, such as bearings, impellers, seals, and pipelines. The root mean square (RMS) values ​​and peak factors extracted window by window are weighted and mapped to the corresponding topological mesh nodes to complete the meshed distribution of the feature space. Finally, a mechanical degradation feature matrix with accurate spatial component attributes is generated. For example, a continuous and slow increase in the vibration RMS value can correspond to uniform bearing wear and aging degradation of seals, while a sudden increase in the peak factor can correspond to impact failures such as impeller blade impact and bearing rolling element pitting. It can be understood that the uniform degradation and sudden impact failure characteristics of the pump unit's mechanical components can be modeled through this step. In one implementation, the discrete wavelet transform can use the db4 wavelet basis, and the number of decomposition layers can be adaptively set to 3 to 5 layers according to the sampling frequency, balancing frequency domain decomposition accuracy and real-time computation efficiency. The sliding time window is set to 200ms by default, ensuring that it can capture short-term impact degradation features while avoiding distortion of weak fault features due to an excessively large window. For example, a 5-layer wavelet decomposition is used at a sampling frequency of 1kHz, which can accurately distinguish subtle high-frequency hydraulic disturbances; a 3-layer wavelet decomposition is used at a sampling frequency of 500Hz, which can reduce computational overhead and ensure real-time performance. It can be understood that the number of wavelet decomposition layers can be adaptively adapted to the hardware sampling parameters. Through the above technical solutions, hydraulic transient anomalies are analyzed in detail from the frequency domain dimension and mechanical degradation distribution is quantitatively characterized from the time domain spatial dimension. This enables structured and gridded modeling of fluid excitation characteristics and mechanical degradation characteristics, providing standardized and high-dimensional feature inputs for subsequent high-precision fluid-structure interaction analysis.

[0020] In some embodiments, the health quantification process of the root health index output module specifically includes the following steps: Step 301: Deep Hidden Feature Extraction; Specifically, the preprocessed purified degradation feature matrix is ​​completely input into the graph convolution feature extraction layer of the health state evolution model. The feature extraction layer automatically mines the hidden component degradation correlation rules, long-term evolution trends and spatial distribution differences within the feature matrix through multi-layer graph convolution nonlinear transformation and neighborhood feature weighted aggregation. It integrates multi-dimensional hidden features of hydraulic disturbance, mechanical wear and time-series degradation, and outputs a hidden layer state vector with fixed dimensions and complete feature information. For example, it can mine the coupled degradation correlation rules of long-term cavitation hydraulic disturbance plus continuous bearing vibration wear, and can also identify the long-term evolution trend of impeller fatigue accumulation degradation caused by pipeline resonance. It can be understood that all hidden degradation rules of multi-component coupling and time-series accumulation can be mined through this step. Of course, this is not limited to the above-mentioned coupled degradation scenarios. Step 302: Nonlinear mapping output of the health index; Specifically, the extracted hidden layer state vector is input into the fully connected layer at the end of the model. The fully connected layer completes the nonlinear mapping from high-dimensional latent features to one-dimensional quantized values ​​through a preset weight matrix, bias parameters, and nonlinear activation function, and outputs a value normalized to the range of 0 to 1. This normalized value is defined as the pump group health index, realizing a refined and continuous quantitative assessment of the overall degradation degree of the equipment. For example, for the coupled degradation scenario of slight cavitation + slight bearing wear, a health index of 0.82 can be output; for the scenario of moderate cavitation plus impeller fatigue degradation, a health index of 0.45 can be output. It can be understood that different degrees of composite degradation scenarios can output accurate corresponding continuous quantified values. By using the above technical solution, relying on the autonomous feature mining capability of the depth map convolution model, the subjectivity and one-sidedness of the traditional manual fixed threshold judgment are avoided. It can accurately distinguish the subtle differences between slight degradation, moderate degradation and severe degradation, and the evaluation results are consistent with the actual degradation mechanism of the equipment.

