Health condition diagnosis method, system, device and medium for centrifugal pump

Abstract

This invention provides a method, system, device, and medium for diagnosing the health status of a centrifugal pump. The method includes: acquiring acoustic signals from the centrifugal pump during operation; wherein the acoustic signals characterize the operating state of the centrifugal pump; inputting the acoustic signals into a denoising model, and denoising the acoustic signals through a back-diffusion process to obtain a denoised signal; wherein the denoising model is obtained through global parameter optimization using the Hannibal Bacca optimization algorithm; extracting the time-domain features, frequency-domain features, and multi-scale slope entropy features of the denoised signal; inputting the time-domain features, frequency-domain features, and multi-scale slope entropy features into a health status diagnosis model to generate a health status diagnosis result for the centrifugal pump; wherein the health status diagnosis model is obtained through global parameter optimization using the Hannibal Bacca optimization algorithm. This application achieves efficient and accurate diagnosis of the operating status of a centrifugal pump.

Description

Technical Field

[0001] This application relates to the field of intelligent fault diagnosis technology, and in particular to a method, system, device and medium for diagnosing the health status of a centrifugal pump. Background Technology

[0002] Centrifugal pumps, as core power equipment in fluid transport systems, are widely used in critical fields such as petrochemicals, aerospace, military, and civilian applications. Due to prolonged operation in harsh environments including high temperature, high pressure, corrosive media, and variable operating conditions, centrifugal pumps are highly susceptible to cavitation, rotor misalignment, loose foundation bolts, and bearing wear. Failure to promptly and accurately identify these early signs of failure can lead to anything from decreased equipment performance and increased energy consumption to severe vibrations, seal failures, and even catastrophic safety accidents. Therefore, researching high-precision and robust centrifugal pump fault diagnosis technologies has significant engineering application value and economic importance.

[0003] Currently, fault diagnosis methods based on acoustic signal analysis are the mainstream technology. However, in practical industrial applications, existing feature extraction methods are mostly limited to a single domain. For example, simple time-domain statistical indicators (such as mean and variance) are not sensitive to early, subtle faults. Simple frequency-domain analysis (such as Fast Fourier Transform (FFT)) is difficult to distinguish faults with similar spectral structures (for example, high-frequency broadband noise caused by cavitation and high-order harmonic aliasing caused by mechanical loosening often exhibit similar energy distributions in the power spectrum). Traditional single-scale entropy features (such as sample entropy and permutation entropy) ignore the dynamic evolution of signals at different time scales and cannot comprehensively characterize nonlinear fault mechanisms, resulting in a single feature representation of hybrid coupling faults. Therefore, there is a need to provide a method, system, equipment, and medium for diagnosing the health status of centrifugal pumps. Summary of the Invention

[0004] This invention provides a method, system, device, and medium for diagnosing the health status of centrifugal pumps, in order to solve the technical problem that existing centrifugal pump acoustic signal fault diagnosis methods suffer from low accuracy due to the limited number of extracted features.

[0005] This invention provides a method for diagnosing the health status of a centrifugal pump. The method includes: acquiring acoustic signals from the centrifugal pump during operation; wherein the acoustic signals characterize the operating status of the centrifugal pump; inputting the acoustic signals into a denoising model and denoising the acoustic signals through a back-diffusion process to obtain a denoised signal; wherein the denoising model is obtained by global parameter optimization using the Hannibal Bacca optimization algorithm; extracting time-domain features, frequency-domain features, and multi-scale slope entropy features from the denoised signal; inputting the time-domain features, frequency-domain features, and multi-scale slope entropy features into a health status diagnosis model to generate a health status diagnosis result for the centrifugal pump; wherein the health status diagnosis model is obtained by global parameter optimization using the Hannibal Bacca optimization algorithm.

[0006] In one embodiment of the present invention, the denoising model is a denoising diffusion probability model. The step of inputting the acoustic signal into the denoising model and denoising the acoustic signal through a back-diffusion process to obtain a denoised signal includes: inputting the acoustic signal into the denoising diffusion probability model for a preset number of denoising processes; for each denoising process: inputting the diffused signal and the current number of denoising processes into the denoising network of the denoising diffusion probability model to generate the predicted noise corresponding to the number of back-diffusion steps; wherein, the diffused signal during the first denoising process is the acoustic signal; updating the diffused signal through back-diffusion based on the predicted noise to obtain a new diffused signal; determining whether the number of denoising processes has reached a preset denoising number threshold: if yes, then the diffused signal is used as the denoised signal; otherwise, the next denoising process is performed until the denoising number threshold is reached.

[0007] In one embodiment of the present invention, the denoising diffusion probability model is obtained by parameter optimization using the Hannibal Barca optimization algorithm and self-supervised training based on pre-acquired acoustic signal samples. The training steps of the denoising diffusion probability model include: superimposing two independent random noises on the pre-acquired acoustic signal samples to generate two noisy signal samples; using one noisy signal sample as the input sample and the other noisy signal sample as the target sample; initializing and saving a number of preset DDPM candidate parameter combinations; performing parameter optimization on the denoising diffusion probability model for a preset first number of optimizations based on the Hannibal Barca optimization algorithm; and performing the following steps in each parameter optimization: for each DD... PM candidate parameter combinations: Construct and save the corresponding denoising diffusion probability model based on the DDPM candidate parameter combination. Perform self-supervised training on the denoising diffusion probability model for a preset number of training rounds based on the input sample and the target sample, and determine and save the fitness of the DDPM candidate parameter combination based on the training results. Based on the Hannibal Barca optimization algorithm, generate and save new DDPM candidate parameter combinations according to the fitness of each DDPM candidate parameter combination. Determine whether the number of parameter optimization attempts has reached the preset first parameter optimization attempt threshold: If yes, determine the corresponding denoising diffusion probability model based on the DDPM candidate parameter combination with the best fitness; otherwise, continue to perform the next parameter optimization.

[0008] In one embodiment of the present invention, for each round of training, the step of performing self-supervised training on the denoising diffusion probability model based on the input sample and the target sample includes: randomly generating a diffusion step number within a preset diffusion step number range, and inputting the input sample and the diffusion step number into the denoising diffusion probability model to generate corresponding prediction noise; wherein the upper limit of the diffusion step number range corresponds to the denoising number threshold; and updating the parameters of the denoising diffusion probability model based on the difference between the prediction noise and the target sample.

[0009] In one embodiment of the present invention, each candidate parameter combination includes an initial learning rate, a maximum number of training rounds, the number of hidden layer nodes in the denoising diffusion probability model, and the network depth of the denoising diffusion probability model.

[0010] In one embodiment of the present invention, the health status diagnosis model is a wide learning system. The steps of inputting time-domain features, frequency-domain features, and multi-scale slope entropy features into the health status diagnosis model to generate the health status diagnosis result of the centrifugal pump include: concatenating the time-domain features, frequency-domain features, and multi-scale slope entropy features according to a preset concatenation order to generate a combined feature characterizing the operating state of the centrifugal pump; inputting the combined feature into the feature mapping layer of the wide learning system, projecting the comprehensive feature onto a high-dimensional feature space to generate a mapped feature; inputting the mapped feature into the enhancement layer of the wide learning system, performing a nonlinear transformation on the mapped feature to generate an enhanced feature; wherein the number of nodes in the feature mapping layer and the number of nodes in the enhancement layer are obtained by global parameter optimization based on the Hannibal Baca optimization algorithm; concatenating the mapped feature and the enhanced feature, and inputting the combined feature formed by the concatenation into the output layer of the wide learning system to generate the health status diagnosis result of the centrifugal pump.

