A pump fault identification method and system for a pump house multi-sound source environment

Abstract

The present application relates to the technical field of equipment fault diagnosis and acoustic signal processing, and particularly relates to a pump fault identification method and system for a pump house multi-sound source environment, which comprises: collecting multi-channel sound signals, mechanical vibration signals and working condition parameters in the pump house and time synchronizing; according to the mechanical vibration signal quality evaluation result, adaptively selecting a semi-blind source extraction or FastICA blind separation strategy to process the multi-channel sound signals to obtain candidate sound sources; and performing multi-modal confidence fusion on the candidate sound sources to automatically identify target pump operation sound signals; performing abnormality detection based on a working condition grading dynamic baseline model; and using a lightweight 1D-CNN model quantized by INT8 to perform fault type classification. The present application forms a closed-loop processing flow from sound source separation, target identification, abnormality detection to fault classification, solves the problems of difficulty in separation, unclear composition correspondence, poor environmental adaptability and edge real-time performance under multi-sound source aliasing, realizes real-time and accurate identification, and has a continuous learning ability.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis and acoustic signal processing technology, specifically to a method and system for identifying water pump faults in a multi-sound-source environment of a pumping station. Background Technology

[0002] Water pumps are indispensable core power equipment in industrial production and urban water supply systems, and their stable and reliable operation is directly related to the continuity of production and the quality of life of residents. With the advancement of industrial and smart city construction, equipment condition monitoring and fault early warning technologies have developed rapidly. Non-contact fault diagnosis technology based on acoustic signals, with its unique advantages such as not requiring equipment disassembly, not affecting normal production, and being able to capture early and weak fault characteristics, has become an important research direction in the field of water pump condition monitoring.

[0003] Currently, the industry has developed a variety of water pump fault diagnosis solutions based on acoustic analysis, involving multiple technical aspects such as sound signal acquisition, preprocessing, feature extraction, and pattern recognition, providing technical support for the preventive maintenance of water pumps.

[0004] However, despite the progress made by existing solutions in the field of acoustic fault diagnosis of water pumps, they still face the following significant drawbacks in practical applications where multiple water pumps are centrally deployed in a pump house: 1) The sound signals generated by multiple water pumps operating simultaneously in the pump room are superimposed on each other. At the same time, there are various environmental interferences such as pipeline vibration, valve operation, and electrical equipment operation. Existing solutions are difficult to accurately separate the independent operating sound of a single target water pump from the complex mixed sound signals. 2) Most existing solutions treat sound source separation and fault identification as two separate technical steps. The multiple signal components obtained by separation cannot be automatically correlated with the physical water pump, requiring manual intervention for judgment, and thus cannot achieve fully automated diagnosis. 3) The acoustic environment inside the pump room will change dynamically with the combination of pump start-up and shutdown and the change of operating conditions. Existing solutions generally use fixed thresholds or static baselines for anomaly detection, which is prone to generating a large number of false alarms or missed alarms due to environmental fluctuations. 4) High-precision fault classification algorithms are usually computationally complex and difficult to implement in real time on edge computing devices with limited resources; while cloud-based solutions suffer from network latency and data transmission security issues, failing to meet the requirements of low-latency monitoring in industrial settings. 5) Most existing blind source separation technologies adopt unsupervised pure blind separation methods, which do not effectively combine the prior information that is easily obtained on site, such as pump mechanical vibration signals and operating condition parameters, to constrain the separation process, resulting in insufficient accuracy and robustness of the separation results.

[0005] The aforementioned problems collectively result in poor performance of existing acoustic fault diagnosis technologies for water pumps in complex environments with multiple sound sources in pump rooms, making it difficult to simultaneously meet the requirements of industrial sites for diagnostic accuracy, real-time performance, and automation. Summary of the Invention

[0006] To address the above technical problems, this invention provides a technical solution for a method and system for identifying pump faults in multi-sound-source environments in pump rooms.

[0007] The technical problem solved by this invention can be achieved by the following technical solution: a method for identifying pump faults in a multi-source environment of a pumping station, comprising: step S1, collecting multimodal data in the pumping station and performing time synchronization processing on the multimodal data; wherein, the multimodal data includes multi-channel sound signals in the pumping station, mechanical vibration signals of each pump, and real-time operating parameters of each pump; step S2, processing the multi-channel sound signals using a hybrid sound source adaptive separation strategy based on the quality assessment results of the mechanical vibration signals to obtain at least one candidate sound source signal; step S3, performing multimodal confidence fusion verification on each candidate sound source signal to identify the target pump operating sound signal from the multiple candidate sound source signals; step S4, performing anomaly detection processing on the target pump operating sound signal to obtain a fault detection result; step S5, classifying the target pump operating sound signal by fault type based on the fault detection result to obtain a fault diagnosis result.

[0008] Preferably, step S2 includes: step S21, performing frame-by-frame processing on the multi-channel audio signal using preset parameters, and performing adaptive wavelet threshold denoising processing on each frame of the multi-channel audio signal to obtain multiple denoised time-domain signal frames; step S22, calculating the average frequency domain coherence coefficient of the mechanical vibration signal and the multi-channel audio signal in the target frequency band, and judging the quality of the mechanical vibration signal based on the average frequency domain coherence coefficient, wherein, if the average frequency domain coherence coefficient is greater than a first threshold, the mechanical vibration signal is judged to be of good quality; if the average frequency domain coherence coefficient is not greater than the first threshold and not less than a second threshold, the mechanical vibration signal is judged to be of average quality; if the average frequency domain coherence coefficient is not greater than the first threshold and not less than a second threshold, the mechanical vibration signal is judged to be of average quality; if the average frequency domain coherence coefficient is less ... If the average coherence coefficient of the domain is less than the second threshold, the mechanical vibration signal is determined to be of poor quality. In step S23, based on the quality determination result of the mechanical vibration signal, an adaptive separation algorithm is selected to perform sound source separation on the denoised time-domain signal frame. If the mechanical vibration signal quality is good, a semi-blind source extraction algorithm is selected to perform sound source separation on the denoised time-domain signal frame, extracting sound source components statistically related to the mechanical vibration signal, and using the sound source components as candidate sound source signals. If the mechanical vibration signal quality is average or poor, a FastICA blind separation algorithm is selected to perform sound source separation on the denoised time-domain signal frame, separating multiple independent components, and using the multiple independent components as candidate sound source signals.

[0009] Preferably, in step S21, adaptive wavelet thresholding denoising is performed on each frame of multi-channel audio signal, including: step S211, using wavelet decomposition to perform multi-level decomposition on each frame of multi-channel audio signal to obtain the approximation coefficients and detail coefficients of each layer; step S212, estimating the noise standard deviation based on the median absolute deviation of the detail coefficients of the first layer to obtain the noise standard deviation estimate; step S213, using the Birgé-Massart strategy to calculate the adaptive threshold corresponding to the detail coefficients of each layer; step S214, using a soft thresholding function to perform thresholding on the detail coefficients of each layer; and step S215, using inverse discrete wavelet transform to reconstruct the processed detail coefficients and the original approximation coefficients of each layer to obtain the multiple denoised time-domain signal frames.

[0010] Preferably, step S22 includes: step S221, aligning the mechanical vibration signal and the multi-channel sound signal with timestamps, performing frame segmentation processing according to the same frame length and frame shift, and windowing processing for each frame signal; step S222, calculating the auto-power spectral density of the multi-channel sound signal, the auto-power spectral density of the mechanical vibration signal, and the cross-power spectral density between the multi-channel sound signal and the mechanical vibration signal, specifically using the following formula: , , , in, The power spectral density of the multi-channel audio signal; The power spectral density of the mechanical vibration signal is given. The cross-power spectral density between the multi-channel sound signal and the mechanical vibration signal; The Fourier transform of the k-th frame of the multi-channel audio signal; The Fourier transform of the k-th frame of the mechanical vibration signal. for The complex conjugate of ; K is the total number of frames; f is the frequency; Step S223: Calculate the frequency domain coherence coefficient at each frequency point within the target frequency band based on the self-power spectral density and the cross-power spectral density. The specific calculation formula is as follows: , in, It is a frequency domain coherence function; Step S224: Calculate the average frequency domain coherence coefficient of all frequency points within the target frequency band. The specific calculation formula is as follows: , in, This represents the average frequency domain coherence coefficient within the target frequency band. This is the set of effective frequency points after filtering.