[0021] In some embodiments, the pre-training process of the health state evolution model specifically includes the following steps: Step 401: Construct a labeled historical sample dataset. Specifically, collect pump set operation data throughout its entire lifecycle, covering all operating conditions including normal operation, minor cavitation, bearing micro-wear, severe faults, and shutdown failures. Perform corresponding standardized feature extraction operations on the collected historical time-series data to generate historical hydraulic transient tensors and historical mechanical degradation feature matrices in batches. Combine equipment disassembly and inspection reports, official maintenance records, and fault logs to label each feature sample with a continuous and precise health status label from 0 to 1, with normal equipment approaching 1 and failed equipment approaching 0, thus constructing a large-scale historical sample dataset with ground truth labels. For example, a new pump set steady-state operation sample is labeled with 0.98, a sample with minor wear after one year of operation is labeled with 0.85, a sample with moderate cavitation wear is labeled with 0.50, a sample with severe jamming fault is labeled with 0.10, and a sample with complete shutdown failure is labeled with 0.02. This can be understood as all operating states throughout the pump set's entire lifecycle corresponding to precise and continuous labels. Step 402: Forward propagation prediction of the model; specifically, the batch of historical hydraulic transient tensors and historical mechanical degradation feature matrices are input into the initial graph convolutional network model in batches, and feature fusion, deep graph convolutional mining, and state mapping are completed sequentially, and the predicted health status index corresponding to each group of samples is output in batches; for example, a batch of mixed samples containing cavitation degradation, bearing wear, and steady-state operation are input into the model, and the model can output the predicted health index corresponding to each group of samples in batches, realizing parallel inference training; it can be understood that batch samples with any combination of working conditions can complete forward prediction; Step 403: Construct the cavitation perception penalty loss function; Specific calculation process: First, calculate the basic mean square error between the predicted health index and the manually labeled true value, obtain the square of the difference between the predicted value and the true value, and calculate the batch mean as the basic supervisory loss of the model to constrain the overall prediction bias; simultaneously, analyze the historical hydraulic transient tensor, statistically analyze the ratio of the total energy of the high-frequency band to the total energy of the entire frequency band, obtain the high-frequency energy disturbance coefficient, accurately characterizing the severity of high-risk fluid anomalies such as pump cavitation and hydraulic shock; use this high-frequency disturbance coefficient as a dynamic penalty weight, and weightedly fuse it with the basic mean square error to construct a dedicated penalty loss function; the stronger the cavitation disturbance, the greater the penalty weight, and the stronger the model's constraint on the prediction error of cavitation-induced degradation, realizing the embedding of prior knowledge of the fault mechanism; for example, under strong cavitation conditions, the high-frequency disturbance coefficient can reach above 0.7, and the model's penalty for the prediction error of this type of sample is greatly improved; under weak cavitation conditions, the disturbance coefficient is only about 0.2, and the penalty intensity is moderately reduced to ensure the prediction accuracy of steady-state conditions; this can be understood as the penalty weight dynamically and adaptively adjusting with the intensity of hydraulic anomalies; Step 404: Iterative convergence training of the model; specifically, based on the inverse chain derivative of the penalty loss function, the gradient values ​​of the weights and biases of each layer of the network are solved. The Adam adaptive gradient descent algorithm is used to update the model parameters layer by layer and dynamically adjust the learning rate; continuous iterative training is carried out, and the training set loss and the validation set prediction error are monitored in real time. When the penalty loss function stably converges to the preset minimum interval, the validation set error no longer decreases and there is no overfitting, the parameter update is stopped, the optimal model parameters are fixed and saved, and the training of the health state evolution model is completed; for example, after 1000 rounds of iterative training, the loss value stably converges to within 0.005, and the validation set health index prediction error is stably below 2%, with no overfitting, the model parameters can be fixed; this can be understood as the convergence threshold and the number of iterations can be finely adjusted according to the equipment monitoring accuracy requirements.

[0022] By introducing the prior knowledge of pump cavitation failure mechanism during model training and adding dynamic penalty constraints for hydraulic anomalies, the above technical solutions solve the problems of traditional models ignoring the influence of fluid excitation and having large deviations in degradation assessment, thereby improving the model's assessment accuracy and generalization ability for hydraulic-mechanical coupling degradation scenarios.