[0011] In one embodiment of the present invention, the health status diagnosis model is trained based on pre-acquired comprehensive features and corresponding health status labels. The comprehensive features are formed by concatenating time-domain features, frequency-domain features, and multi-scale slope entropy features. The training steps of the health status diagnosis model include: initializing a number of preset BLS candidate parameter combinations; performing parameter optimization on the health status diagnosis model for a preset second optimization number based on the Hannibal Bacca optimization algorithm; and performing the following steps during each parameter optimization: for each BLS candidate parameter combination: constructing and saving the corresponding width learning system based on the BLS candidate parameter combination. The comprehensive features are input into the width learning system, and the width learning system is trained for a preset number of training rounds using the ridge regression method. The fitness of the BLS candidate parameter combination is determined and saved based on the training results. Based on the Hannibal Barca optimization algorithm, new BLS candidate parameter combinations are generated and saved according to their fitness. It is then determined whether the number of parameter optimization attempts has reached a preset second parameter optimization attempt threshold: if so, the width learning system is used as the health status diagnosis model based on the BLS candidate parameter combination with the best fitness; otherwise, the next parameter optimization attempt is performed.

[0012] This invention also provides a health status diagnosis system for a centrifugal pump. The system includes: a signal acquisition module for acquiring acoustic signals during the operation of the centrifugal pump; wherein the acoustic signals are used to characterize the operating status of the centrifugal pump; a denoising module for inputting the acoustic signals into a denoising model and denoising the acoustic signals through a back-diffusion process to obtain a denoised signal; wherein the denoising model is obtained by global parameter optimization using the Hannibal Bacca optimization algorithm; a multi-domain feature extraction module for extracting time-domain features, frequency-domain features, and multi-scale slope entropy features of the denoised signal; and a diagnosis module for inputting the time-domain features, frequency-domain features, and multi-scale slope entropy features into the health status diagnosis model to generate a health status diagnosis result for the centrifugal pump; wherein the health status diagnosis model is obtained by global parameter optimization using the Hannibal Bacca optimization algorithm.

[0013] The present invention also provides an electronic device, comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device enables the centrifugal pump health status diagnosis method described above.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a computer processor, causes the computer to perform any of the above-described methods for diagnosing the health status of a centrifugal pump.

[0015] The beneficial effects of this invention are as follows: The present invention proposes a method, system, device, and medium for diagnosing the health status of a centrifugal pump. By acquiring the acoustic signal during the operation of the centrifugal pump and using a denoising model obtained through global parameter optimization using the Hannibal Bacca optimization algorithm to perform back-diffusion denoising on the acoustic signal, it can effectively suppress background noise interference in the industrial environment and improve signal quality. By extracting the time-domain features, frequency-domain features, and multi-scale slope entropy features of the denoised signal, a multi-dimensional feature representation of the centrifugal pump's operating status is achieved, improving the distinguishability between different fault states. Furthermore, inputting the above-mentioned multi-domain features into the health status diagnosis model, which is also obtained through global parameter optimization using the Hannibal Bacca optimization algorithm, for classification and recognition can improve the recognition accuracy and generalization ability of the diagnosis model, achieving efficient and accurate diagnosis of the centrifugal pump's operating status. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0017] In the attached diagram:

[0018] Figure 1 A flowchart illustrating a method for diagnosing the health status of a centrifugal pump according to an embodiment of the present invention;

[0019] Figure 2 This is a schematic diagram of the network structure of a width learning system provided in an embodiment of the present invention;

[0020] Figure 3a A time-domain waveform diagram of a noisy signal provided in an embodiment of the present invention;

[0021] Figure 3b for Figure 3a Spectrum diagram;

[0022] Figure 3c This is a time-domain waveform diagram of a denoised signal provided in an embodiment of the present invention;

[0023] Figure 3d for Figure 3c Spectrum diagram;

[0024] Figure 4a The time-domain waveform of the original acoustic signal under normal operating conditions is provided in an embodiment of the present invention.

[0025] Figure 4b This is a time-domain waveform of a normal signal after HBO-DDPM noise reduction, provided in an embodiment of the present invention.

[0026] Figure 4c for Figure 4a The spectrum of the original acoustic signal under normal operating conditions;

[0027] Figure 4d for Figure 4b The spectrum of the original acoustic signal under normal operating conditions;

[0028] Figure 5 This is the result of the fault diagnosis confusion matrix of a health status diagnosis model provided in an embodiment of the present invention on a test set;

[0029] Figure 6 The result of the fault diagnosis confusion matrix of traditional decision trees on the test set;

[0030] Figure 7 This is the result of the fault diagnosis confusion matrix of the traditional KNN algorithm on the test set;

[0031] Figure 8 This is a structural block diagram of a centrifugal pump health status diagnosis system provided in one embodiment of the present invention;

[0032] Figure 9 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0033] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0034] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0035] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0036] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0037] Research has revealed that, in addition to the aforementioned problems, existing centrifugal pump fault diagnosis technologies face the following severe challenges in practical industrial applications: First, the difficulty of extracting weak features in strong noise environments. Industrial environments are extremely complex, and the collected centrifugal pump acoustic signals are often submerged in fluid turbulence noise, environmental background noise, and electromagnetic interference. Traditional signal denoising methods have significant limitations: wavelet thresholding relies excessively on the selection of basis functions and threshold functions, making it difficult to adapt to non-stationary signals; while Empirical Mode Decomposition (EMD) and its variants (such as Time-Varying Filter Empirical Mode Decomposition (TVFEMD)) possess adaptability, they are prone to mode aliasing and endpoint effects, and key parameters (such as bandwidth threshold and B-spline order) are usually set manually based on experience, lacking adaptability. Although deep generative models (such as generative adversarial networks and diffusion models) that have emerged in recent years have performed excellently in image denoising, in industrial one-dimensional signal processing, most existing denoising models fall under the category of supervised learning, requiring a large amount of "clean-noisy" paired data for training. However, in real-world industrial scenarios, it is almost impossible to obtain absolutely pure, ideal sound signals as labels, which limits the practical application of deep noise reduction technology. Secondly, there is a contradiction between the training efficiency and accuracy of diagnostic models. With the development of deep learning, models such as convolutional neural networks and Transformers are widely used. While these have improved diagnostic accuracy, their deep network layers and massive parameter counts (often in the millions) make training extremely time-consuming and demanding on hardware computing power, making it difficult to meet the real-time diagnostic needs of industrial IoT edge devices.

[0038] To address the aforementioned issues, this invention provides a method for diagnosing the health status of centrifugal pumps. By acquiring the acoustic signals generated during the pump's operation and utilizing a denoising model obtained through global parameter optimization using the Hannibal Barca optimization algorithm, back-diffusion denoising is applied to the acoustic signals. This effectively suppresses background noise interference in industrial environments and improves signal quality. By extracting the time-domain, frequency-domain, and multi-scale slope entropy features of the denoised signal, a multi-dimensional feature representation of the centrifugal pump's operating status is achieved, enhancing the distinguishability between different fault states. Furthermore, inputting these multi-domain features into a health status diagnosis model, also obtained through global parameter optimization using the Hannibal Barca optimization algorithm, for classification and identification improves the model's recognition accuracy and generalization ability, enabling efficient and accurate diagnosis of the centrifugal pump's operating status.

[0039] like Figure 1 As shown, the health status diagnosis method for centrifugal pumps includes the following steps:

[0040] S100. Acquire the acoustic signal of the centrifugal pump during operation; wherein the acoustic signal is used to characterize the operating status of the centrifugal pump.

[0041] Acoustic sensors installed at preset locations on the centrifugal pump can acquire the acoustic signals generated during its operation. For example, high-sensitivity piezoelectric accelerometers are installed on the pump casing and bearing housing to continuously collect the vibration and acoustic radiation signals generated during pump operation at different operating states. Specifically, a higher sampling frequency can be set to sample the acoustic signals during centrifugal pump operation in real time, for example, setting the sampling frequency to 51200Hz, i.e., collecting 51200 sample values ​​per second, to obtain the original acoustic signals that reflect changes in the centrifugal pump's operating state. It is understood that the acoustic signal in this application is an acoustic signal sequence X formed by discretely sampling a continuous acoustic signal at a preset sampling frequency; the acoustic signal sequence consists of multiple sample values ​​arranged sequentially.