[0011] Preferably, in step S23, the specific steps for separating the sound sources of the denoised time-domain signal frame using the semi-blind source extraction algorithm are as follows: Constructing the covariance matrix of the denoised time-domain signal frame and the cross-covariance vector between the denoised time-domain signal frame and the mechanical vibration signal; constructing a generalized eigenvalue equation, solving for the eigenvector corresponding to the largest eigenvalue of the generalized eigenvalue equation, and obtaining the optimal unmixing vector; multiplying the optimal unmixing vector with the denoised time-domain signal frame to obtain the candidate sound source signal; The specific steps for separating the sound sources of the denoised time-domain signal frame using the FastICA blind separation algorithm are as follows: The denoised time-domain signal frame is preprocessed with whitening to obtain a whitened signal matrix, the specific formula of which is: , in, This is the signal matrix after whitening; This is the whitening matrix; This is the time-domain signal matrix after noise reduction; The unmixing matrix is ​​solved using a fixed-point iterative algorithm based on maximizing negative entropy. The specific solution formula is as follows: , , , , in, This is the unmixed vector after iterative updates; The unmixing vector; It is a nonlinear function; After iterative convergence, the unmixed matrix is ​​multiplied by the whitened signal matrix to obtain the candidate sound source signal.

[0012] Preferably, step S3 includes: step S31, performing physical causality verification, operating mode verification, and spatial consistency verification on each candidate sound source signal to obtain the coherence score, operating condition matching score, and spatial consistency score corresponding to each candidate sound source signal; step S32, weightedly fusing the coherence score, operating condition matching score, and spatial consistency score of each candidate sound source signal to obtain the comprehensive confidence level corresponding to each candidate sound source signal, specifically calculated using the following formula: , in, The overall confidence level; The score is for coherence. Score for matching operating conditions; Spatial consistency score; , , All are weighting coefficients, which are adaptively adjusted according to the quality changes of the mechanical vibration signal; in step S33, the comprehensive confidence of each candidate sound source signal is compared with a preset threshold. If the comprehensive confidence is not less than the preset threshold, the candidate sound source signal is determined to be the target water pump operating sound signal; otherwise, it is discarded.

[0013] Preferably, step S4 includes: step S41, extracting features from the target water pump's operating sound signal to obtain an acoustic feature vector, the acoustic feature vector including Mel frequency cepstral coefficient features, wavelet packet energy features, and spectral features; step S42, during the normal operation phase of the water pump, dividing it into multiple operating condition intervals according to the real-time operating parameters of the target water pump, collecting normal operation sound data for each operating condition interval, and constructing a dynamic baseline model corresponding to each operating condition interval; step S43, during the online monitoring phase, selecting the corresponding optimal dynamic baseline model according to the current real-time operating parameters, and inputting the acoustic feature vector into the optimal dynamic baseline model for anomaly detection processing to obtain the fault detection result.

[0014] Preferably, in step S5, a deep learning model is used to classify the target water pump's operating sound signal into fault types to obtain the acoustic event types of the target water pump. The acoustic event types include normal state, cavitation fault, bearing wear fault, and impeller abnormal fault. The deep learning model is a 1D-CNN model, and the parameters of the 1D-CNN model are quantized using INT8.

[0015] A pump fault identification system for multi-source environments in pumping stations, used to implement the pump fault identification method described above, includes: a signal acquisition module for acquiring multimodal data within the pumping station and performing time synchronization processing on the multimodal data; an adaptive sound source separation module connected to the signal acquisition module for processing the multi-channel sound signals using a hybrid sound source adaptive separation strategy based on the quality assessment results of the mechanical vibration signals, to obtain multiple candidate sound source signals; a multimodal confidence fusion module connected to the adaptive sound source separation module for performing multimodal confidence fusion verification on each candidate sound source signal, and identifying the target pump operating sound signal from the multiple candidate sound source signals; an anomaly detection module connected to the multimodal confidence fusion module for performing anomaly detection processing on the target pump operating sound signal to obtain a fault detection result; and a fault classification module connected to the anomaly detection module for classifying the target pump operating sound signal into fault types based on the fault detection result to obtain a fault diagnosis result.

[0016] Preferably, it further includes: a multi-threaded pipeline scheduling module, connected to the signal acquisition module, the adaptive sound source separation module, the multimodal confidence fusion module, the anomaly detection module, and the fault classification module, for controlling the operation of the system, wherein the multi-threaded pipeline scheduling module includes processing threads, acquisition threads, and communication threads with decreasing priority; and a model inference acceleration module, connected to the fault classification module, for accelerating inference processing of the deep learning model based on the inference framework.

[0017] Beneficial Effects: This invention effectively overcomes the shortcomings of traditional blind separation methods that lack prior guidance by simultaneously acquiring multi-channel sound, mechanical vibration, and real-time operating parameters, and using vibration signal quality assessment results to constrain the adaptive separation process of mixed sound sources. It can accurately extract the sound signal of a single water pump even in multi-pump superposition and complex environmental interference. Subsequently, through multi-modal confidence fusion verification, the correspondence between the separated sound sources and the physical water pumps is automatically established, achieving full automation of sound source separation and fault identification without manual intervention. Simultaneously, by integrating operating parameters for anomaly detection and fault classification, it can dynamically adapt to changes in the pump room's operating environment, significantly reducing the risk of false alarms and missed alarms caused by fixed thresholds. Furthermore, by using prior information guidance and cascaded processing strategies to reduce computational complexity, it ensures the real-time operation of high-precision diagnostic results on edge computing devices, comprehensively addressing the stringent requirements of industrial sites for diagnostic accuracy, robustness, real-time performance, and automation. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system architecture diagram of the present invention; Figure 3 This is a schematic diagram of the sensor layout in the pump room according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0022] Reference Figure 1This invention provides a method for identifying pump faults in a multi-source environment of a pumping station, comprising: Step S1, collecting multimodal data in the pumping station and performing time synchronization processing on the multimodal data; wherein, the multimodal data includes multi-channel sound signals in the pumping station, mechanical vibration signals of each pump, and real-time operating parameters of each pump; Step S2, processing the multi-channel sound signals using a hybrid sound source adaptive separation strategy based on the quality assessment results of the mechanical vibration signals to obtain at least one candidate sound source signal; Step S3, performing multimodal confidence fusion verification on each candidate sound source signal to identify the target pump operating sound signal from the multiple candidate sound source signals; Step S4, performing anomaly detection processing on the target pump operating sound signal to obtain a fault detection result; Step S5, classifying the target pump operating sound signal by fault type based on the fault detection result to obtain a fault diagnosis result.

[0023] Specifically, in this embodiment of the invention, to address the problems of multiple sound sources overlapping due to the simultaneous operation of multiple water pumps in a pumping station, the lack of a complete closed loop from sound source separation to fault classification in existing solutions, and the poor adaptability of fixed thresholds due to dynamic changes in the acoustic environment, this invention simultaneously collects multi-channel sound signals, mechanical vibration signals, and real-time operating parameters. It adaptively selects a hybrid sound source separation strategy based on the mechanical vibration signal quality assessment results to obtain candidate sound sources. Then, it automatically identifies the target water pump's operating sound signal through multimodal confidence fusion verification. Subsequently, it sequentially performs anomaly detection and fault type classification, constructing a complete processing closed loop from sound source separation, target identification, anomaly detection, and fault classification. This avoids the shortcomings of existing technologies, such as the lack of prior constraints in blind source separation leading to unclear component correspondences, the inability of fixed baselines to adapt to dynamic environmental changes, and insufficient real-time performance due to cloud reliance. It achieves accurate extraction of the target water pump's operating sound and real-time accurate identification of fault types in complex scenarios with multiple sound sources overlapping and dynamic changes in the acoustic environment of a pumping station.

[0024] Specifically, refer to Figure 2 In actual deployment, a microphone array 11, an accelerometer 12, an environmental reference microphone 13, and a working condition parameter acquisition interface 16 are deployed inside the pump room. Specifically, refer to... Figure 3 First, multiple water pumps are arranged in the pump room, such as: first water pump 141, second water pump 142, third water pump 143, etc.

[0025] Next, the microphone array 11 is installed on a bracket 1.0m-2.5m above the ground and 2m-5m away from the water pump in the pump room. It is connected to the edge computing device through a digital audio interface to collect multi-channel sound signals of the water pump in the pump room in a non-contact manner, that is, the mixed signal of the sound of all the water pumps running, so as to provide multi-channel spatial information for subsequent sound source separation.

[0026] Secondly, an acceleration sensor 12 is installed vertically on the drive end bearing housing of each water pump, referring to... Figure 3 A first accelerometer 121 is installed vertically to the drive end bearing housing of the first water pump 141, a second accelerometer 122 is installed vertically to the drive end bearing housing of the second water pump 142, a third accelerometer 123 is installed vertically to the drive end bearing housing of the third water pump 143, and so on. The accelerometer 121 is preferably an IEPE type piezoelectric accelerometer. The output of each accelerometer is converted into a digital signal by a 24-bit ADC acquisition module and communicates with the edge computing device through the Modbus TCP protocol to collect the mechanical vibration signal of the corresponding water pump, which serves as the reference signal for subsequent semi-blind source separation and the physical causal verification basis for target sound source identification.