[0023] In some embodiments, the feature purification process of the mask noise reduction module specifically includes the following steps: Step 501: Locate high-weight interference feature regions. Specifically, traverse all elements of the coupling weight shielding matrix and compare the matrix element values ​​with the system's preset shielding threshold. If the element value is greater than or equal to the shielding threshold, the location is determined to be a false high-response feature region formed by hydraulic pseudo-response or environmental clutter interference, and marked as a noise region to be corrected. If the element value is less than the shielding threshold, it is determined to be a real degradation feature region, and the original value is directly retained. For example, hydraulic pseudo-response may include instantaneous false vibration peaks generated by short-term water flow impact, and environmental clutter may include sudden feature peaks caused by workshop equipment start-up and shutdown or electromagnetic interference. It can be understood that all high-weight false feature regions not caused by real equipment degradation can be accurately located. Step 502: Construct the mechanical feature neighborhood topology; specifically, with each pixel node of the mechanical degradation feature matrix as the center, establish an eight-neighbor topology adjacency graph, quantify the spatial relationship between the central node and the surrounding neighboring nodes; calculate the local surface curvature through the neighborhood gradient difference. The larger the curvature value, the more severe the feature abrupt change in the local area and the higher the probability of noise interference; the smaller the curvature, the more stable the feature distribution and the higher the probability of true degradation features; for example, noise interference points will show single-point feature abrupt changes and a sudden increase in neighborhood curvature values, while the true feature of uniform bearing wear will show a gradual and stable change in neighborhood curvature; it can be understood that the difference in curvature can accurately distinguish between noise abrupt changes and true degradation gradual features. Step 503: Construct an adaptive smoothing Laplacian operator. Specifically, for the selected high-weight interference feature elements, traverse their eight neighboring nodes and select stable and effective neighboring nodes whose local curvature is not higher than a preset smoothing threshold. Based on the spatial distance weight and feature amplitude of the effective neighboring nodes, adaptively construct a Laplacian smoothing operator that fits the current local surface shape. The operator weights are dynamically updated with the local feature stability, avoiding the limitations of global uniform filtering. For example, if the interference node is surrounded by stable wear features, a high smoothing weight operator can be constructed to complete accurate correction. If there are real weak degradation mutation features around the interference node, the smoothing weights can be reduced while retaining effective mutation information. This can be understood as the operator parameters being adaptable to all local feature distribution scenarios. Step 504: Manifold Interpolation Repair of Feature Values; Specifically, using the constructed adaptive Laplacian operator, manifold interpolation smoothing is performed on the high-abrupt-interference feature nodes; based on the effective neighborhood stationary features, the true feature amplitude of the center node is fitted with spatial weights, noise abrupt peaks are removed, and repair feature values ​​that conform to the continuous degradation law of the equipment are fitted; for example, the instantaneous ultra-high vibration peaks caused by electromagnetic interference are removed, and feature values ​​that are continuous and conform to the true degradation trend of the equipment are fitted based on the neighborhood stationary wear features; it can be understood that all discrete abrupt noise points can be accurately repaired through this step; Step 505: Update and generate the purified degradation feature matrix. Specifically, using a point-by-point precise replacement method, the original interference noise values ​​in the mechanical degradation feature matrix are replaced with fitted repair feature values, fully preserving the original effective degradation features in the low-noise region. Finally, the entire matrix is ​​adaptively denoised to generate a purified degradation feature matrix with high signal-to-noise ratio and high fidelity. For example, only the noise mutation points identified in the matrix are replaced, fully preserving the continuous degradation features of bearing wear and impeller fatigue, avoiding excessive smoothing of effective features. This step can be understood as achieving precise correction of noise points and full preservation of effective features. Of course, this is not limited to the above-mentioned degradation features. The above technical solution abandons the coarse noise reduction method of traditional global unified filtering and achieves pixel-level adaptive local noise reduction based on manifold curvature and neighborhood topology. While filtering out hydraulic pseudo-features and industrial clutter noise, it fully preserves the early weak degradation mutation features and avoids the effective defect features being over-smoothed, thus solving the problem of low fidelity of traditional noise reduction algorithms.

[0024] In some embodiments, the AR visualization rendering process for maintaining guidance signaling specifically includes the following steps: Step 601: Receive terminal coordinate request instruction; Specifically, the system background continuously listens for communication requests from external AR augmented reality terminals. When maintenance personnel hold an AR terminal and approach the pump set equipment and trigger the maintenance viewing function, the system receives the equipment identification code and coordinate request instruction uploaded by the terminal in real time, and starts the AR visualization analysis process; For example, when maintenance personnel hold AR glasses and approach monitoring equipment such as chemical centrifugal pumps and municipal water supply pump sets, and click the terminal maintenance query button, the system can receive the instruction and respond; It can be understood that all AR terminals connected to the system and all monitoring pump set equipment can adapt to this interaction process; Step 602: Maintenance Signal 3D Spatial Parsing; Specifically, the system performs structured data parsing on the maintenance guidance signal carrying the location coordinates of the faulty component, component identification, and fault range, converting the two-dimensional plane point coordinates into a three-dimensional spatial mapping matrix adapted for real-time AR rendering. The matrix parameters accurately include the three-dimensional spatial coordinates, size boundaries, and rotational orientation of the faulty component, achieving precise mapping of the fault location from the data plane to the physical three-dimensional space. For example, the two-dimensional fault coordinates of the impeller stored in the background are parsed into a three-dimensional spatial matrix of the impeller's length, width, height, installation angle, and spatial position adapted for AR rendering. This can be understood as the planar coordinates of all key components of the pump unit being able to complete the three-dimensional spatial mapping transformation. Step 603: AR terminal 3D bounding box rendering; Specifically, the parsed 3D spatial mapping matrix is ​​sent to the external augmented reality terminal. The terminal, combined with its built-in device 3D model library, accurately overlays and renders the 3D geometric model bounding box corresponding to the faulty component in the real-time captured device scene. By highlighting the selection and using differentiated color markings, the faulty component is intuitively distinguished from the normal component, achieving millimeter-level visual positioning of the fault location. For example, a red highlighted bounding box marks a bearing with severe wear, a yellow highlighted bounding box marks a slightly cavitated impeller, and a green bounding box marks a normal and intact component. This can be understood as a bounding box rendering effect that can be adapted to different colors and transparency according to the fault level. The above technical solution realizes the mapping of fault location from implicit data coordinates in the background to real-world 3D visualization in the front end, which changes the problem of traditional text alarms and data alarms being unintuitive, provides on-site maintenance personnel with fault location guidance with zero troubleshooting cost, and reduces the difficulty of manual troubleshooting.