[0042] Furthermore, in order to eliminate the influence of different acquisition conditions or differences in sensor sensitivity, after acquiring the original sound signal sequence, for the i-th sample value in the original sound signal sequence X... Amplitude normalization is performed, and the normalized sample values ​​are arranged sequentially to obtain the normalized sound signal. For example, the max-min normalization method can be used to normalize the original sample values... The amplitude is mapped to a preset range (e.g.) or To eliminate the influence of dimensions, normalized sample values ​​are obtained. As shown in formula (1):

[0043] (1)

[0044] in, For the normalized i-th sample value, Let be the i-th sample value in the original acoustic signal sequence X. The minimum sample value in the original acoustic signal sequence X. The maximum sample value in the original acoustic signal sequence X.

[0045] Furthermore, to facilitate subsequent signal processing, the normalized acoustic signal can be segmented. Specifically, a sliding window method can be used to segment the normalized acoustic signal, dividing a long acoustic signal sequence into multiple acoustic signal segments of preset lengths, with a preset overlap region between adjacent segments. Each acoustic signal segment is then used as the final acoustic signal for subsequent processing. For example, the length of each acoustic signal segment can be set to... Each sample value corresponds to a signal duration of 0.05 seconds, which facilitates subsequent rapid signal denoising, feature extraction, and health status diagnosis.

[0046] S200. Input the acoustic signal into the denoising model and denoise the acoustic signal through a reverse diffusion process to obtain a denoised signal; wherein, the denoising model is obtained by global parameter optimization using the Hannibal Barca optimization algorithm.

[0047] Because acoustic signals collected in industrial settings are typically affected by various factors such as fluid turbulence noise, ambient background noise, and electromagnetic interference, they often contain a large amount of noise components unrelated to the fault. Therefore, processing the acoustic signals using a denoising model can effectively reduce noise interference and enhance the effective information reflecting the centrifugal pump's operating status, resulting in a denoised signal with a high signal-to-noise ratio. Furthermore, using the Hannibal Barca optimization algorithm to globally optimize the parameters of the denoising model can reduce the need for manual parameter setting based on experience, thereby improving the rationality of parameter selection and enhancing the denoising performance of the model.

[0048] In an optional embodiment of the present invention, the denoising model is a denoising diffusion probabilistic model (DDPM). Step S200 includes the following process: inputting the acoustic signal into the denoising diffusion probabilistic model for a preset number of denoising processes; for each denoising process: inputting the diffusion signal and the current number of denoising processes into the denoising network of the denoising diffusion probabilistic model to generate the predicted noise corresponding to the number of back diffusion steps; wherein, the diffusion signal during the first denoising process is the acoustic signal; updating the diffusion signal through back diffusion according to the predicted noise to obtain a new diffusion signal; determining whether the number of denoising processes has reached a preset denoising number threshold: if so, then using the diffusion signal as the denoised signal; otherwise, continuing the next denoising process until the denoising number threshold is reached.

[0049] Specifically, in each denoising process, the current diffused signal and the corresponding number of denoising processes are input into the denoising network of the denoising diffusion probability model. The denoising network estimates the noise components in the signal to obtain the corresponding predicted noise. In the first denoising process, the diffused signal is an acoustic signal. The diffused signal is updated using backdiffusion based on the predicted noise to gradually reduce the noise components while retaining effective information reflecting the centrifugal pump's operating status, thus obtaining a new diffused signal. After one denoising process is completed, it is determined whether the number of denoising processes has reached a preset threshold. If not, the updated diffused signal is used as the input for the next denoising process, and the noise estimation and backdiffusion update process is repeated to gradually suppress random noise in the signal. Conversely, when the number of denoising processes reaches the preset threshold, the final diffused signal is used as the denoised signal.

[0050] It should be noted that in the denoising diffusion probability model, each denoising process corresponds to a time step in the back diffusion process, that is, the number of denoising processes corresponds one-to-one with the back diffusion time steps. As the back diffusion time step decreases, the signal gradually completes the denoising process, which is shown in formula (2):

[0051] (2)

[0052] in, For the current denoising diffusion probability model parameters Under the conditions, based on the current diffusion signal The updated diffusion signal was inferred. The conditional probability distribution, This represents the diffused signal corresponding to the t-th reverse diffusion time step, i.e., the diffused signal in the current denoising step. Based on The diffused signal obtained after one backdiffusion update, where t is the backdiffusion time step, corresponding to the number of denoising processes. ( The distribution is Gaussian. Given a diffusion signal After the current diffusion time step t, the parameters of the denoising diffusion probability model are: At that time, the predicted mean, The parameters of the current diffusion probability model are: When the predicted covariance is obtained, it represents the given spread signal. After the current diffusion time step t, the denoising diffusion probability model for Uncertainty.

[0053] In an optional embodiment of the present invention, the denoising diffusion probability model is obtained by parameter optimization using the Hannibal Barca optimization algorithm and self-supervised training based on pre-acquired acoustic signal samples. The training steps of the denoising diffusion probability model include: superimposing two independent random noises on the pre-acquired acoustic signal samples to generate two noisy signal samples; using one noisy signal sample as the input sample and the other noisy signal sample as the target sample; initializing and saving a number of preset DDPM candidate parameter combinations; performing parameter optimization on the denoising diffusion probability model for a preset first optimization number of times based on the Hannibal Barca optimization algorithm; and performing the following steps during each parameter optimization: for each DDPM candidate parameter combination: constructing and saving the corresponding denoising diffusion probability model based on the DDPM candidate parameter combination; performing self-supervised training on the denoising diffusion probability model for a preset number of training rounds based on the input sample and the target sample; and determining and saving the fitness of the DDPM candidate parameter combination based on the training results. Based on the Hannibal Barca optimization algorithm, new DDPM candidate parameter combinations are generated and saved according to their fitness. It is then determined whether the number of parameter optimization attempts has reached the preset first parameter optimization attempt threshold. If so, the corresponding denoising diffusion probability model is determined based on the DDPM candidate parameter combination with the best fitness. Otherwise, the next parameter optimization attempt is performed.

[0054] Specifically, in order to achieve effective noise reduction of the centrifugal pump acoustic signal in the absence of clean labeled data, the denoising diffusion probability model is trained using a self-supervised training method. Specifically, two independent random noises are superimposed on the pre-collected acoustic signal samples to generate two noisy signal samples. One of the noisy signal samples is used as the input sample, and the other as the target sample, thus constructing a training sample pair. The above training sample pair can be expressed as shown in formula (3):

[0055] (3)

[0056] in, For potentially pure sound signals, and The noise is obtained through independent sampling. Using the above method, training samples can be constructed without obtaining clean acoustic signal labels, enabling the denoising diffusion probability model to learn effective structural information and noise distribution characteristics in the acoustic signal during training.

[0057] In the denoising diffusion probability model, the training process gradually introduces random noise into the data through a forward diffusion mechanism, causing the signal to gradually transform into a Gaussian noise state. Specifically, from the acoustic signal... Initially, Gaussian noise is added to the signal at each diffusion time step to obtain isotropic white Gaussian noise. Suppose that the noise intensity sequence satisfies ,make , T is the preset maximum diffusion time step. The forward diffusion process can then be expressed as shown in formula (4):

[0058] (4)

[0059] in, The Gaussian white noise corresponding to the diffusion time step t, ( The distribution is Gaussian. It is the identity matrix. The cumulative noise scheduling factor is the sum of the initial signal and the diffusion time step t. Given an initial signal Under the condition of t-step diffusion, the diffusion signal obtained The conditional probability distribution.

[0060] Based on the above diffusion process, the denoising diffusion probability model is trained under self-supervised supervision to predict the noise added during the diffusion process. The training loss function can be expressed as shown in formula (5):

[0061] (5)

[0062] in, In the denoising diffusion probability model, the predicted difference between noise and the target sample is used. For acoustic signal samples Random noise and And the expectation of the joint component at diffusion time step t, The noise variable is obtained by random sampling. ( ) is the predicted noise obtained by the denoising network. These are the parameters for the denoising diffusion probability model. This represents the preset noise scheduling coefficient during the diffusion process. By minimizing the aforementioned difference, the denoising diffusion probability model gradually learns the structural information and noise distribution characteristics in the acoustic signal, thereby achieving effective noise reduction of the centrifugal pump acoustic signal without requiring clean labeled data.