[0027] Then, a single omnidirectional microphone is installed at the farthest point from the main sound source in the corner of the pump room as an environmental reference microphone 13. It is connected to the edge computing device 15 via a USB sound card to collect environmental background noise separately and provide reference information on the environmental noise level for the dynamic baseline model. Finally, through the operating condition parameter acquisition interface 16 set at the PLC of the water pump control cabinet, the real-time speed, outlet flow rate and outlet pressure data of each water pump are read from the PLC of the water pump control cabinet using the communication protocol, so as to provide the current operating condition information of each water pump, which is used for the operating condition classification of dynamic baseline and the calculation of theoretical acoustic characteristics in the verification of operating mode.

[0028] In particular, all sensor data in this embodiment of the invention are synchronized via NTP (Network Time Protocol) with a synchronization accuracy better than 1ms, to ensure that the data of each modality are aligned with the time axis in subsequent processing, thus guaranteeing the accuracy of separation and identification.

[0029] In a preferred embodiment of the present invention, step S2 includes: step S21, performing frame segmentation processing on the multi-channel audio signal using preset parameters, and performing adaptive wavelet threshold noise reduction processing on each frame of the multi-channel audio signal to obtain multiple noise-reduced time-domain signal frames; step S22, calculating the average frequency domain coherence coefficient of the mechanical vibration signal and the multi-channel audio signal in the target frequency band, and judging the quality of the mechanical vibration signal based on the average frequency domain coherence coefficient, wherein if the average frequency domain coherence coefficient is greater than a first threshold, the mechanical vibration signal is judged to be of good quality; if the average frequency domain coherence coefficient is not greater than the first threshold and not less than a second threshold, the mechanical vibration signal is judged to be of poor quality. Generally; if the average value of the frequency domain coherence coefficient is less than the second threshold, the mechanical vibration signal quality is determined to be poor; step S23, according to the quality determination result of the mechanical vibration signal, an adaptive selection separation algorithm is used to perform sound source separation on the denoised time domain signal frame, wherein, if the mechanical vibration signal quality is good, a semi-blind source extraction algorithm is selected to perform sound source separation on the denoised time domain signal frame, extracting sound source components statistically related to the mechanical vibration signal, and using the sound source components as candidate sound source signals; if the mechanical vibration signal quality is average or poor, a FastICA blind separation algorithm is selected to perform sound source separation on the denoised time domain signal frame, separating multiple independent components, and using the multiple independent components as candidate sound source signals.

[0030] Specifically, in this embodiment of the invention, the quality of the mechanical vibration signal is quantified by pre-evaluating the frequency domain coherence between the mechanical vibration signal of each water pump and the multi-channel sound signal, and this is used as an adaptive switching criterion. When the quality of the mechanical vibration signal is good, semi-blind source extraction is used to improve the separation accuracy by utilizing vibration prior information. When the quality of the mechanical vibration signal deteriorates, the Fast Independent Component Analysis (FastICA) blind separation method is used to ensure the robustness of the system, thereby achieving stable and accurate extraction of the sound source components of the target water pump in the multi-source aliasing environment of the pump room.

[0031] Specifically, firstly, the multi-channel audio signal is processed by frame segmentation using preset parameters such as a sampling rate of 16kHz, a frame length of 256 (corresponding to 16ms), and a frame shift of 128 (corresponding to 8ms), so that each frame signal is analyzed under the quasi-steady-state assumption; wherein, the number of channels M of the multi-channel audio signal is greater than or equal to 4.

[0032] Next, adaptive wavelet thresholding preprocessing is performed on each frame of multi-channel audio signal. Through wavelet decomposition, hierarchical adaptive threshold calculation and soft thresholding, the steady-state background noise of the environment is suppressed while preserving the acoustic characteristics of the water pump, thereby improving the input signal-to-noise ratio of the subsequent separation algorithm.

[0033] Then, the mechanical vibration signal and the multi-channel sound signal are time-stamped and framed and windowed. The quality of the mechanical vibration signal is graded and evaluated by calculating the average frequency domain coherence coefficient of the two signals in the target frequency band (e.g., 100Hz~2000Hz).

[0034] In this way, the separation strategy is adaptively selected based on the quality assessment results. When the mechanical vibration signal quality is good, the target sound source is accurately locked through semi-blind source extraction. When the mechanical vibration signal quality is average or poor, the robustness of the separation is ensured through FastICA blind separation. This effectively solves the problem that the existing technology has a single separation strategy and cannot balance accuracy and robustness.

[0035] More specifically, in this embodiment of the invention, in addition to using the above-mentioned semi-blind source extraction algorithm for sound source separation, the Independent Vector Analysis (IVA) algorithm based on reference signals can also be used to replace the semi-blind source extraction. The IVA algorithm processes in the frequency domain and can better handle the frequency domain arrangement ambiguity problem of broadband acoustic signals, but the computational load is large, so it can be used as a high-quality offline analysis option. Similarly, the FastICA blind separation algorithm can be replaced by a deep clustering method based on time-frequency masking (such as the lightweight version of DEMUCS), which has higher separation accuracy when there is sufficient training data, but requires additional deployment of neural network models and higher requirements for the computing power of edge devices.

[0036] In a preferred embodiment of the present invention, step S21 involves performing adaptive wavelet threshold denoising on each frame of multi-channel audio signal, including: step S211, using wavelet decomposition to perform multi-level decomposition on each frame of multi-channel audio signal to obtain approximation coefficients and detail coefficients of each layer; step S212, estimating the noise standard deviation based on the median absolute deviation of the detail coefficients of the first layer to obtain the noise standard deviation estimate; step S213, using the Birgé-Massart strategy to calculate the adaptive threshold corresponding to the detail coefficients of each layer; step S214, using a soft thresholding function to perform thresholding on the detail coefficients of each layer; and step S215, using inverse discrete wavelet transform to reconstruct the processed detail coefficients and the original approximation coefficients of each layer to obtain the plurality of denoised time-domain signal frames.

[0037] Specifically, in this embodiment of the invention, adaptive wavelet threshold denoising is performed on each frame of multi-channel audio signal to suppress steady-state background noise in the environment while preserving the acoustic characteristics of the water pump, thus providing a higher signal-to-noise ratio input signal for subsequent separation algorithms.

[0038] Specifically, firstly, the time-domain signal of the multi-channel audio signal is transformed to the wavelet domain using Discrete Wavelet Transform (DWT) to obtain data at different scales and frequency bands. Then, the Daubechies 4 (db4) wavelet basis is used for 5-level decomposition, dividing each level of the signal into approximation coefficients representing the low-frequency profile and detail coefficients representing high-frequency details, thus obtaining the 5th level of approximation coefficients. And the detail coefficients cD1, cD2, …, cD5 of layers 1 to 5 corresponding to different high frequency bands.

[0039] Next, based on the detail coefficients cD1 of the first layer, the noise standard deviation is estimated, and the median absolute deviation of the first layer detail coefficients is calculated to obtain the noise standard deviation estimate. The specific formula is as follows: , , in, is the noise standard deviation estimate; MAD is the median absolute deviation of the level 1 detail coefficients; 0.6745 is the ratio coefficient of MAD to standard deviation under Gaussian distribution.

[0040] Then, considering that the amplitude of the wavelet coefficients of noise decreases layer by layer with the increase of the decomposition level, a different threshold is further determined for each level of detail coefficients to achieve hierarchical adaptive processing. Specifically, the Birgé-Massart strategy is used to determine the adaptive threshold for each level of detail coefficients, where the adaptive threshold of the j-th level detail coefficients is... Defined as: , in, Let j be the length of the detail coefficients at the j-th level; This is a universal threshold factor based on the theory of Gaussian noise extremum distribution; This is the layer attenuation factor, which reflects the characteristic that noise energy is halved as the number of layers increases.

[0041] Then, a soft thresholding function is applied to each detail coefficient ω to perform thresholding, in order to maintain the overall continuity of the signal while eliminating small-amplitude noise, specifically: , in, The detail coefficients after thresholding. This is the adaptive threshold corresponding to this layer. It is a symbolic function.

[0042] Finally, using the processed detail coefficients cD1~cD5 of layers 1 to 5 and the unprocessed approximation coefficients cA5 of layer 5, the inverse Mallat algorithm is used to upsample layer by layer and convolve with the reconstruction filter to achieve inverse discrete wavelet transform, and the denoised time-domain signal frame is reconstructed.

[0043] As can be seen, through the above adaptive wavelet threshold denoising process, the present invention can effectively suppress steady-state background noise in the environment while preserving the acoustic characteristics of the water pump, and significantly improve the input signal-to-noise ratio of the subsequent separation algorithm.

[0044] In particular, in addition to using the wavelet threshold denoising method described above for signal preprocessing, multi-channel Wiener filtering can also be used as an alternative. This multi-channel Wiener filtering utilizes the statistical correlation between multi-channel signals for denoising, and its effect is better when the number of channels is large, but the computational load is slightly greater than that of the wavelet threshold denoising method.