[0025] In some embodiments, the adaptive threshold warning process of the signaling generation module specifically includes the following steps. This embodiment fully discloses the calculation logic, correlation, and engineering examples of the threshold compensation amount and the adaptive adjustment amount: Step 701: Real-time acquisition of equipment operating parameters; Specifically, the system acquires core operating parameters of the pump set's main shaft speed and outlet flow rate in real time at millisecond levels through the BMC acquisition unit, high-precision speed sensor, and flow sensor, accurately characterizing the current load intensity and operating range of the equipment; For example, it can acquire full-load parameters of rated operating conditions of 2900 r / min and 100 m³ / h, and it can also acquire low-load variable operating parameters of 1500 r / min and 40 m³ / h; It can be understood that the pump set's full-range speed and flow operating parameters can be acquired in real time for load determination. Of course, this is not limited to acquiring only two types of operating parameters, and parameters such as pressure and head can be added to assist in the determination. Step 702: Match and adaptively correct the alarm threshold for operating conditions; Specifically, the system pre-constructs an operating condition-threshold mapping surface trained with massive historical data of all operating conditions. The surface stores the normal baseline fluctuation range of the pump group health index and the basic alarm threshold Base_Th under different speed and flow combination operating conditions. To solve the problems of large fluctuations and easy false alarms under high load conditions and insufficient sensitivity and easy missed alarms under low load conditions, this system sets up a dynamic compensation mechanism for operating conditions. The threshold compensation amount is positively correlated with the load intensity. The higher the load, the greater the compensation amount and the higher the adaptive upward adjustment of the threshold. Specifically: First, calculate the operating condition load factor K, K = (real-time speed / rated speed + real-time flow / rated flow) / 2, K ranges from 0 to 1, used to quantify the real-time load level of the equipment; then calculate the dynamic compensation amount ΔTh = Base_Th × K × preset compensation factor; finally, the adaptive alarm threshold Adap_Th = Base_Th + ΔTh; where: ΔTh is the threshold dynamic compensation amount, which is also equal to the threshold adaptive upward adjustment amount, the two values ​​are completely equivalent, the load factor K directly determines the upward adjustment amount, realizing the dynamic correction of the operating condition adaptive threshold; Quantitative engineering example: Set the pump set rated speed to 2900 r / min, rated flow rate to 100 m³ / h, base alarm threshold Base_Th=0.6, and fixed compensation coefficient=0.08; High load condition: real-time speed 2800r / min, flow rate 95m³ / h, calculated K≈0.96, compensation amount ΔTh≈0.6×0.96×0.08≈0.046, adaptive threshold Adap_Th=0.646, the threshold is adaptively adjusted upward with high load to avoid false alarms caused by operating condition fluctuations; Low load steady-state conditions: real-time speed 1500 r / min, flow rate 40 m³ / h, calculated K≈0.47, compensation ΔTh≈0.022, adaptive threshold Adap_Th=0.622, threshold slightly increased to ensure detection sensitivity; Step 703: Calculate the health status degradation slope; Specifically, the system continuously collects the pump group health index for 10 fixed sampling periods, calculates the deviation between the health index and the current adaptive alarm threshold for each period, and fits the time-series deviation curve with the time sequence as the horizontal axis and the deviation value as the vertical axis using the least squares method to solve for the curve slope; a continuously negative slope indicates that the health index is continuously below the benchmark threshold and the equipment is in a state of continuous degradation; the larger the absolute value of the slope, the faster the degradation rate; for example, a deviation slope of -0.02 / cycle for 10 consecutive periods indicates slow and gradual degradation of the equipment; a slope of -0.08 / cycle indicates rapid deterioration of the equipment and an extremely high risk of failure; it can be understood that the slope value can accurately quantify the degradation rate of the equipment; Step 704: Hysteresis Window Determination and Maintenance Signal Generation; Specifically, the system presets an adjustable hysteresis time window of 3s to 10s. When the deviation slope is continuously negative and the duration is greater than or equal to the hysteresis time window, it excludes temporary index fluctuations caused by instantaneous operating condition fluctuations and data disturbances, accurately determines that the pump unit has entered the continuous fault evolution stage, and then generates a refined maintenance guidance signal carrying the three-dimensional position coordinates of the target component and the fault level; For example, a 5s hysteresis window is preset. If the degradation slope is continuously negative and the duration reaches 5s, it is determined to be a real degradation and an operation and maintenance alarm is triggered; if the fluctuation lasts only 1s, it is determined to be an instantaneous interference and no alarm is triggered; It can be understood that the hysteresis window duration can be adaptively adjusted according to the on-site monitoring accuracy requirements; Through the above technical solutions, the system effectively distinguishes between normal operating condition fluctuations and real gradual degradation by using a triple judgment mechanism of adaptive threshold dynamic compensation adjustment, least squares slope fitting, and time-domain hysteresis. This solves the problems of frequent false alarms and missed alarms caused by traditional fixed thresholds, and improves the operating condition adaptability and operational stability of the early warning system.