[0063] By minimizing the above loss function, the denoising diffusion probability model gradually learns the structural information and noise distribution characteristics in the acoustic signal, thereby achieving effective noise reduction of the centrifugal pump acoustic signal without the need for clean label data.

[0064] Furthermore, to obtain a superior denoising diffusion probability model, this application utilizes the Hannibal Barca optimization algorithm to globally optimize the model parameters. During parameter optimization, multiple DDPM candidate parameter combinations are first initialized. In each round of parameter optimization, a corresponding denoising diffusion probability model is constructed based on each DDPM candidate parameter combination and subjected to self-supervised training to obtain the corresponding fitness value. Based on the fitness of each DDPM candidate parameter combination, the Hannibal Barca optimization algorithm is used to update the DDPM candidate parameter combinations, thereby generating new DDPM candidate parameter combinations. Through multiple rounds of parameter optimization iterations, when the number of parameter optimizations reaches a preset threshold, the parameter combination with the optimal fitness is determined from all DDPM candidate parameter combinations, and the final denoising diffusion probability model is determined accordingly.

[0065] Since the performance of the denoising diffusion probability model is closely related to the network structure and training parameters, in this embodiment, each DDPM candidate parameter combination includes the initial learning rate, the maximum number of training epochs, the number of hidden layer nodes in the denoising diffusion probability model, and the network depth.

[0066] The following is a detailed explanation of the Hannibal Barca optimization algorithm: Each DDPM candidate parameter combination is represented as Y = (lr, Epochs, ,Layers), where lr is the initial learning rate of the denoising diffusion probability model, and Epochs is the number of training epochs of the model. Let represent the number of hidden layer nodes in the denoising diffusion probability model, and Layers represent the network depth. During parameter optimization, the corresponding denoising diffusion probability model is first trained based on each DDPM candidate parameter combination, and the fitness value is calculated based on the model training error. Individuals in the population are ranked according to their fitness, and divided into Roman army individuals with higher fitness and Carthaginian army individuals with lower fitness. The algorithm mainly includes an encirclement attack phase and a strategic retreat phase. In the encirclement attack phase, the positions of DDPM candidate parameter combinations are updated through multi-directional attacks, thereby achieving a search of the parameter space. Specifically, this includes three update methods: right-wing attack, left-wing attack, and center attack.

[0067] During the right-wing attack, the candidate parameter combination of DDPM is updated by referencing the center of the plane, and the update formula is shown in formula (6):

[0068] (6)

[0069] in, For the updated DDPM candidate parameter combinations, For the current DDPM candidate parameter combinations, The reference plane center position for all current DDPM candidate parameter combinations. It is a random vector.

[0070] During the left-wing attack, the candidate parameter combination of DDPM is updated by mirror search, and the update formula is shown in formula (7):

[0071] (7)

[0072] in, The location center of the candidate parameter combination for the opposing group DDPM.

[0073] During the central attack process, the candidate parameter combinations of DDPM are guided to approach the current optimal solution, and the update formula is shown in formula (8):

[0074] (8)

[0075] in, This represents the current globally optimal combination of candidate parameters for DDPM. This represents the current locally optimal candidate parameter combination for DDPM.

[0076] In the parameter optimization iteration process, in order to avoid the algorithm getting stuck in local optima, a random walk is performed on some DDPM candidate parameter combinations in the middle of the iteration through a strategic retreat phase. The update formula is shown in formula (9):

[0077] (9)

[0078] in, and These represent the upper and lower bounds of the values ​​of each parameter in the DDPM candidate parameter combination, respectively. Through the above update mechanism, the Hannibal Barca optimization algorithm can continuously search and update the DDPM candidate parameter combination in the parameter space, thereby obtaining the optimal parameter combination suitable for the denoising diffusion probability model.

[0079] Furthermore, in an optional embodiment of the present invention, the step of performing self-supervised training on the denoising diffusion probability model based on the input sample and the target sample for each round of training includes: randomly generating a diffusion step number within a preset diffusion step number range, and inputting the input sample and the diffusion step number into the denoising diffusion probability model to generate corresponding prediction noise; wherein the upper limit of the diffusion step number range corresponds to the denoising number threshold; and updating the parameters of the denoising diffusion probability model based on the difference between the prediction noise and the target sample.

[0080] Specifically, after each round of parameter optimization using the Hannibal Barca optimization algorithm, a corresponding denoising diffusion probability model is constructed based on the current DDPM candidate parameter combination, and the model is self-supervised trained according to the preset number of training rounds. During the training process, diffusion time steps within the preset diffusion step range are randomly generated, and the input samples and diffusion time steps are input together into the denoising diffusion probability model to generate the corresponding prediction noise. The difference between the prediction noise and the target sample is calculated according to formula (5), and the model parameters are updated based on the difference, thereby completing the training process of this round.

[0081] Furthermore, in each round of parameter optimization, the denoising diffusion probability model corresponding to each DDPM candidate parameter combination is trained for a short period. The training error of the model is calculated based on data from the validation set, such as by using Mean Squared Error (MSE) to evaluate the difference between the model output and the target sample. This training error is then used as the fitness value of the corresponding DDPM candidate parameter combination for subsequent parameter updates. Through multiple rounds of parameter optimization iterations, when the number of optimizations reaches a preset threshold, the parameter combination with the optimal fitness is determined from all DDPM candidate parameter combinations, and the final denoising diffusion probability model is determined accordingly. The obtained optimal model is then used to perform sliding window denoising on the original acoustic signal to obtain a denoised signal for subsequent feature extraction and health status diagnosis.

[0082] As can be seen, the above-mentioned denoising and self-supervised training methods are referred to as the HBO-DDPM method. This application innovatively combines a diffusion model with a self-supervised denoising strategy. Compared with traditional supervised learning denoising models, it does not require the prior acquisition of a clean "standard signal" and can train a high-performance denoiser using only the noisy data itself, greatly reducing the cost and difficulty of industrial data annotation.

[0083] S300 extracts the time-domain features, frequency-domain features, and multi-scale slope entropy features of the denoised signal.

[0084] To fully characterize the acoustic signal variations of a centrifugal pump under different operating conditions, time-domain, frequency-domain, and multi-scale slope entropy features were extracted from the denoised signal. Specifically, in the time domain, statistical analysis was performed on the amplitude sequence of the denoised signal to extract time-domain statistical features reflecting the overall amplitude distribution and impact characteristics of the signal. In the frequency domain, the denoised signal underwent spectral transformation processing, such as obtaining the signal's spectral information through Fast Fourier Transform, and frequency-domain feature parameters were extracted based on the spectral distribution to reflect the frequency distribution characteristics of the acoustic signal energy under different fault conditions. Furthermore, multi-scale slope entropy calculation was performed on the denoised signal to reflect the nonlinear dynamic characteristics of the centrifugal pump's acoustic signal.