[0045] In a preferred embodiment of the present invention, step S22 includes: Step S221: The mechanical vibration signal and the multi-channel sound signal are time-stamped and then framed according to the same frame length and frame shift, and each frame signal is windowed. Step S222: Calculate the auto-power spectral density of the multi-channel sound signal, the auto-power spectral density of the mechanical vibration signal, and the cross-power spectral density between the multi-channel sound signal and the mechanical vibration signal. The specific calculation formula is as follows: , , , in, The power spectral density of the multi-channel audio signal; The power spectral density of the mechanical vibration signal is given. The cross-power spectral density between the multi-channel sound signal and the mechanical vibration signal; The Fourier transform of the k-th frame of the multi-channel audio signal; The Fourier transform of the k-th frame of the mechanical vibration signal. for The complex conjugate of ; K is the total number of frames; f is the frequency; Step S223: Calculate the frequency domain coherence coefficient at each frequency point within the target frequency band based on the self-power spectral density and the cross-power spectral density. The specific calculation formula is as follows: , in, It is a frequency domain coherence function; Step S224: Calculate the average frequency domain coherence coefficient of all frequency points within the target frequency band. The specific calculation formula is as follows: , in, This represents the average frequency domain coherence coefficient within the target frequency band. This is the set of effective frequency points after filtering.

[0046] Specifically, in this embodiment of the invention, by calculating the average frequency domain coherence coefficient of the mechanical vibration signal and the multi-channel sound signal in the target frequency band, the correlation between the mechanical vibration signal and the target water pump sound is further quantified, thereby providing objective criteria for the adaptive selection of the subsequent separation algorithm.

[0047] Specifically, firstly, the mechanical vibration signal a(t) collected by the accelerometer 12 is compared with the sound signal from the first channel of the microphone array 11. Alignment is achieved through NTP timestamps to ensure strict synchronization on the timeline; then, framing is performed with the same frame length and frame shift, for example, a frame length of 256 points (16ms) and a frame shift of 128 points (50% overlap); then, a Hanning window is applied to each frame signal to suppress spectral leakage.

[0048] In this process, the data for each second is processed and divided into several overlapping frames, resulting in approximately 124 frames per second.

[0049] Accordingly, the specific process of adding windows is as follows: First, the first channel audio signal The data is divided into frames, and then a Hanning window is applied to each frame to prevent frequency leakage. The specific formula is as follows: ; in, The frame length is 256 points (16ms). The definition of a Hanning window is: ; Next, a Fast Fourier Transform (FFT) is performed on the windowed k-th frame signal to obtain the spectrum of each frame. The specific formula is as follows: , in, For FFT points, Or larger, frequency interpolation is achieved by padding with zeros, which can be chosen to be the same as N; m is the frequency index, corresponding to the frequency. ; Using complex exponential basis functions, the time-domain signal is projected onto sine waves of different frequencies; It is a complex number, representing frequency. The amplitude and phase.

[0050] for The continuous representation, i.e. .

[0051] In particular, mechanical vibration signals The calculation method and Similarly, I will not go into details here.

[0052] Based on the spectrum of each frame, the auto-power spectral density of the first channel audio signal is calculated respectively. The self-power spectral density of the mechanical vibration signal and the cross-power spectral density between the two The specific calculation formula is as follows: , , , in, The power spectral density of the multi-channel audio signal; The power spectral density of the mechanical vibration signal is given. The cross-power spectral density between the multi-channel sound signal and the mechanical vibration signal; The Fourier transform of the k-th frame of the multi-channel audio signal; The Fourier transform of the k-th frame of the mechanical vibration signal. for The complex conjugate of ; K is the total number of frames; f is the frequency.

[0053] Then, based on the self-power spectral density and cross-power spectral density, the frequency domain coherence coefficient at each frequency point within the target frequency band is calculated. The specific formula is expressed as follows: .

[0054] Next, calculate the average frequency domain coherence coefficient of all effective frequency points within the target frequency band (e.g., 100Hz~2000Hz), expressed by the following formula: , in, This represents the average frequency domain coherence coefficient within the target frequency band. This is the set of effective frequency points after filtering.

[0055] Finally, the quality of the mechanical vibration signal is judged based on the average value of the frequency domain coherence coefficient, specifically as follows: like >0.7 indicates that the mechanical vibration signal quality is good; If 0.4≤ ≤0.7 indicates that the mechanical vibration signal quality is generally poor; like A value <0.4 indicates poor mechanical vibration signal quality.

[0056] In this way, the mechanical vibration signal quality is graded and evaluated by the calculated average MSC value, providing an objective basis for subsequent adaptive selection of semi-blind source extraction or FastICA blind separation.

[0057] In particular, in addition to using the method of calculating the average value of the frequency domain coherence coefficient as described above to evaluate the quality of mechanical vibration signals, the signal-to-noise ratio of mechanical vibration signals can also be estimated as a quality criterion. That is, the ratio of the peak value to the noise floor in the power spectrum of mechanical vibration signals can be calculated to determine the quality. This method is simpler to calculate, but its robustness to non-stationary noise is slightly worse.

[0058] In a preferred embodiment of the present invention, in step S23, The specific steps for source separation of the denoised time-domain signal frame using the semi-blind source extraction algorithm are as follows: Construct the covariance matrix of the denoised time-domain signal frame, and the cross-covariance vector between the denoised time-domain signal frame and the mechanical vibration signal; Construct a generalized eigenvalue equation, solve for the eigenvector corresponding to the largest eigenvalue of the generalized eigenvalue equation, and obtain the optimal solution mixing vector; The optimal unmixing vector is multiplied by the denoised time-domain signal frame to obtain the candidate sound source signal; The specific steps for source separation of the denoised time-domain signal frame using the FastICA blind separation algorithm are as follows: The denoised time-domain signal frame is preprocessed with whitening to obtain a whitened signal matrix, the specific formula of which is: , in, This is the signal matrix after whitening; This is the whitening matrix; This is the time-domain signal frame after noise reduction; The unmixing matrix is ​​solved using a fixed-point iterative algorithm based on maximizing negative entropy. The specific solution formula is as follows: , , , , in, This is the unmixed vector after iterative updates; The unmixing vector; It is a nonlinear function; After iterative convergence, the unmixed matrix is ​​multiplied by the whitened signal matrix to obtain the candidate sound source signal.

[0059] Specifically, in this embodiment of the invention, based on the quality assessment results of the mechanical vibration signal, a semi-blind source extraction algorithm or a FastICA blind separation algorithm is adaptively selected to perform sound source separation on the denoised time-domain signal frame.

[0060] Specifically, when the mechanical vibration signal quality is good, that is... When the value is greater than 0.7, a semi-blind source extraction algorithm is used, which specifically includes: First, let the M-channel signal matrix of the microphone array 11 in the current frame be X = [x1, x2, …, xM]^T, where X is an M×N matrix, N is the number of sampling points in the frame, and the mechanical vibration signal vector is a, where a is a 1×N vector.

[0061] Next, calculate the covariance matrix of X: The array It is an M×M matrix.

[0062] Simultaneously, calculate the cross-covariance vector between X and a: ,Should It is an M×1 vector.

[0063] Next, we construct the generalized eigenvalue equations: , in, This is the demixing vector (spatial filter). It is an M×1 vector, each This indicates the importance or contribution of the i-th microphone in extracting the target sound source; It is a scalar, a scalar eigenvalue of the generalized eigenvalue equation, and the final value of the maximizing objective.

[0064] Then, the eigenvector corresponding to the largest eigenvalue of the generalized eigenvalue equation is solved to obtain the optimal solution mixing vector. Specifically, the solution process is as follows: 1) To The Cholesky decomposition is expressed by the following formula: , Where L is a lower triangular matrix; 2) Solve the trigonometric equations: This yields an M×1 vector y; 3) Solve the trigonometric equations ,get ; 4) To Normalization is performed to obtain the optimal unmixing vector. The specific formula is expressed as follows: ; 5) Calculation The specific formula is expressed as follows: .

[0065] Then, the optimal unmixing vector is multiplied by the denoised time-domain signal frame to obtain the target sound source estimate. The specific formula is expressed as follows: ,Should It is a 1×N vector; That is, the target water pump sound signal is extracted from the current frame and used as a candidate sound source signal.

[0066] This semi-blind source extraction algorithm uses mechanical vibration signals as reference constraints to transform the blind source separation problem into generalized eigenvalue decomposition with prior knowledge. It has low computational cost, stable convergence, and the separation results directly correspond to the target water pump, eliminating the need for subsequent complex component identification.

[0067] Correspondingly, when the quality of the mechanical vibration signal is poor, i.e. When the value is ≤0.7, the algorithm reverts to the FastICA blind separation algorithm, which includes: First, X is preprocessed by whitening to obtain the whitened signal matrix, as shown in the following formula: , in, This is the signal matrix after whitening; Let be the whitening matrix, such that The rows are uncorrelated and have unit variance; This is the time-domain signal matrix after noise reduction.