[0026] In some embodiments, the data preprocessing and matrix construction process of the runtime parameter acquisition module specifically includes the following steps: Step 801: Synchronously acquire analog sensor signals; Specifically, the system sets a unified high-frequency sampling frequency, and uses pressure sensors installed at the inlet and outlet of the pump set and vibration sensors on the bearing base to synchronously acquire the fluid end pressure analog signal and the mechanical end vibration analog signal using the same sampling clock, ensuring that the timing start and sampling interval of the two types of signals are completely aligned and there is no phase deviation; For example, pressure sensors are arranged at the low-pressure end of the centrifugal pump inlet and the high-pressure end of the outlet, and vibration sensors are arranged at the front and rear bearing bases and the pump frame to synchronously acquire analog voltage signals; It can be understood that all sensor analog signals that can characterize the fluid working condition and mechanical operating state of the pump set can be synchronously acquired, and of course, the sensor installation location and type are not limited here; Step 802: Analog-to-digital conversion to obtain the original digital sequence; specifically, a high-precision AD conversion module is used to perform discrete equal-interval sampling and quantization on the continuous analog signal, converting the continuous voltage analog quantity into a discrete digital quantity, and obtaining the pressure original digital signal sequence and vibration original digital signal sequence with strict time alignment respectively; for example, the pressure and vibration continuous analog voltage signals from 0 to 5V are converted into discrete digital quantized values ​​in the range of 0 to 4095, forming a computable time-series digital sequence; it can be understood that all standard sensor analog voltage signals can be digitized and analyzed through AD conversion; Step 803: Outlier removal and processing; Specifically, preset the upper and lower limits of the sensor's standard range, traverse the original digital signal sequence, and identify abrupt changes in data exceeding the upper and lower limits as invalid outliers caused by electromagnetic interference or sensor jitter; use the neighborhood mean replacement method to repair the outliers: take the values ​​of the three valid sampling points before and after the outlier point, calculate the arithmetic mean, and replace the outlier point with the mean center value to completely eliminate single-point abnormal noise; for example, for pressure sensors with a standard range of 0 to 1.6 MPa, instantaneous abrupt changes exceeding 1.6 MPa are identified as outliers and replaced with the mean of valid neighboring data; for vibration sensors with a range of 0 to 20 g, instantaneous changes exceeding the range are uniformly repaired; it can be understood that all single-point invalid data exceeding the range or with abnormal changes can be repaired, but this is not limited to the above two types of sensor ranges; Step 804: Wavelet Thresholding Refined Denoising; Specifically, wavelet soft thresholding refined denoising is performed on the regular signal sequence after outlier removal. High-frequency random Gaussian noise components are stripped away through multi-layer wavelet decomposition, while retaining effective fault feature components such as pressure transient fluctuations and vibration impact pulses. This purifies the original signal without damaging weak defect features, resulting in highly pure pressure and vibration signals. For example, high-frequency Gaussian noise generated by circuit thermal noise and random environmental vibrations is stripped away, while pressure pulses caused by cavitation and vibration impact features caused by bearing wear are fully preserved. This step can be understood as accurately distinguishing random noise from effective fault features. Step 805: Time-series alignment to construct the operating parameter matrix; Specifically, using a unified global timestamp as a benchmark, the denoised pressure time-series data and vibration time-series data are aligned moment by moment and concatenated column by column to construct a standardized two-dimensional pump group operating parameter matrix in which rows correspond to sampling times and columns correspond to sensing channels. The data is well-organized and the time series is unbiased, providing reliable data support for backend algorithm calculations. For example, using a 1ms global timestamp as a benchmark, the pressure and vibration data at the same moment are concatenated one by one to form a time-synchronized and dimension-regular N×2 parameter matrix. This can be understood as all preprocessed dual-source time-series data being able to construct a standardized matrix according to this rule. The above technical solution completes the standardized preprocessing of the entire process from raw analog signal acquisition to structured parameter matrix, eliminating sensor anomalies, random noise, and timing misalignment issues layer by layer, ensuring the consistency and reliability of input feature data.