[0085] Specifically, to comprehensively characterize the dynamic characteristics of centrifugal pumps under different faults, especially to distinguish between cavitation and loosening faults with similar spectral characteristics, in one specific embodiment, 25 dimensions of features were extracted from the acoustic signal X across three domains: time domain, frequency domain, and nonlinear dynamics. The time domain features reflect the amplitude distribution characteristics and energy changes of the signal. Specifically, the mean value is extracted from the denoised signal. Standard deviation Root mean square value Peak skewness , cliff Peak factor and waveform factor There are a total of 8 features. Among them, kurtosis can be expressed as shown in formula (10):

[0086] (10)

[0087] in, Let N be the amplitude of the i-th sampling point, and N be the number of signal sampling points. Kurtosis is highly sensitive to impact signals and can effectively characterize the mechanical impact caused by bolt loosening. The peak factor and waveform factor are shown in formulas (11) and (12), respectively:

[0088] (11)

[0089] (12)

[0090] The aforementioned time-domain features can reflect the energy changes and impact characteristics of the centrifugal pump's acoustic signal from a time dimension. Furthermore, considering the difference between cavitation (broadband high-frequency noise) and misalignment (harmonic energy), a frequency domain analysis was performed on the acoustic signal, and five-dimensional frequency domain features were extracted, including the centroid frequency FC, mean square frequency MSF, frequency variance RMSF, spectral flatness SF, and the main frequency amplitude (Max Amp). Among these, the centroid frequency increases significantly with the increase in high-frequency components during cavitation, as shown in formula (13):

[0091] (13)

[0092] in, P(k) is the frequency value corresponding to the k-th frequency component after frequency domain processing of the acoustic signal, and P(k) is the power spectral density value of this frequency component. In addition, the mean square frequency is used to describe the rate of change of spectral energy. Spectral flatness is used to distinguish whether the signal is noise-like (such as cavitation, in which case the SF value is high but the spectrum is flat) or harmonic-like (such as misalignment, in which case the SF value is low but the peak is obvious), as shown in formula (14):

[0093] (14)

[0094] Where K is the number of spectrum sampling points. In order to further represent the nonlinear dynamic characteristics of the centrifugal pump acoustic signal, this embodiment also introduces multi-scale slope entropy (MSlpEn) as a nonlinear feature. Multi-scale slope entropy combines multi-scale analysis and slope entropy, and can effectively quantify the complexity of the signal at different time scales. Specifically, the denoised signal is subjected to multi-scale coarse-grained processing. For scale s (values ​​from 1 to 12), the denoised signal is divided into non-overlapping windows of length s, and the mean of each window is calculated, so as to obtain the coarse-grained sequence as shown in formula (15). :

[0095] (15)

[0096] in, Let be the i-th sample value in the denoised signal, s be the current scale, and j be the j-th sample value in the coarse-grained sequence. Window number, Let be the mean of the j-th window at scale s. After obtaining the coarse-grained sequence, set the embedding dimension. and threshold Calculate the difference between adjacent elements in the coarse-grained sequence, and apply this to the sequence. Perform the difference operation as shown in formula (16):

[0097] (16)

[0098] in, This represents the difference between the k-th element and the (k+1)-th element in a coarse-grained sequence at scale s. and These are two adjacent elements in the coarse-grained sequence at scale s. Based on the difference... The magnitude of the sequence is symbolically represented to indicate the trend of sequence change, and a set of symbols is assigned to it. , respectively representing significant increase, slight increase, stable, slight decrease and significant decrease, and their symbolization rules are shown in formula (17):

[0099] (17)

[0100] in, The difference The corresponding symbol encoding result, A preset threshold is used to distinguish the amplitude of signal changes. Continuous... A pattern composed of symbols Statistically analyze the probability distribution of each symbol pattern in the sequence. And calculate the corresponding Shannon entropy to obtain the Shannon entropy at this scale, as shown in formula (18):

[0101] (18)

[0102] in, Let x be the multiscale slope entropy value of signal x at scale s. Symbolic pattern The probability of occurrence in a symbol sequence.

[0103] By scale s=1 Calculating the entropy value within a range of 12 yields a 12-dimensional multi-scale slope entropy feature. Combining the time-domain, frequency-domain, and multi-scale slope entropy features creates a 25-dimensional feature vector for centrifugal pump health status diagnosis.

[0104] It should be noted that existing technologies often struggle to distinguish between cavitation (high-frequency broadband noise) and mechanical loosening (nonlinear impact). This application introduces frequency-domain centroid frequency, spectral flatness, and multi-scale slope entropy. Centroid frequency and spectral flatness can keenly capture the high-frequency energy distribution characteristics of cavitation signals, while multi-scale slope entropy can deeply mine the nonlinear dynamic abrupt changes of loosening faults at multiple scales. The complementary use of multi-domain features significantly improves the clarity of classification boundaries. The problem of confusion between similar faults is solved through global feature fusion.

[0105] S400. Input the time-domain features, frequency-domain features, and multi-scale slope entropy features into the health status diagnosis model to generate the health status diagnosis results of the centrifugal pump; wherein, the health status diagnosis model is obtained by global parameter optimization through the Hannibal-Bacca optimization algorithm.

[0106] In an optional embodiment of the present invention, the health status diagnosis model is a wide learning system, and step S400 includes the following processes: concatenating time-domain features, frequency-domain features, and multi-scale slope entropy features according to a preset concatenation order to generate combined features characterizing the operating status of the centrifugal pump; inputting the combined features into the feature mapping layer of the wide learning system, projecting the comprehensive features into a high-dimensional feature space to generate mapped features; inputting the mapped features into the enhancement layer of the wide learning system, performing a nonlinear transformation on the mapped features to generate enhanced features; wherein the number of nodes in the feature mapping layer and the number of nodes in the enhancement layer are obtained by global parameter optimization based on the Hannibal Barca optimization algorithm; concatenating the mapped features and the enhanced features, and inputting the comprehensive features formed after concatenation into the output layer of the wide learning system to generate the health status diagnosis result of the centrifugal pump.

[0107] Furthermore, the health status diagnosis model is trained based on pre-acquired comprehensive features and corresponding health status labels. The comprehensive features are formed by concatenating time-domain features, frequency-domain features, and multi-scale slope entropy features. The training steps of the health status diagnosis model include: initializing several preset BLS candidate parameter combinations; performing parameter optimization on the health status diagnosis model for a preset second optimization number based on the Hannibal Bacca optimization algorithm; and performing the following steps during each parameter optimization: for each BLS candidate parameter combination: constructing and saving the corresponding width learning system based on the BLS candidate parameter combination, inputting the comprehensive features into the width learning system, training the width learning system for a preset number of training rounds using the ridge regression method, and determining and saving the fitness of the BLS candidate parameter combination based on the training results; generating and saving new BLS candidate parameter combinations based on the fitness of each BLS candidate parameter combination using the Hannibal Bacca optimization algorithm; determining whether the parameter optimization number has reached the preset second parameter optimization number threshold: if so, then using the BLS candidate parameter combination with the best fitness as the corresponding width learning system as the health status diagnosis model; otherwise, continuing to perform the next parameter optimization.

[0108] The Broad Learning System (BLS) is an efficient learning algorithm based on the Random Vector Functional-Link Neural Network (RVFLNN). Unlike deep learning models such as convolutional neural networks, BLS does not require continuous backpropagation to update the parameters of the feature extraction layer. Instead, it directly calculates the output layer weights by solving the pseudo-inverse or regularized regression form of the linear system. Its core principle is to enhance the model's expressive power by horizontally expanding "feature nodes" and "booster nodes," thereby approximating complex functional relationships without increasing network depth. Therefore, it has advantages such as simple structure and fast training speed. Its network structure diagram is shown below. Figure 2 As shown. Figure 2 Mid-bottom This represents the input feature data. In this embodiment, the input feature is a combined feature formed by concatenating time-domain features, frequency-domain features, and multi-scale slope entropy features in a preset order. The input feature is first projected onto the feature mapping layer through a mapping function to generate multiple mapped feature nodes, denoted as... , that is, Mapped Feature 1, Mapped Feature 2, … Mapped Feature n in the figure. Its calculation form is shown in formula (19):

[0109] (19)

[0110] in, Represents the combined feature matrix. The feature mapping layer is represented by the first... The mapping weights of each node, This indicates the corresponding bias term. Represents a non-linear activation function. Indicates the first There are 10 feature nodes. These mapped feature nodes are further input to the enhancement layer, where a nonlinear transformation generates enhancement nodes, denoted as . The nodes are labeled as EnhancementNodes in the diagram. The calculation method for enhancement nodes is shown in formula (20):

[0111] (20)

[0112] in, This represents the feature node matrix obtained from the feature mapping layer. Indicates the enhancement layer number 1 The mapping weights corresponding to each node This indicates the corresponding bias term. Represents a non-linear activation function. Indicates the generated first Enhanced nodes. Further, the feature node matrix... With enhanced node matrix The features are concatenated to obtain the comprehensive feature matrix. The comprehensive feature matrix is ​​input into the output layer of the width learning system to obtain the health status diagnosis results of the centrifugal pump. ,in, Indicates the diagnosis results of health status. This represents the output layer weight matrix. The output layer weights are analytically solved using the ridge regression method, and their calculation method is shown in formula (21):

[0113] (twenty one)

[0114] in, It is the identity matrix. This is the regularization coefficient.