[0068] Next, a fixed-point iterative algorithm based on maximizing negative entropy is used to solve for the unmixing matrix. The specific solution formula is as follows: , , Its derivative is , in, This is the unmixed vector after iterative updates; The unmixing vector; It is a nonlinear function.

[0069] Then, repeat the above iterative steps until convergence, and obtain M unmixed vectors to form the unmixed matrix W.

[0070] Finally, the demixing matrix is ​​multiplied by the whitened signal matrix to obtain the separated independent components, specifically expressed as: Y = W·Z, where Y is an M×N matrix, and these independent components are used as candidate sound source signals.

[0071] The FastICA blind separation algorithm decomposes the mixed signal into multiple independent source signal estimates by maximizing the statistical independence of each output component when there is no reliable prior information. It has fast convergence speed and high computational efficiency, making it suitable for real-time processing at the edge.

[0072] It is evident that by dynamically selecting the separation strategy based on the quality of the mechanical vibration signal, high-precision separation can be achieved using semi-blind source extraction when the quality is good, and the separation robustness can be ensured by reverting to FastICA when the quality is poor. This effectively solves the problem of insufficient adaptability of a single separation strategy in complex pump room environments.

[0073] In particular, in addition to using the semi-blind source extraction algorithm mentioned above for sound source separation, an independent vector analysis algorithm based on reference signals can also be used as an alternative. This algorithm processes in the frequency domain and can better handle the frequency domain arrangement ambiguity problem of broadband acoustic signals, but it has a large computational load and can be used as a high-quality offline analysis option.

[0074] Similarly, in addition to the FastICA blind separation algorithm mentioned above, a deep clustering method based on time-frequency masking can be used as an alternative. It has higher separation accuracy when there is sufficient training data, but requires additional deployment of neural network models and higher computing power requirements for edge devices.

[0075] In a preferred embodiment of the present invention, step S3 includes: Step S31: Perform physical causality verification, operating mode verification and spatial consistency verification on each candidate sound source signal to obtain the coherence score, operating condition matching score and spatial consistency score corresponding to each candidate sound source signal. Step S32: The coherence score, working condition matching score, and spatial consistency score of each candidate sound source signal are weighted and fused to obtain the comprehensive confidence score corresponding to each candidate sound source signal. The specific calculation formula is as follows: , in, The overall confidence level; The score is for coherence. Score for matching operating conditions; Spatial consistency score; , , All of these are weighting coefficients, which are adaptively adjusted according to the quality changes of the mechanical vibration signal. Step S33: Compare the overall confidence level of each candidate sound source signal with a preset threshold. If the overall confidence level is not less than the preset threshold, the candidate sound source signal is determined to be the target water pump operating sound signal; otherwise, it is discarded.

[0076] Specifically, considering that the correspondence between the components after blind separation and the physical sound source in the existing technology requires manual judgment or relies on only a single criterion, resulting in poor reliability, this embodiment of the invention proposes a multimodal confidence evaluation framework that integrates physical causal verification, operation mode verification, and spatial consistency verification. The following three evaluations are performed on each candidate sound source signal s_candidate output in step S2, thereby achieving automatic and robust identification of the target water pump operation sound signal.

[0077] In the semi-blind source extraction case, there is only one candidate sound source signal s_candidate, while in the FastICA blind separation case, there are M candidate sound source signals s_candidate.

[0078] Specifically, firstly, physical causality verification is performed on the candidate sound source signal s_candidate to obtain a coherence score. Specifically, the average frequency domain coherence coefficients of the candidate sound source signal s_candidate and the mechanical vibration signal a within the target frequency band (e.g., 100Hz~2000Hz) are calculated. And the average value of the frequency domain coherence coefficient As a coherence score The value range is [0, 1].

[0079] This physical causality verification verifies from the perspective of physical causality whether the candidate sound source is caused by the mechanical vibration of the target water pump. If the two are highly coherent, it indicates that the sound wave originates from the vibration at the location of the accelerometer.

[0080] Next, the candidate sound source signal s_candidate is validated for its operating mode to obtain the operating condition matching score. Specifically, the spectrum of the candidate sound source signal s_candidate is extracted using short-time Fourier transform (STFT), and its fundamental frequency is identified through peak search. and the amplitudes of the first three harmonics, A1, A2, and A3; At the same time, the theoretical fundamental frequency is calculated based on the current real-time speed n (unit: rpm) of the water pump. The specific calculation formula is as follows: .

[0081] The working condition matching score can be obtained. : , in, The total energy across the entire frequency band is given. If the extracted fundamental frequency deviates from the theoretical fundamental frequency by less than 5%, and the proportion of harmonic energy conforms to the normal operating mode of the water pump, then... Close to 1; It is worth mentioning that when verifying operating modes, in cases where rotational speed data is lacking, the fundamental frequency can be estimated directly from the acoustic signal using the harmonic product spectrum method and matched with historical normal operating modes. Alternatively, dual verification can be performed by combining the impeller pass frequency (fundamental frequency × number of blades).

[0082] This operation mode verification verifies, from the perspective of operation mechanism, whether the spectral characteristics of the candidate sound source conform to the theoretical acoustic mode of the target water pump under the current operating conditions. The sound of the water pump has a fundamental frequency and harmonic structure that is strictly related to the rotational speed, which is a key feature that distinguishes it from the noise of other equipment.

[0083] Then, spatial consistency verification is performed on the candidate sound source signal s_candidate to obtain a spatial consistency score. Specifically, using microphone array 11, the generalized cross-correlation-phase transform method (GCC-PHAT) is employed to estimate the direction of arrival (DOA) azimuth angle of the candidate sound source signal. Let the actual installation azimuth angle of the target water pump be... ,Should This can be calibrated during system deployment. Then, the spatial consistency score is calculated. for: .

[0084] It is worth mentioning that when the microphone array does not support DOA estimation, the received signal energy ratio can be used as a simplified scheme, that is, to compare the energy difference of different microphones receiving the candidate sound source and match it with the distance attenuation model from the target water pump to each microphone.

[0085] In the pump room, different water pumps are located in different spaces, and the direction of the sound source is an important basis for distinguishing each water pump. This operation mode verification verifies from the perspective of spatial location whether the physical source of the candidate sound source is consistent with the actual location of the target water pump.

[0086] Then, the coherence score, operating condition matching score, and spatial consistency score of each candidate sound source signal are weighted and fused, and the specific calculation formula is as follows: , in, The overall confidence level; The score is for coherence. Score for matching operating conditions; Spatial consistency score; , , All of these are weighting coefficients, which are adaptively adjusted according to the quality changes of the mechanical vibration signal.

[0087] Among them, the weighting coefficient , , The dynamic adjustment rules are shown in Table 1 below: Table 1. Rules for Dynamic Adjustment of Weights

[0088] Finally, the calculated overall confidence level C is compared with the preset threshold of 0.65. When C is not less than the preset threshold, that is... The candidate sound source signal was determined to be the sound signal of the target water pump in operation.

[0089] Thus, by integrating physical causal verification, operational mode verification, and spatial consistency verification into a multimodal confidence assessment framework, and dynamically adjusting the weights of each verification dimension based on the quality of the mechanical vibration signal, physical causal verification is used as the primary basis when the acceleration signal is reliable, and operational mode verification is used as the primary basis when the acceleration signal is unreliable. This achieves automatic and robust identification of the target water pump's operating sound signal, solving the problem of difficult component identification after blind separation in existing technologies.

[0090] In a preferred embodiment of the present invention, step S4 includes: Step S41: Extract features from the target water pump's operating sound signal to obtain an acoustic feature vector, which includes Mel frequency cepstral coefficient features, wavelet packet energy features, and spectral features. Step S42: During the normal operation of the water pump, the target water pump is divided into multiple operating condition intervals according to the real-time operating condition parameters. Normal operation sound data of each operating condition interval is collected, and a dynamic baseline model corresponding to each operating condition interval is constructed. Step S43: During the online monitoring phase, select the corresponding optimal dynamic baseline model based on the current real-time operating parameters, and input the acoustic feature vector into the optimal dynamic baseline model for anomaly detection processing to obtain the fault detection result.

[0091] Specifically, considering that existing technologies mostly use static thresholds or fixed baselines and cannot adapt to the dynamic changes in pump room environment and equipment status over time, this embodiment of the invention proposes a dynamic baseline model based on operating condition classification, and combines it with a single-class support vector machine for anomaly detection to significantly reduce the false alarm rate.