[0027] In some embodiments, the health state evolution model uses a graph convolutional network structure for feature extraction, and the specific implementation steps are as follows: Step 901: Feature Matrix Graph Structure Topology Mapping; Specifically, the input purified degradation feature matrix is ​​transformed into an undirected graph structure data. Each feature element at a spatial location in the matrix is ​​defined as an independent graph node, with the node's feature value being the corresponding device degradation feature response value. A node adjacency matrix is ​​constructed based on the row and column spatial adjacency relationships of the matrix, establishing spatial topological associations between all nodes to achieve accurate conversion from gridded features to topology graph data. For example, grid feature nodes corresponding to bearings, impellers, and pipelines are topologically connected according to spatial adjacency relationships to achieve the association and binding of degradation features of adjacent components. This can be understood as topology mapping being completed for all spatially distributed component features of the pump group. Step 902: Graph Convolution Neighborhood Feature Aggregation and Update; Specifically, a preset graph convolution kernel is invoked, all graph nodes are traversed, and with a single node as the center, the features of surrounding associated nodes are weighted and aggregated according to the spatial weights of the adjacency matrix, fusing multi-region spatial degradation association information; the aggregated neighborhood features are nonlinearly fused and iteratively updated with the original features of the center node, weakening the weight of isolated noise nodes, strengthening the feature association of continuous degradation regions, and accurately mining the coupling degradation law of multiple components; for example, the impeller cavitation degradation features are fused with the vibration degradation features of adjacent bearings to mine the coupling degradation law of cavitation impact aggravating bearing wear; the pipeline resonance and pump body vibration features are fused to identify the linkage degradation trend; it can be understood that all multi-component spatial coupling degradation association laws can be mined through this step; Step 903: Output the deep hidden layer state vector; Specifically, after multi-layer graph convolutional iterative aggregation and topological correlation mining, the update features of all graph nodes are globally fused and normalized, and the dimensions are unified to output a deep hidden layer state vector that can characterize the spatiotemporal correlation law of global equipment degradation and the coupling characteristics of multiple components, which is used for subsequent nonlinear mapping calculation of health index; For example, the final output hidden vector can simultaneously carry multi-dimensional hidden information such as component degradation location, degradation degree, temporal evolution rate, and spatial coupling correlation; It can be understood that this vector completely covers the global degradation feature information of the pump group, but it is not limited to including only the above feature dimensions here; By utilizing the advantages of graph convolutional networks in naturally adapting to gridded and spatially distributed features, the limitations of traditional fixed convolutional kernels are overcome, the spatial correlation characteristics of degradation of multiple pump components are explored, and the characterization ability and evaluation accuracy of hidden degradation features of equipment under complex working conditions are improved.

[0028] In some embodiments, the AR terminal adaptive information rendering process specifically includes the following steps: Step 1001: Respond to terminal interaction operations; specifically, when maintenance personnel click on the fault bounding box area on the AR terminal interface, the system responds to the touch interaction operation in real time, triggering the maintenance file parsing, spare parts matching, and adaptive rendering process; for example, when maintenance personnel click on the bearing fault bounding box marked in the AR interface, the system immediately retrieves the maintenance data and suitable spare parts for that bearing; clicking on the impeller fault bounding box displays the corresponding impeller maintenance-related information; it can be understood that the AR interaction area of ​​all faulty components can respond accurately; Step 1002: Parse maintenance records and spare parts information; specifically, the system searches the background maintenance database based on the three-dimensional location coordinates and unique part number of the faulty component, accurately retrieving the historical maintenance records, fault statistics, standardized repair processes, and compatible replacement spare parts list and parameters corresponding to that component; for example, it retrieves the wear failure records, disassembly and repair process standards, and compatible bearing models and size parameters of the target bearing for the past 3 years; it retrieves the cavitation repair process and replacement impeller specifications; this can be understood as the accurate retrieval of maintenance records and spare parts data for all standardized components of the pump set; Step 1003: Obtain real-time terminal posture parameters; specifically, collect the AR terminal's camera pose matrix, world space 3D coordinates, field of view, and depth parameters in real time to accurately lock the terminal's real-time shooting angle, spatial position, and visible area; for example, when maintenance personnel move or rotate the AR terminal, the system updates posture parameters such as camera pitch angle, horizontal rotation angle, shooting distance, and visible range in real time; this can be understood as all dynamic posture changes of the terminal being collected, adapted, and rendered in real time. Step 1004: Construct 3D spatial information anchor points. Specifically, the retrieved maintenance text, process descriptions, spare parts parameters, and image data are converted from 2D data to 3D spatial coordinates based on the camera's real-time pose matrix. 3D information anchor points are then constructed and bound to the visible area surrounding the faulty component model. For example, the bearing maintenance procedure text and spare parts specification images are bound to the visible area to the side of the bearing model; the impeller cavitation repair process is bound to the visible position above the impeller. This can be understood as all maintenance text and image information being able to construct 3D anchor points and adaptively bind them; however, the anchor point binding position is not limited here. Step 1005: Spatial Occlusion and Perspective Adaptive Rendering; Specifically, the spatial occlusion relationship and depth distance between the 3D information anchor points, the fault bounding box, and the device body model are calculated in real time. Combined with the terminal's field of view and perspective rules, the display level, transparency, scaling ratio, and display position of the information anchor points are automatically and dynamically adjusted. Unoccluded areas are displayed normally with high brightness, partially occluded areas are displayed with low brightness, and completely occluded areas are hidden, achieving an adaptive dynamic rendering effect that conforms to human visual habits. For example, when the terminal is facing the faulty component without occlusion, the information anchor points are fully highlighted; when the device pipeline is slightly occluded, the information is displayed semi-transparently with low brightness; when the component is completely occluded, it is automatically hidden to avoid overlapping and confusion in the image. This can be understood as dynamic rendering that can adapt to all shooting angles and occlusion scenes.