[0115] It should be noted that although wide learning systems have advantages such as flat structure, fast training speed, and strong incremental learning ability, their generalization performance largely depends on the setting of key parameters such as the number of nodes in the feature mapping layer, the number of nodes in the enhancement layer, and the regularization coefficient. Existing methods usually rely on manual experience to adjust parameters, lacking an effective global optimization mechanism, which can easily lead to the model getting stuck in local optima or overfitting. Therefore, in this application, the Hannibal Bacca optimization algorithm is introduced to globally optimize the key parameters of the wide learning system to obtain a better model structure. Specifically, the number of feature nodes... Increase the number of nodes Regularization coefficient These three key parameters are optimized. The fitness function of the BLS parameter combination is shown in formula (22):

[0116] (twenty two)

[0117] in, This represents the model's classification accuracy on the test set. The Hannibal Barca optimization algorithm performs a global search in the parameter space by simulating search strategies such as encirclement, attack, and strategic retreat, thereby quickly obtaining the parameter combination that minimizes the test error and improving the recognition accuracy and generalization ability of the centrifugal pump health status diagnosis model. It is understood that the specific implementation of the Hannibal Barca optimization algorithm in this embodiment is the same as that of the aforementioned denoising diffusion probability model, and will not be elaborated upon here.

[0118] It should be noted that the data used in training and validating the model in this application comes from a centrifugal pump comprehensive fault simulation test bench. This test bench is a closed-loop system that can simulate typical fault states of centrifugal pumps under different operating conditions under controllable conditions, thus providing an experimental data basis for research on acoustic signal acquisition and fault diagnosis of centrifugal pumps.

[0119] Specifically, this embodiment uses a single-stage cantilever centrifugal pump as the test object, which mainly consists of three parts: a three-phase asynchronous motor, a coupling, and a chemical process pump. The three-phase asynchronous motor serves as the drive unit, and the coupling acts as the power transmission unit, transmitting the power output from the motor to the centrifugal pump rotor, thereby driving the impeller inside the pump to rotate. The main technical parameters of this centrifugal pump are: rated flow rate... Rated head Rated speed For acoustic signal acquisition, high-precision acoustic sensors were selected and installed near the bearings at both the suspended and unsuspended ends of the pump body to monitor the acoustic radiation signals generated during the operation of the centrifugal pump. Before the experiment, the acoustic sensors were calibrated for sensitivity using a standard acoustic calibrator to ensure the accuracy and consistency of the acquired data. During the experiment, a mobile workstation equipped with LMS Test Lab professional software was used for real-time visual monitoring and data recording of the acquired signals. Auxiliary equipment for the experimental system also included a non-contact photoelectric tachometer, an adjustable sensor bracket, and low-noise shielded cables to achieve speed measurement, stable sensor installation, and reduction of external electromagnetic interference.

[0120] To fully verify the effectiveness of the method of this invention, this application designed and simulated four typical centrifugal pump operating states, including normal state, cavitation failure, loose anchor bolt failure, and shaft misalignment failure, as detailed below:

[0121] (1) Normal state: The centrifugal pump operates stably under the design conditions, with the inlet and outlet valves fully open, the foundation bolts tightened, and the shaft system properly aligned.

[0122] (2) Cavitation: The pump inlet pressure is artificially reduced until it is lower than the saturated vapor pressure of the liquid being pumped by adjusting the vacuum pump on the inlet pipe of the centrifugal pump or adjusting the opening of the inlet valve. As bubbles are generated and collapse, fluid vibration is induced. The degree of cavitation is confirmed by observing bubbles in the transparent pipe section and monitoring the inlet and outlet pressures in the experiment.

[0123] (3) Bolt Loosening: Use a torque wrench to control the preload of the bolts to 50% of the rated value to simulate the working condition of loose bolts. Loosening will cause the pump body stiffness to decrease, causing vibration at a specific frequency.

[0124] (4) Misalignment: The position of the motor shaft is adjusted by using a laser alignment instrument to simulate the misalignment fault.

[0125] After signal denoising and feature extraction, the resulting 25-dimensional global feature dataset was randomly divided into training and testing sets in an 8:2 ratio. The experimental sample size was sufficient, with at least 800 samples for each operating condition. During data labeling, the four operating conditions were assigned labels 1, 2, 3, and 4 in the order of "normal," "cavitation," "loose bolts," and "misalignment," respectively.

[0126] During the model optimization process, the parameters of the Hannibal Barca optimization algorithm are set as follows: population size N=20, maximum number of iterations T=30.

[0127] Furthermore, in the noise reduction performance verification experiment, a simulated signal was generated as the original signal, and Gaussian white noise was superimposed on the simulated signal to construct a noisy signal. To evaluate the signal noise reduction effect, signal-to-noise ratio (SNR) and mean square error (MSE) were introduced as performance evaluation indicators to quantitatively measure the quality of signal recovery after noise reduction. Figure 3a As shown, it displays the time-domain waveform of the noisy signal. It can be seen that after adding Gaussian white noise, the original signal is submerged by a large amount of random noise, and the time-domain waveform exhibits obvious random fluctuations. Figure 3b The spectrum of the noisy signal shows a significant broadband noise component. After processing by the method of this application, the following result is obtained: Figure 3c The denoised waveform is shown below. It can be seen that the denoised time-domain signal waveform is smoother and its periodic characteristics are clearer. (See diagram below.) Figure 3d As shown, the main characteristic frequency peaks in the denoised spectrum are more prominent, the background noise is significantly reduced, and the signal-to-noise ratio is significantly improved, indicating that the proposed denoising method can effectively recover the effective feature information in the signal. Figure 3a and Figure 3c This is a time-domain waveform diagram, where the horizontal axis represents time (in seconds) and the vertical axis represents signal amplitude (in Pa). Figure 3b and Figure 3d The graph shows the spectrum, with the horizontal axis representing frequency (Hz) and the vertical axis representing amplitude (Pa). The proposed method was used to denoise the noisy signal. The signal-to-noise ratio and mean square error of the signal before and after denoising are shown in Table 1 below.

[0128] Table 1. Comparison of Signal-to-Noise Ratio and Mean Square Error Before and After Noise Reduction

[0129]

[0130] Based on the changes in the time-frequency plots of the analog signals before and after noise reduction, as well as the calculations of the signal-to-noise ratio and mean square error, the effectiveness of the noise reduction method can be demonstrated. Furthermore, noise reduction processing is applied to the acquired acoustic signals. Taking an acoustic signal acquired under healthy conditions as an example, its time-domain and frequency-domain plots before and after noise reduction are shown below. Figures 4a to 4d As shown. Among them, Figure 4a The time-domain waveform of the original sound signal under normal operating conditions is shown. It can be seen that there are certain random fluctuations and noise interference in the signal, and the time-domain waveform is relatively messy. Figure 4b The time-domain waveform of the normal signal after HBO-DDPM noise reduction is shown. It can be seen that, compared with... Figure 4a In comparison, the signal waveform after noise reduction is smoother, and the random noise component is significantly reduced. Figure 4cThe spectrum of the original sound signal under normal operating conditions is shown. It can be seen that there are many noise components in a wide frequency range and the spectrum distribution is relatively scattered. Figure 4b The signal spectrum after noise reduction using the HBO-DDPM method is shown. It can be seen that, compared to... Figure 4c Compared to the original signal, the main characteristic frequencies are more prominent, high-frequency noise components are significantly reduced, and the spectral structure is clearer. Therefore, the signal after noise reduction is cleaner and has a smoother spectrum than the original signal. Consequently, the extracted features are more accurate, the influence of interference is removed, and this has a positive impact on subsequent model building.