[0092] Specifically, a 24-dimensional lightweight acoustic feature vector is first extracted from each frame of the target water pump's operating sound signal. This 24-dimensional lightweight acoustic feature vector includes 13-dimensional Mel-frequency cepstral coefficients (MFCC) features, 8-dimensional wavelet packet energy features, and 3-dimensional spectral features, where: The 13-dimensional frequency cepstral coefficient (MFCC) feature is sensitive to changes in the timbre of the water pump sound and is a key feature for distinguishing different fault types. The 13-dimensional frequency cepstral coefficient (MFCC) feature has 40 filter banks and a frequency range of 20Hz-8000Hz, which is used to simulate the characteristics of human hearing. The 8-dimensional wavelet packet energy feature is particularly sensitive to the increase in high-frequency energy caused by bearing wear, etc., and is used to capture the changes in the energy distribution of the signal in different frequency bands; For the energy characteristics of this 8-dimensional wavelet packet, it is preferable to use the db4 wavelet basis for 3-level decomposition to obtain the normalized energy of 8 frequency bands; This 3D spectral feature includes the spectral centroid, spectral bandwidth, and spectral roll-off point, which is used to describe the overall morphological changes of the spectrum. It requires little computation and can be used for rapid screening.

[0093] Next, considering that the normal range of acoustic characteristics of water pumps varies significantly at different speeds, segmented modeling can avoid misjudgments caused by changes in operating conditions. Based on this, a dynamic baseline model is further constructed. Specifically, during the healthy operation phase of the water pump, such as the normal period confirmed by experts after new installation or major overhaul, normal operating sound data under various typical operating conditions are collected, and multiple operating condition intervals are divided according to real-time operating parameters such as speed or flow rate.

[0094] Taking speed as an example, the operating conditions are divided into several intervals, such as [1450, 1460) rpm, [1460, 1470) rpm, etc. Within each operating condition interval, the mean vector μ and standard deviation vector σ of each dimension of the feature vector are calculated using a sliding window with a width of 100 frames, and stored as the normal baseline parameters of the current operating condition interval.

[0095] In this way, a normal acoustic characteristic reference system associated with the operating conditions can be established, thereby providing a dynamic baseline standard that matches the current operating conditions for subsequent anomaly detection. This enables the system to automatically adapt to changes in acoustic characteristics under different operating conditions of the water pump and effectively distinguish between normal acoustic fluctuations caused by changes in operating conditions and abnormal acoustic changes caused by actual faults.

[0096] It's worth noting that when sufficient historical data is available, Gaussian process regression (GPR) can be used to establish a continuous mapping model from "operating parameters to normal eigenvectors". This GPR can provide the uncertainty of the prediction, but it has a large computational cost, making it suitable for deployment to edge devices after periodically updating the baseline offline.

[0097] Then, anomaly detection is performed during the online monitoring phase. Specifically, a linear one-class support vector machine (One-Class SVM) is used as the anomaly detector, and a linear kernel is chosen to ensure the inference speed at the edge. During the training phase, feature vectors of normal samples are collected for each operating condition range, and the corresponding linear single-class support vector machine model is trained. The parameter ν is set to 0.01, which means that 1% of normal samples are allowed to be misclassified as abnormal.

[0098] During the detection phase, the corresponding single-class support vector machine model is selected based on the current real-time operating parameters, and the distance from the current acoustic feature vector to the decision boundary of the single-class support vector machine model is calculated. If the distance is negative, it indicates that the feature vector is outside the decision boundary, and the frame is determined to be an abnormal frame.

[0099] Furthermore, a sliding window confirmation mechanism is adopted to avoid false alarms caused by transient interference. That is, when more than M frames in K consecutive frames are judged as abnormal frames, the existence of abnormal noise is finally confirmed. For example, K is 10 and M is 8.

[0100] In this way, by using a dynamic baseline model based on operating conditions and single-class support vector machine anomaly detection, the system can learn the complex boundaries of the "normal" state and automatically adapt to the dynamic changes in pump room background noise and operating conditions. Compared with fixed threshold or statistical threshold (such as 3σ) methods, it is more accurate in detecting anomalies of multi-feature joint distribution and effectively reduces the false alarm rate.

[0101] In a preferred embodiment of the present invention, in step S5, a deep learning model is used to classify the fault types of the target water pump's operating sound signal to obtain the acoustic event types of the target water pump. The acoustic event types include normal state, cavitation fault, bearing wear fault, and impeller abnormal fault. The deep learning model is a 1D-CNN model, and the parameters of the 1D-CNN model are quantized using INT8.

[0102] Specifically, this step is triggered immediately upon detecting abnormal noise in step S4 to classify the fault type. Since existing fault classification models rely on cloud processing, which suffers from network latency, or employ complex algorithms with high computational demands, making real-time inference difficult to achieve on edge devices, this embodiment of the invention employs a lightweight deep learning model—the lightweight 1D-CNN model—combined with INT8 quantization, to achieve high-precision, low-latency fault classification on edge computing devices.

[0103] Specifically, firstly, a lightweight 1D-CNN model is constructed, the structure of which is shown in Table 2 below: Table 2 Lightweight 1D-CNN Model Architecture

[0104] As shown in Table 2 above, this lightweight 1D-CNN model takes the 24-dimensional acoustic feature vector extracted in step S4 as input, and extracts local temporal features, mid-level combined features and high-level abstract features through three layers of one-dimensional convolution. The temporal dimension is compressed to a fixed length through global average pooling. Finally, the fully connected layer outputs the probabilities of four categories through the Softmax function, which correspond to normal state, cavitation failure, bearing wear failure and impeller abnormal failure, respectively, thereby achieving fine differentiation of each failure type.

[0105] Specifically, the model structure described above can also be replaced with a 1D adapted version of MobileNetV2 (depthseparable convolution), which has fewer parameters. When higher accuracy is required, a 2D-CNN channel with a CWT time-frequency map as input can be added, and the output can be fused after paralleling with the 1D-CNN channel. The classifier can be replaced with a lightweight gradient booster (LightGBM), which can achieve accuracy comparable to CNN and faster inference when feature engineering is sufficient.

[0106] Next, the model was trained and quantized. Specifically, the collected historical failure data and normal operation data were used for training. The training configuration used the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 training epochs.

[0107] Among them, the normal operation data is preferably no less than 500 segments in each category, and each segment is 1 second long.

[0108] After training, the model parameters were quantized from FP32 precision to INT8 precision using ONNX Runtime's static quantization tool. This INT8 quantization is a key technology for achieving real-time inference at the edge, which can reduce the model size by about 70%, increase inference speed by 2 to 3 times, and keep the accuracy loss within 2%.

[0109] In addition, considering the scarcity of fault samples in industrial settings and the limited generalization ability of the model after deployment, this embodiment also includes an incremental learning and voiceprint database update mechanism. Specifically, when the fault classification confidence score output by the 1D-CNN model is higher than 0.85, the acoustic feature vector and classification result corresponding to the abnormal segment are saved to the local voiceprint cache, and the cache samples are remotely reviewed by maintenance experts periodically (e.g., weekly) to confirm that they are correct and then marked as valid samples.

[0110] After accumulating a certain number of new samples, incremental training is started during the idle period of the edge computing device to update the model parameters. The idle period can be determined based on the time period and CPU and memory usage data, such as the time period being 2-5 am and the CPU usage being less than 100% and the total memory usage being less than 40%.

[0111] In this way, by extracting temporal patterns from acoustic feature vectors through multi-layer one-dimensional convolution, fine distinctions between different fault types can be achieved. Furthermore, INT8 quantization significantly reduces model size and inference latency, enabling the system to achieve high-precision, low-latency real-time fault classification on edge computing devices. At the same time, the incremental learning mechanism enables the system to have continuous learning capabilities. As the running time accumulates, the model's accuracy in identifying fault features of specific pump rooms and specific water pumps continuously improves, effectively solving the problems of scarce fault samples and limited model generalization ability in industrial fields.

[0112] It is worth mentioning that, in addition to adopting the above-mentioned method of incremental learning and voiceprint database update on the edge computing device, a unified voiceprint database can also be established in the cloud, and multiple edge devices can share the incremental learning results to form a federated learning architecture, so as to further improve the generalization ability and evolution efficiency of the model in the distributed deployment scenario.

[0113] As a preferred embodiment of the present invention, the system adopts an edge real-time processing architecture, and achieves end-to-end real-time processing on the edge computing device 15 through a multi-threaded pipeline architecture and model inference acceleration technology.

[0114] Specifically, in the embodiments of the present invention, Regarding the multi-threaded pipeline architecture, this system employs a collaborative approach involving acquisition threads, processing threads, and communication threads, specifically as follows: The acquisition thread continuously reads multi-channel audio signals from the microphone array 11 at a preset sampling rate (e.g., 16kHz) and writes them into a circular buffer with a capacity of 10 frames, blocking and waiting for condition variables. The acquisition thread is set to normal priority to ensure that audio data is not lost and to decouple it from the processing thread; The processing thread triggers a processing cycle every time a preset number of sampling points (e.g., 256 sampling points, i.e., 16ms) are accumulated, retrieves a frame of data from the circular buffer, and executes steps S2 to S5 as described above in sequence. This processing thread is set to the SCHED_FIFO real-time scheduling strategy via the sched_setscheduler system call; The priority of this processing thread is set to the highest, such as 99, to ensure that the processing thread gets enough CPU time slices and avoids processing delays caused by system load fluctuations. The communication thread asynchronously reports the fault diagnosis results to the monitoring center via a communication protocol (such as MQTT or Modbus TCP protocol), thus not blocking the processing flow; The communication thread is set to low priority to ensure that diagnostic results are transmitted in a timely manner without affecting real-time processing.