[0029] The above technical solution enables integrated AR visualization display of fault location, historical maintenance traceability, spare parts recommendation, and process guidance, solving the problem of cumbersome traditional maintenance data query.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A pump set health monitoring and predictive maintenance system, characterized in that, include: The operating parameter acquisition module acquires the pump set operating parameter matrix, which includes the fluid end pressure time series and the mechanical end vibration time series. The feature acquisition module acquires the frequency domain energy distribution features of the fluid end pressure time series and generates a hydraulic transient tensor; it also acquires the time domain envelope features of the mechanical end vibration time series and generates a mechanical degradation feature matrix. The sparsification module performs sparsification on the hydraulic transient tensor, setting elements with energy variance no higher than the preset noise floor to zero in order to reduce the data size of subsequent processing. The coupling weight generation module constructs a time-delay cross-correlation analysis window, which slides on the time axis of the sparsed hydraulic transient tensor to determine the cross-correlation sequence between the hydraulic transient tensor and the mechanical degradation feature matrix under different time lag factors. The time lag factor corresponding to the peak with the largest absolute value in the cross-correlation coefficient sequence is selected as the optimal fluid-structure interaction time lag. Based on the optimal fluid-structure interaction time lag, the hydraulic transient tensor is shifted along the time axis to complete the calibration. The calibrated hydraulic transient tensor is then input into a nonlinear activation function for mapping to obtain the hydraulic modulation intensity tensor. The hydraulic modulation intensity tensor and the mechanical degradation feature matrix are subjected to Hadamard fusion processing to generate a coupling weight masking matrix. The masking and denoising module performs masking and denoising on the mechanically degraded feature matrix based on the coupled weight masking matrix to generate a purified degraded feature matrix; The health index output module inputs the purified degradation feature matrix into the pre-trained health state evolution model and outputs the corresponding pump group health index. The maintenance signal generation module generates a maintenance guidance signal with the location coordinates of the target component when the pump group health index is less than or equal to the preset alarm threshold.

2. The pump set health monitoring and predictive maintenance system according to claim 1, characterized in that, The feature acquisition module acquires the frequency domain energy distribution features of the fluid end pressure time series and generates a hydraulic transient tensor; Obtain the temporal envelope features of the mechanical end vibration time series to generate a mechanical degradation feature matrix, including: Perform discrete wavelet transform on the fluid end pressure time series to obtain multiple pressure subsequences in different frequency bands; determine the local energy variance of each pressure subsequence; concatenate the local energy variances into a one-dimensional feature vector according to the preset frequency from low to high, and then convert the one-dimensional feature vector into a two-dimensional hydraulic transient tensor. Obtain the root mean square value and peak factor of the mechanical end vibration time series within a preset sliding time window; map the root mean square value and peak factor to a preset geometric topology grid to generate a mechanical degradation feature matrix with spatial distribution attributes.

3. The pump set health monitoring and predictive maintenance system according to claim 1, characterized in that, The health index output module inputs the purified degradation feature matrix into the pre-trained health state evolution model and outputs the corresponding pump group health index, including: The purified degradation feature matrix is ​​input into the feature extraction layer of the healthy state evolution model to obtain the hidden layer state vector; The hidden layer state vector is input to the fully connected layer for nonlinear mapping, and the output value reflecting the degree of equipment degradation is used as the pump group health index.

4. The pump set health monitoring and predictive maintenance system according to claim 1, characterized in that, Before obtaining the pump set operating parameter matrix, a pre-training process for the health state evolution model is also included, including: Obtain the historical sample matrix; the historical sample matrix contains historical hydraulic transient tensors and historical mechanical degradation feature matrices labeled with actual health status. The historical hydraulic transient tensor and the historical mechanical degradation feature matrix are input into the initial network model for forward propagation, and the predicted health status index is output. Determine the basic mean square error between the predicted health status index and the actual health status label; obtain the high-frequency perturbation coefficients in the historical hydraulic transient tensor that indicate fluid cavitation; fuse the high-frequency perturbation coefficients with the basic mean square error to generate a penalty loss function; The gradient value is determined based on the penalty loss function. The weight parameters of the initial network model are updated according to the gradient descent direction. When the penalty loss function converges to the preset interval, the parameter update is stopped and the model parameters are saved to obtain the health state evolution model.