[0131] Figure 5 The results show the fault diagnosis confusion matrix of the health status diagnosis model in this application on the test set. It can be seen that the overall recognition accuracy of the health status diagnosis model based on global feature fusion is 97.50%. Specifically, 160 samples from both Class 1 and Class 4 were correctly identified, with a recognition rate of 100.0%; 148 samples from Class 2 were correctly identified, and 11 were misclassified as Class 3, corresponding to a recognition rate of 93.1%; 155 samples from Class 3 were correctly identified, and 5 were misclassified as Class 2, with a recognition rate of 96.9%. From the prediction results, the prediction accuracy for Class 1 and Class 4 is 100.0%, while the prediction accuracy for Class 2 and Class 3 is 96.7% and 93.4%, respectively, with misclassification rates of 3.3% and 6.6%, respectively. Overall, the results indicate that the health status diagnosis model optimized using the Hannibal Barca algorithm can accurately distinguish different centrifugal pump operating states and has high fault identification accuracy.

[0132] Figure 6 This is the confusion matrix result for fault diagnosis on the test set using a traditional decision tree. From... Figure 6 As can be seen, the overall recognition accuracy of the decision tree model is 93.3%. Among them, 159 samples of Class 1 were correctly identified, and 1 was misclassified as Class 3; 145 samples of Class 2 were correctly identified, 14 were misclassified as Class 3, and 1 was misclassified as Class 4; 142 samples of Class 3 were correctly identified, 1 was misclassified as Class 1, 14 were misclassified as Class 2, and 3 were misclassified as Class 4; and 151 samples of Class 4 were correctly identified, 2 were misclassified as Class 1, 1 was misclassified as Class 2, and 6 were misclassified as Class 3.

[0133] Figure 7 This is the confusion matrix result for fault diagnosis on the test set using the traditional KNN algorithm. Figure 7As can be seen, the overall recognition accuracy of the KNN model is 94.7%. Among them, 159 samples of Class 1 were correctly identified and 1 was misclassified as Class 3; 149 samples of Class 2 were correctly identified and 11 were misclassified as Class 3; 145 samples of Class 3 were correctly identified and 15 were misclassified as Class 2; and 153 samples of Class 4 were correctly identified, 2 were misclassified as Class 1, 2 were misclassified as Class 2, and 3 were misclassified as Class 3.

[0134] The classification accuracy of the three models on the test set was compared, and the results are shown in Table 2:

[0135] Table 2. Comparison of classification accuracy of the three models

[0136]

[0137] It can be seen that the health status diagnosis model (denoted as HBO-BLS model) proposed in this invention, which optimizes the parameters of the wide learning system based on the Hannibal Battle optimization algorithm, has a classification accuracy that is about 4% higher than that of decision trees and about 3% higher than that of KNN, demonstrating its superior classification accuracy.

[0138] In summary, this invention proposes a high-precision fault diagnosis method that integrates a self-supervised denoising model optimized by the Hannibal Bacca optimization algorithm, deep fusion of multi-dimensional features across the entire domain, and classification of a health status diagnostic model optimized by the Hannibal Bacca optimization algorithm. Specifically, the system achieves automated parameter optimization through a dual Hannibal Bacca optimization mechanism. This involves simultaneously optimizing the parameters of both the front-end denoising model and the back-end health status diagnostic model using the Hannibal Bacca optimization algorithm, thereby avoiding the subjectivity and blindness associated with manually adjusting parameters such as the number of deep network layers, learning rate, and the number of nodes in the health status diagnostic model. This allows the system to automatically search for optimal parameter combinations under different operating conditions, improving the accuracy and generalization ability of the diagnostic model. Experimental results show that the fault diagnosis accuracy of the method proposed in this invention can reach over 98.75%.

[0139] Furthermore, this invention employs a width learning system as the health status diagnosis model. Compared to deep convolutional neural networks or Transformer models, the width learning system enhances the model's expressive power by horizontally expanding the number of nodes and directly solves for the output weights using a pseudo-inverse matrix, eliminating the need for complex and time-consuming backpropagation iterations. Therefore, while ensuring diagnostic accuracy, it significantly improves the training and inference speed of the model, making it more suitable for industrial online monitoring scenarios. In summary, the method of this invention can achieve high-fidelity reconstruction of strongly noisy signals even in the absence of clean labeled data, effectively improves the distinguishing ability between similar faults through multi-domain complementary features, and achieves efficient and accurate construction of the fault diagnosis model by combining a dual optimization strategy.

[0140] like Figure 8 As shown, the centrifugal pump health status diagnosis system includes: a signal acquisition module 810, a denoising module 820, a multi-domain feature extraction module 830, and a diagnosis module 840. The signal acquisition module 810 acquires the acoustic signal during the centrifugal pump's operation; the acoustic signal characterizes the centrifugal pump's operating status. The denoising module 820 inputs the acoustic signal into a denoising model and performs denoising processing through a back-diffusion process to obtain a denoised signal; the denoising model is obtained through global parameter optimization using the Hannibal Bacca optimization algorithm. The multi-domain feature extraction module 830 extracts the time-domain features, frequency-domain features, and multi-scale slope entropy features of the denoised signal. The diagnosis module 840 inputs the time-domain features, frequency-domain features, and multi-scale slope entropy features into the health status diagnosis model to generate the centrifugal pump's health status diagnosis result; the health status diagnosis model is obtained through global parameter optimization using the Hannibal Bacca optimization algorithm.

[0141] Specific limitations regarding the centrifugal pump health status diagnosis system can be found in the limitations of the centrifugal pump health status diagnosis method described above, and will not be repeated here. Each module in the aforementioned centrifugal pump health status diagnosis system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware format within or independently of the processor in a computer device, or stored in software format in the memory of a computer device, so that the processor can call the corresponding operations of each module.

[0142] It should be noted that, in order to highlight the innovative aspects of this invention, this embodiment does not include modules that are not closely related to solving the technical problems proposed by this invention, but this does not mean that there are no other modules in this embodiment.

[0143] like Figure 9 As shown, the electronic device 9 may include a memory 91, a processor 92, and a bus, and may also include a computer program stored in the memory 91 and executable on the processor 92, such as a health status diagnostic program for a centrifugal pump.

[0144] The memory 91 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 91 can be an internal storage unit of the electronic device 9, such as a portable hard drive. In other embodiments, the memory 91 can be an external storage device of the electronic device 9, such as a plug-in portable hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 9. Furthermore, the memory 91 can include both internal and external storage units of the electronic device 9. The memory 91 can be used not only to store application software and various types of data installed on the electronic device 9, such as health status diagnostic codes for centrifugal pumps, but also to temporarily store data that has been output or will be output.

[0145] In some embodiments, the processor 92 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 92 is the control unit of the electronic device 9, connecting various components of the entire electronic device 9 through various interfaces and lines. It executes programs or modules stored in the memory 91 (such as a health status diagnostic program for a centrifugal pump) and calls data stored in the memory 91 to perform various functions and process data in the electronic device 9.

[0146] The processor 92 executes the operating system of the electronic device 9 and various installed application programs. The processor 92 executes the application programs to implement the steps in the above-described centrifugal pump health status diagnosis method.

[0147] For example, a computer program can be divided into one or more modules, one or more of which are stored in memory 91 and executed by processor 92 to complete this application. One or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in electronic device 9. For example, the computer program can be divided into a signal acquisition module 810, a noise reduction module 820, a multi-domain feature extraction module 830, and a diagnostic module 840.

[0148] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium, which can be non-volatile or volatile. The software functional module stored in the storage medium includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute some functions of the centrifugal pump health status diagnosis method of the various embodiments of this application.