[0115] Regarding model inference acceleration, the fault type classification processing in step S5 above adopts a deep learning model, which is deployed on the ONNX Runtime inference framework and accelerated by the AVX2 instruction set on the operating system (such as Debian 11 system) of the edge computing device 15. Among them, the CPU backend of the ONNX Runtime inference framework automatically utilizes the vectorized instructions of modern x86 processors, significantly improving the efficiency of matrix operations and keeping the inference latency of lightweight models at around 3ms, leaving ample time margin for the entire processing flow.

[0116] More specifically, on the hardware platform of this embodiment of the invention, the average processing time of each stage of the system is further measured, with 256 points per frame, in milliseconds. The measurement results are shown in Table 3 below: Table 3. Processing Stage - Average Time Mapping Table

[0117] As can be seen, in the above processing stages, the total time for selecting the optimal path for semi-blind source extraction is approximately 10.5ms, which is less than the audio frame length of 16ms, thus meeting the requirements for real-time processing. In terms of memory usage, the physical memory usage during system operation is approximately 180MB, which fully complies with the resource constraints of edge devices.

[0118] Reference Figure 2 The present invention also provides a pump fault identification system for multi-source environments in pump rooms, used to implement the pump fault identification method for multi-source environments in pump rooms as described above, including: Signal acquisition module 1 is used to acquire multimodal data in the pump room and perform time synchronization processing on the multimodal data; The adaptive sound source separation module 2 is connected to the signal acquisition module 1 and is used to process the multi-channel sound signal using a hybrid sound source adaptive separation strategy based on the quality assessment result of the mechanical vibration signal to obtain multiple candidate sound source signals. The multimodal confidence fusion module 3 is connected to the adaptive sound source separation module 2 and is used to perform multimodal confidence fusion verification on each candidate sound source signal to identify the target water pump operation sound signal from multiple candidate sound source signals. Anomaly detection module 4, connected to the multimodal confidence fusion module 3, is used to perform anomaly detection processing on the target water pump's operating sound signal to obtain fault detection results; The fault classification module 5 is connected to the anomaly detection module 4 and is used to classify the target water pump's operating sound signal according to the fault detection results to obtain fault diagnosis results. A multi-threaded pipeline scheduling module 6 is connected to the signal acquisition module 1, the adaptive sound source separation module 2, the multimodal confidence fusion module 3, the anomaly detection module 4, and the fault classification module 5, and is used to control the operation of the system. The multi-threaded pipeline scheduling module includes processing threads, acquisition threads, and communication threads with decreasing priority. The model inference acceleration module 7 is connected to the fault classification module 5 and is used to accelerate the inference processing of the deep learning model based on the inference framework. The signal acquisition module 1 includes: Microphone array 11, located inside the pump room, is used to collect multi-channel sound signals within the pump room; An accelerometer 12 is installed in the drive end bearing housing of each water pump to collect the mechanical vibration signals of each water pump. An environmental reference microphone 13 is placed in a corner of the pump room to collect environmental noise signals inside the pump room. The operating condition parameter acquisition interface 16 is located at the water pump control cabinet and is used to collect real-time operating condition parameters of each water pump.

[0119] Specifically, in the embodiments of the present invention, reference is made to... Figure 2 The system is architecturally divided into four main functional layers, specifically including: The signal acquisition layer includes a microphone array 11, an accelerometer 12, an environmental reference microphone 13, and a working condition parameter acquisition interface 16. These devices together constitute the signal acquisition module 1, which is responsible for synchronously acquiring multimodal signals in the pump room. The sound source separation and target recognition layer includes an adaptive sound source separation module 2 and a multimodal confidence fusion module 3, which are responsible for separating and recognizing the operating sound of the target water pump from the mixed sound; The anomaly detection and fault classification layer includes anomaly detection module 4 and fault classification module 5, which are responsible for determining whether there is abnormal noise and identifying the fault type. The edge real-time processing architecture layer includes a multi-threaded pipeline scheduling module 6 and a model inference acceleration module 7, which are responsible for implementing low-latency real-time processing on the edge computing device 15.

[0120] Accordingly, the multi-threaded pipeline scheduling module 6 employs three threads—a data acquisition thread, a processing thread, and a communication thread—working collaboratively. Among them, The acquisition thread continuously reads data from the microphone array 11 at a sampling rate of 16kHz, writes it to a circular buffer with a capacity of 10 frames, and blocks while waiting for a condition variable. This acquisition thread is set to normal priority to ensure that audio data is not lost and to decouple it from the processing thread.

[0121] The processing thread triggers a processing cycle every 256 sampling points (i.e., 16ms) to retrieve a frame of data from the circular buffer and execute steps S2 to S5 in sequence. The processing thread is set to the SCHED_FIFO real-time scheduling strategy through the sched_setscheduler system call, with the priority set to the highest (e.g., 99) to ensure that the processing thread gets enough CPU time slices and avoids processing delays caused by system load fluctuations.

[0122] The communication thread asynchronously reports the fault diagnosis results to the monitoring center via MQTT or Modbus TCP protocol without blocking the processing flow. This communication thread is set to low priority so that the diagnosis results can be transmitted in a timely manner without affecting real-time processing.

[0123] Furthermore, in the model inference acceleration module 7, the fault classification model adopts the ONNX Runtime inference framework and enables AVX2 instruction set acceleration on the Debian 11 system. The CPU backend of ONNX Runtime automatically utilizes the vectorized instructions of modern x86 processors to significantly improve matrix operation efficiency, so that the inference latency of the lightweight model is controlled at about 3ms, leaving sufficient time margin for the entire processing flow.

[0124] In summary, this invention proposes a method and system for water pump fault identification in multi-source environments of pump rooms. It synchronously collects multi-channel sound signals, mechanical vibration signals of each water pump, and real-time operating parameters of each water pump in the pump room and performs time synchronization processing. Based on the quality assessment results of the mechanical vibration signals, it adaptively selects a semi-blind source extraction or FastICA blind separation strategy to obtain candidate sound sources. Then, it automatically identifies the target water pump's operating sound signal through multimodal confidence verification that integrates physical causality, operating mode, and spatial consistency. Subsequently, it sequentially performs dynamic baseline anomaly detection based on operating condition classification and fault type classification based on a lightweight 1D-CNN model, thus constructing a complete processing closed loop from sound source separation, target identification, anomaly detection, and fault classification.

[0125] Compared with the prior art, the present invention has the following significant advantages: By using the "semi-blind source extraction + FastICA" adaptive separation strategy, the independent sound source of the target water pump can be accurately extracted from the mixed sound of multiple pumps running simultaneously, thus solving the problem of multi-source aliasing in the pump room. A multimodal confidence assessment framework that integrates physical causality, operating mode, and spatial consistency can reliably identify the sound of the target water pump even when the quality of the accelerometer signal fluctuates or the operating conditions change, thus improving the robustness of target identification. Based on the dynamic baseline model with different operating conditions and the single-class SVM anomaly detection mechanism, it can automatically adapt to changes in pump room background noise and operating conditions, reduce false alarms, and enhance dynamic environmental adaptability. By employing a lightweight 1D-CNN model and INT8 quantization technology, combined with multi-threaded real-time scheduling and ONNX Runtime acceleration, real-time processing requirements can be met on edge computing devices without relying on the cloud or high-performance servers, thus achieving high-precision real-time diagnosis at the edge. By automatically storing high-confidence samples and using an incremental learning mechanism, the system's diagnostic accuracy can be continuously improved over time, enabling it to continuously learn and evolve.