5. The pump set health monitoring and predictive maintenance system according to claim 1, characterized in that, The masking and denoising module performs masking and denoising on the mechanically degraded feature matrix based on the coupled weight masking matrix, generating a purified degraded feature matrix, including: Obtain the feature elements in the coupling weight masking matrix whose values ​​are not lower than a preset masking threshold; Construct an adjacency graph of the mechanical degradation feature matrix and determine the local curvature of each feature element and its neighboring feature elements; For feature elements whose values ​​are not lower than a preset masking threshold, a Laplacian operator is constructed using feature elements whose local curvature in their neighborhood is not higher than a preset smoothing threshold. Based on the Laplacian operator, manifold interpolation is performed on feature elements whose values ​​are not lower than a preset masking threshold to generate repair feature values. The purified degradation feature matrix is ​​obtained by replacing the corresponding element values ​​in the mechanical degradation feature matrix with the repair feature values.

6. The pump set health monitoring and predictive maintenance system according to claim 1, characterized in that, After generating the maintenance guidance signaling with the target component's location coordinates, the following is also included: Obtain coordinate request commands sent by external augmented reality terminals; The maintenance guidance signaling will be parsed into a three-dimensional spatial mapping matrix; The three-dimensional spatial mapping matrix is ​​sent to an external augmented reality terminal, and the bounding box of the three-dimensional geometric model of the faulty component is overlaid and rendered on the display interface of the external augmented reality terminal.

7. The pump set health monitoring and predictive maintenance system according to claim 1, characterized in that, When the pump group's health index is less than or equal to a preset alarm threshold, the maintenance signal generation module generates a maintenance guidance signal with the location coordinates of the target component, including: Real-time acquisition of the pump set's current operating parameters, which include at least the spindle speed and outlet flow rate; The operating condition parameters are input into the pre-constructed operating condition-threshold mapping surface to determine the current adaptive alarm threshold; the operating condition-threshold mapping surface records the baseline fluctuation range of the pump group health index under different operating conditions. Determine the slope of the deviation between the pump unit health index and the adaptive alarm threshold; When the deviation slope remains negative for a duration not less than the preset hysteresis time window, the pump unit is determined to have entered the fault evolution stage, and a maintenance guidance signal with the location coordinates of the target component is generated.

8. The pump set health monitoring and predictive maintenance system according to claim 1, characterized in that, The operating parameter acquisition module acquires the pump set operating parameter matrix, including: Simultaneously acquire simulated pressure signals at the fluid end and simulated vibration signals at the mechanical end of the pump unit using a preset sampling frequency; The pressure analog signal and the vibration analog signal are converted from analog to digital respectively to obtain the original digital signal sequence; Outlier removal is performed on the original digital signal sequence, replacing outliers whose values ​​exceed the preset sensor range with the center value of the signal at adjacent time points. The signal after removing outliers is subjected to wavelet threshold denoising, and the denoised pressure signal and vibration signal are aligned and spliced ​​according to the timestamp to construct the pump group operation parameter matrix.

9. A pump set health monitoring and predictive maintenance system according to claim 3, characterized in that, The feature extraction layer of the health state evolution model adopts a graph convolutional network structure; The purified degradation feature matrix is ​​input into the feature extraction layer of the healthy state evolution model, including: The purified degenerate feature matrix is ​​treated as graph-structured data, where each element of the matrix is ​​a graph node. The features of neighboring nodes are aggregated using graph convolution kernels, the feature representation of the current node is updated, and the hidden layer state vector is obtained.

10. A pump set health monitoring and predictive maintenance system according to claim 6, characterized in that, After overlaying and rendering the bounding box of the 3D geometric model of the faulty component onto the display interface of the external augmented reality terminal, it also includes: In response to user interaction on an external augmented reality terminal, the historical maintenance records and recommended spare parts list corresponding to the location coordinates of the target component are analyzed. Obtain the current camera pose parameters and field of view of the external augmented reality terminal; Based on camera pose parameters, the text and graphic data of historical maintenance records and recommended spare parts lists are converted into three-dimensional spatial anchor points, and the spatial occlusion relationship between the three-dimensional spatial anchor points and the bounding box of the three-dimensional geometric model of the faulty component is determined. Based on spatial occlusion and field of view, the historical maintenance records and recommended spare parts list are adaptively rendered in the display interface of the external augmented reality terminal in a manner that conforms to the laws of real visual perspective.

Images