[0149] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for diagnosing the health status of a centrifugal pump, characterized in that, The health status diagnosis method includes: Acquire acoustic signals during the operation of a centrifugal pump; wherein the acoustic signals are used to characterize the operating state of the centrifugal pump; The acoustic signal is input into a denoising model, and the acoustic signal is denoised through a back-diffusion process to obtain a denoised signal; wherein, the denoising model is obtained by global parameter optimization using the Hannibal Barca optimization algorithm; Extract the time-domain features, frequency-domain features, and multi-scale slope entropy features of the denoised signal; The time-domain features, frequency-domain features, and multi-scale slope entropy features are input into the health status diagnosis model to generate the health status diagnosis result of the centrifugal pump; wherein, the health status diagnosis model is obtained by global parameter optimization using the Hannibal Barca optimization algorithm; The denoising model is a denoising diffusion probability model, which is obtained by optimizing parameters using the Hannibal Barca optimization algorithm and by self-supervised training based on pre-acquired acoustic signal samples. The training steps of the denoising diffusion probability model include: Two independent random noises are superimposed on the pre-acquired acoustic signal samples to generate two noisy signal samples. One of the noisy signal samples is used as the input sample, and the other noisy signal sample is used as the target sample. Several preset DDPM candidate parameter combinations are initialized and saved. The parameters of the denoised diffusion probability model are optimized based on the Hannibal Barca optimization algorithm for a preset first optimization number; The following steps are performed during each parameter optimization: For each DDPM candidate parameter combination: construct and save the corresponding denoising diffusion probability model based on the DDPM candidate parameter combination, perform self-supervised training on the denoising diffusion probability model for a preset number of training rounds based on the input sample and the target sample, and determine and save the fitness of the DDPM candidate parameter combination based on the training results. Based on the Hannibal Barca optimization algorithm, new DDPM candidate parameter combinations are generated and saved according to the fitness of each DDPM candidate parameter combination. Determine whether the number of parameter optimization attempts has reached the preset threshold for the first parameter optimization attempt: If so, then the corresponding denoising diffusion probability model is determined based on the optimal combination of DDPM candidate parameters. Otherwise, continue with the next parameter optimization.

2. The method for diagnosing the health status of a centrifugal pump according to claim 1, characterized in that, The steps of inputting the acoustic signal into a denoising model and denoising the acoustic signal through a back-diffusion process to obtain a denoised signal include: The acoustic signal is input into the denoising diffusion probability model for a preset number of denoising cycles. For each noise reduction process: The diffused signal and the current number of denoising processes are input into the denoising network of the denoising diffusion probability model to generate the predicted noise corresponding to the number of back diffusion steps; wherein, the diffused signal during the first denoising process is the acoustic signal; The spread signal is updated by reverse diffusion based on the predicted noise to obtain a new spread signal; Determine whether the number of denoising operations has reached the preset threshold. If so, the diffused signal will be used as the denoised signal; Otherwise, the denoising process continues until the threshold number of denoising iterations is reached.

3. The method for diagnosing the health status of a centrifugal pump according to claim 1, characterized in that, For each training round, the steps for self-supervised training of the denoising diffusion probability model based on the input and target samples include: A number of diffusion steps within a preset range is randomly generated, and the input sample and the number of diffusion steps are input into the denoising diffusion probability model to generate corresponding predicted noise; wherein the upper limit of the diffusion step range corresponds to the denoising number threshold. The parameters of the denoising diffusion probability model are updated based on the degree of difference between the predicted noise and the target sample.

4. The method for diagnosing the health status of a centrifugal pump according to claim 1, characterized in that, Each candidate parameter combination includes the initial learning rate, the maximum number of training epochs, the number of hidden layer nodes in the denoising diffusion probability model, and the network depth of the denoising diffusion probability model.

5. The method for diagnosing the health status of a centrifugal pump according to claim 1, characterized in that, The health status diagnosis model is a width learning system. The steps for generating the health status diagnosis result of the centrifugal pump by inputting the time-domain features, frequency-domain features, and multi-scale slope entropy features into the health status diagnosis model include: The time-domain features, frequency-domain features, and multi-scale slope entropy features are spliced ​​together according to a preset splicing order to generate a combined feature characterizing the operating state of the centrifugal pump. The combined features are input into the feature mapping layer of the width learning system, and the combined features are projected into a high-dimensional feature space to generate mapped features; The mapped features are input into the enhancement layer of the width learning system, and a nonlinear transformation is performed on the mapped features to generate enhanced features; wherein, the number of nodes in the feature mapping layer and the number of nodes in the enhancement layer are obtained by global parameter optimization based on the Hannibal Barca optimization algorithm; The mapped features and the enhanced features are concatenated, and the resulting comprehensive features are input into the output layer of the width learning system to generate the health status diagnosis result of the centrifugal pump.

6. The method for diagnosing the health status of a centrifugal pump according to claim 1, characterized in that, The health status diagnosis model is trained based on pre-acquired comprehensive features and corresponding health status labels. The comprehensive features are formed by concatenating time-domain features, frequency-domain features, and multi-scale slope entropy features. The training steps of the health status diagnosis model include: Initialize and preset several BLS candidate parameter combinations; The parameters of the health status diagnosis model are optimized based on the Hannibal Barca optimization algorithm for a preset second optimization number. The following steps are performed during the optimization of each parameter: For each BLS candidate parameter combination: construct and save the corresponding width learning system based on the BLS candidate parameter combination, input the comprehensive features into the width learning system, train the width learning system for a preset number of training rounds according to the ridge regression method, and determine and save the fitness of the BLS candidate parameter combination based on the training results; Based on the Hannibal Barca optimization algorithm, new BLS candidate parameter combinations are generated and saved according to the fitness of each BLS candidate parameter combination. Determine whether the number of parameter optimization attempts has reached the preset threshold for the number of second parameter optimization attempts: If so, the corresponding width learning system will be used as the health status diagnosis model based on the optimal combination of BLS candidate parameters. Otherwise, continue with the next parameter optimization.

7. A health status diagnosis system for a centrifugal pump, characterized in that, The system includes: A signal acquisition module is used to acquire acoustic signals of the centrifugal pump during operation; wherein the acoustic signals are used to characterize the operating status of the centrifugal pump. The denoising module is used to input the acoustic signal into the denoising model and perform denoising processing on the acoustic signal through a back diffusion process to obtain a denoised signal; wherein, the denoising model is obtained by global parameter optimization using the Hannibal Barca optimization algorithm; The multi-domain feature extraction module is used to extract the time-domain features, frequency-domain features, and multi-scale slope entropy features of the denoised signal. The diagnostic module is used to input the time-domain features, frequency-domain features, and multi-scale slope entropy features into the health status diagnostic model to generate the health status diagnostic results of the centrifugal pump; wherein, the health status diagnostic model is obtained by global parameter optimization using the Hannibal Barca optimization algorithm; The denoising model is a denoising diffusion probability model, which is obtained by optimizing parameters using the Hannibal Barca optimization algorithm and by self-supervised training based on pre-acquired acoustic signal samples. The training steps of the denoising diffusion probability model include: Two independent random noises are superimposed on the pre-acquired acoustic signal samples to generate two noisy signal samples. One of the noisy signal samples is used as the input sample, and the other noisy signal sample is used as the target sample. Several preset DDPM candidate parameter combinations are initialized and saved. The parameters of the denoised diffusion probability model are optimized based on the Hannibal Barca optimization algorithm for a preset first optimization number; The following steps are performed during each parameter optimization: For each DDPM candidate parameter combination: construct and save the corresponding denoising diffusion probability model based on the DDPM candidate parameter combination, perform self-supervised training on the denoising diffusion probability model for a preset number of training rounds based on the input sample and the target sample, and determine and save the fitness of the DDPM candidate parameter combination based on the training results. Based on the Hannibal Barca optimization algorithm, new DDPM candidate parameter combinations are generated and saved according to the fitness of each DDPM candidate parameter combination. Determine whether the number of parameter optimization attempts has reached the preset threshold for the first parameter optimization attempt: If so, then the corresponding denoising diffusion probability model is determined based on the optimal combination of DDPM candidate parameters. Otherwise, continue with the next parameter optimization.

8. An electronic device, characterized in that, The electronic device includes: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the health status diagnosis method for a centrifugal pump as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by the computer's processor, causes the computer to perform the health status diagnosis method for the centrifugal pump according to any one of claims 1 to 6.

Images