[0126] The above description is merely a preferred embodiment of the present invention and does not limit the implementation and protection scope of the present invention. Those skilled in the art should realize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying pump faults in a multi-source environment of a pumping station, characterized in that, include: Step S1: Collect multimodal data in the pump room and perform time synchronization processing on the multimodal data; wherein, the multimodal data includes multi-channel sound signals in the pump room, mechanical vibration signals of each water pump, and real-time operating parameters of each water pump. Step S2: Based on the quality assessment results of the mechanical vibration signal, the multi-channel sound signal is processed using a hybrid sound source adaptive separation strategy to obtain at least one candidate sound source signal. Step S2 includes: Step S21: The multi-channel audio signal is processed into frames using preset parameters, and each frame of the multi-channel audio signal is subjected to adaptive wavelet threshold noise reduction to obtain multiple noise-reduced time-domain signal frames. Step S22: Calculate the average frequency domain coherence coefficient of the mechanical vibration signal and the multi-channel sound signal within the target frequency band, and judge the quality of the mechanical vibration signal based on the average frequency domain coherence coefficient. If the average value of the frequency domain coherence coefficient is greater than the first threshold, the mechanical vibration signal is determined to be of good quality. If the average value of the frequency domain coherence coefficient is not greater than the first threshold and not less than the second threshold, the mechanical vibration signal is judged to be of average quality. If the average value of the frequency domain coherence coefficient is less than the second threshold, the mechanical vibration signal is determined to be of poor quality. Step S23: Based on the quality judgment result of the mechanical vibration signal, an adaptive selection separation algorithm is used to perform sound source separation on the denoised time-domain signal frame, wherein... If the mechanical vibration signal is of good quality, a semi-blind source extraction algorithm is selected to separate the sound sources of the denoised time-domain signal frame, extract the sound source components that are statistically related to the mechanical vibration signal, and use the sound source components as candidate sound source signals. If the quality of the mechanical vibration signal is average or poor, the FastICA blind separation algorithm is selected to perform sound source separation on the noise-reduced time-domain signal frame, separating out multiple independent components, and using the multiple independent components as candidate sound source signals; Step S3: Perform multimodal confidence fusion verification on each candidate sound source signal to identify the target water pump operating sound signal from the multiple candidate sound source signals; Step S3 includes: Step S31: Perform physical causality verification, operating mode verification, and spatial consistency verification on each candidate sound source signal to obtain the coherence score, operating condition matching score, and spatial consistency score corresponding to each candidate sound source signal; wherein, The physical causality verification verifies from the perspective of physical causality whether the candidate sound source is caused by the mechanical vibration of the target water pump; The operation mode verification verifies, from the perspective of operation mechanism, whether the spectral characteristics of the candidate sound source conform to the theoretical acoustic mode of the target water pump under the current operating conditions. The spatial consistency verification verifies, from a spatial location perspective, whether the physical source of the candidate sound source is consistent with the actual location of the target water pump. Step S32: The coherence score, working condition matching score, and spatial consistency score of each candidate sound source signal are weighted and fused to obtain the comprehensive confidence score corresponding to each candidate sound source signal. The specific calculation formula is as follows: ， in, The overall confidence level; The score is based on coherence. Score for matching operating conditions; Spatial consistency score; , , All of these are weighting coefficients, which are adaptively adjusted according to the quality changes of the mechanical vibration signal. Step S33: Compare the overall confidence score of each candidate sound source signal with a preset threshold. If the overall confidence score is not less than the preset threshold, the candidate sound source signal is determined to be the target water pump operating sound signal; otherwise, it is discarded. Step S4: Perform anomaly detection processing on the target water pump's operating sound signal to obtain fault detection results; Step S5: Based on the fault detection results, classify the fault type of the target water pump's operating sound signal to obtain the fault diagnosis results.

2. The method for identifying pump faults in a multi-sound-source environment of a pumping station according to claim 1, characterized in that, In step S21, adaptive wavelet threshold noise reduction processing is performed on each frame of multi-channel audio signal, including: Step S211: Wavelet decomposition is used to perform multi-level decomposition on each frame of multi-channel audio signal to obtain the approximation coefficients and detail coefficients of each level. Step S212: Estimate the noise standard deviation based on the median absolute deviation of the detail coefficients of the first layer to obtain the noise standard deviation estimate. Step S213: The Birgé-Massart strategy is used to calculate the adaptive threshold corresponding to the detail coefficients of each layer; Step S214: Use a soft thresholding function to perform thresholding on the detail coefficients of each layer; Step S215: The inverse discrete wavelet transform is used to reconstruct the detail coefficients and the original approximation coefficients of each layer after processing, so as to obtain the multiple noise-reduced time-domain signal frames.

3. The method for identifying pump faults in a multi-sound-source environment of a pumping station according to claim 1, characterized in that, Step S22 includes: Step S221: The mechanical vibration signal and the multi-channel sound signal are time-stamped and then framed according to the same frame length and frame shift, and each frame signal is windowed. Step S222: Calculate the auto-power spectral density of the multi-channel sound signal, the auto-power spectral density of the mechanical vibration signal, and the cross-power spectral density between the multi-channel sound signal and the mechanical vibration signal. The specific calculation formula is as follows: ， ， ， in, The power spectral density of the multi-channel audio signal; The power spectral density of the mechanical vibration signal is given. The cross-power spectral density between the multi-channel sound signal and the mechanical vibration signal; The Fourier transform of the k-th frame of the multi-channel audio signal; The Fourier transform of the k-th frame of the mechanical vibration signal. for The complex conjugate of ; K is the total number of frames; f is the frequency; Step S223: Calculate the frequency domain coherence coefficient at each frequency point within the target frequency band based on the self-power spectral density and the cross-power spectral density. The specific calculation formula is as follows: ， in, It is a frequency domain coherence function; Step S224: Calculate the average frequency domain coherence coefficient of all frequency points within the target frequency band. The specific calculation formula is as follows: ， in, This represents the average frequency domain coherence coefficient within the target frequency band. This is the set of effective frequency points after filtering.

4. The method for identifying pump faults in a multi-source environment of a pumping station according to claim 1, characterized in that, In step S23 The specific steps for source separation of the denoised time-domain signal frame using the semi-blind source extraction algorithm are as follows: Construct the covariance matrix of the denoised time-domain signal frame, and the cross-covariance vector between the denoised time-domain signal frame and the mechanical vibration signal; Construct a generalized eigenvalue equation, solve for the eigenvector corresponding to the largest eigenvalue of the generalized eigenvalue equation, and obtain the optimal solution mixing vector; The optimal unmixing vector is multiplied by the denoised time-domain signal frame to obtain the candidate sound source signal; The specific steps for source separation of the denoised time-domain signal frame using the FastICA blind separation algorithm are as follows: The denoised time-domain signal frame is preprocessed with whitening to obtain a whitened signal matrix, the specific formula of which is: ， in, This is the signal matrix after whitening; This is the whitening matrix; This is the time-domain signal matrix after noise reduction; The unmixing matrix is ​​solved using a fixed-point iterative algorithm based on maximizing negative entropy. The specific solution formula is as follows: ， ， ， ， in, This is the unmixed vector after iterative updates; The unmixing vector; It is a nonlinear function; After iterative convergence, the unmixed matrix is ​​multiplied by the whitened signal matrix to obtain the candidate sound source signal.

5. The method for identifying pump faults in a multi-sound-source environment of a pumping station according to claim 1, characterized in that, Step S4 includes: Step S41: Extract features from the target water pump's operating sound signal to obtain an acoustic feature vector, which includes Mel frequency cepstral coefficient features, wavelet packet energy features, and spectral features. Step S42: During the normal operation of the water pump, the target water pump is divided into multiple operating condition intervals according to the real-time operating condition parameters. Normal operation sound data of each operating condition interval is collected, and a dynamic baseline model corresponding to each operating condition interval is constructed. Step S43: During the online monitoring phase, select the corresponding optimal dynamic baseline model based on the current real-time operating parameters, and input the acoustic feature vector into the optimal dynamic baseline model for anomaly detection processing to obtain the fault detection result.

6. The method for identifying pump faults in a multi-sound-source environment of a pumping station according to claim 1, characterized in that, In step S5, a deep learning model is used to classify the fault types of the target water pump's operating sound signal to obtain the acoustic event types of the target water pump. The acoustic event types include normal state, cavitation fault, bearing wear fault, and impeller abnormal fault. The deep learning model is a 1D-CNN model, and the parameters of the 1D-CNN model are quantized using INT8.

7. A pump fault identification system for multi-sound-source environments in pump rooms, characterized in that, A method for identifying pump faults in a multi-sound-source environment of a pumping station as described in any one of claims 1-6, comprising: The signal acquisition module is used to acquire multimodal data in the pump room and perform time synchronization processing on the multimodal data; An adaptive sound source separation module, connected to the signal acquisition module, is used to process the multi-channel sound signal using a hybrid sound source adaptive separation strategy based on the quality assessment results of the mechanical vibration signal, to obtain multiple candidate sound source signals; A multimodal confidence fusion module, connected to the adaptive sound source separation module, is used to perform multimodal confidence fusion verification on each candidate sound source signal and identify the target water pump operating sound signal from multiple candidate sound source signals; An anomaly detection module, connected to the multimodal confidence fusion module, is used to perform anomaly detection processing on the target water pump's operating sound signal to obtain fault detection results; The fault classification module, connected to the anomaly detection module, is used to classify the target water pump's operating sound signal according to the fault detection results, and obtain the fault diagnosis results.

8. A pump fault identification system for multi-sound-source environments in pump rooms according to claim 7, characterized in that, Also includes: A multi-threaded pipeline scheduling module, connected to the signal acquisition module, the adaptive sound source separation module, the multimodal confidence fusion module, the anomaly detection module, and the fault classification module, is used to control the operation of the system. The multi-threaded pipeline scheduling module includes processing threads, acquisition threads, and communication threads with decreasing priority. The model inference acceleration module is connected to the fault classification module and is used to accelerate the inference process of the deep learning model based on the inference framework.

Images