An embedded-based method and device for detecting cavitation sound signals of a hydraulic turbine

Abstract

The application relates to an embedded-based water turbine cavitation sound signal detection method and device, which comprises the following steps: collecting a main detection sound channel signal and a reference sound channel signal, and obtaining a digital sound signal; performing variational mode decomposition on the digital signal to obtain K intrinsic mode functions; performing mode reliability evaluation on the intrinsic mode functions, and performing weighted enhancement on the intrinsic mode functions of the main detection sound channel; constructing a multi-channel input sample according to the enhanced intrinsic mode functions, and performing standardization processing on each channel; inputting the standardized multi-channel input into a pre-trained MSCNN network; performing multi-scale convolution feature extraction, feature fusion and classification mapping on the multi-channel input sample by the MSCNN network to obtain a water turbine cavitation state probability; and performing result display, alarm prompting and data back transmission according to the water turbine cavitation state recognition result. Compared with the prior art, the application has the advantages of convenient detection, strong extraction ability of weak initial cavitation characteristics and high recognition stability.

Description

Technical Field

[0001] This invention relates to the field of turbine condition monitoring, and in particular to a method and apparatus for detecting cavitation acoustic signals in turbines based on embedded systems. Background Technology

[0002] When a hydro turbine operates under unfavorable conditions, low load, high flow velocity, or with significant fluctuations in incoming flow, the local pressure within the flow channel can easily fall below the vaporization pressure of water, inducing cavitation. The generation, development, and collapse of cavitation bubbles can lead to increased noise, intensified vibration, and decreased efficiency, as well as erosion of flow components and structural fatigue. In severe cases, this can affect the safe and stable operation of the unit, shorten equipment service life, and increase maintenance costs. Therefore, accurately monitoring the cavitation status of the hydro turbine is of great significance for improving the operational safety and equipment reliability of hydropower units.

[0003] Traditional cavitation detection methods mainly include shutdown inspection, manual judgment, vibration monitoring, and pressure pulsation analysis. These methods are technically mature and can reflect cavitation status from aspects such as component damage, unit vibration, and flow field changes. However, they suffer from insufficient real-time performance, reliance on human experience, and susceptibility to operating environment influences. Acoustic signals, on the other hand, can directly reflect the transient impacts generated by cavitation bubble collapse. They offer advantages such as non-invasiveness, fast response, high sensitivity, and ease of deployment, and have become an important means of online cavitation monitoring.

[0004] Existing methods for detecting cavitation in hydro-turbines based on acoustic signals mainly include threshold discrimination, frequency domain energy analysis, traditional feature engineering, and deep learning recognition. Threshold discrimination methods determine cavitation status by comparing acoustic signal amplitude, root mean square value, or sound pressure level with preset thresholds. They offer advantages such as simple principles, low computational cost, and ease of real-time implementation. However, the detection results are easily affected by operating conditions, background noise, and sensor installation location, resulting in poor threshold universality. Frequency domain energy analysis methods identify cavitation by analyzing spectral amplitude or energy changes within specific frequency bands, reflecting the frequency distribution characteristics of cavitation noise. However, the acoustic disturbances of initial cavitation are weak, and relevant features are easily masked by water flow noise and mechanical noise, leading to misjudgments or missed detections. Traditional feature engineering methods extract time-domain, frequency-domain, or time-frequency-domain features and combine them with classifiers for recognition, comprehensively utilizing multi-dimensional signal information. However, feature design relies on human experience, limiting adaptability and generalization ability under complex operating conditions. To reduce reliance on manual features, deep learning methods are increasingly being applied to cavitation acoustic signal detection, automatically extracting deep features and improving recognition capabilities. However, most existing studies are based on offline analysis on PCs or host computers, which involves a large number of model parameters and computational loads, making it difficult to deploy directly on resource-constrained DSP embedded platforms.

[0005] A method for identifying cavitation in a hydro turbine, as disclosed in Chinese Patent Publication No. CN119152886A, includes: acquiring an audio signal A(t) and a vibration signal V(t) from a hydro turbine in operation; preprocessing the audio signal A(t) and the vibration signal V(t) to obtain processed audio signals A″(t) and vibration signals V″(t); fusing the preprocessed audio signals A″(t) and the vibration signals V″(t) to obtain a fused signal; performing empirical mode decomposition (EMD) on the fused signal Ffusion(t) to obtain a set of intrinsic mode functions (IMFs) of the fused signal, wherein the IMF set includes several IMFs; extracting the time-frequency features of each IMF; inputting the time-frequency features into a pre-constructed support vector machine for classification to obtain a classification result; and determining whether the hydro turbine is cavitating based on the classification result. This method can improve the accuracy of identifying hydro turbine cavitation based on the time-frequency features of the fused signal. However, the existing patent decomposition method uses EMD, which suffers from mode aliasing and endpoint effects, making it unsuitable for weak cavitation. At the same time, although the existing patent is lightweight, the front-end EMD plus manual feature extraction still relies on a PC environment, which does not meet the on-site portability needs of inspection personnel.

[0006] Therefore, the problems that need to be solved include the weak initial cavitation characteristics, the insufficient stability of traditional methods, and the difficulty of embedding deep learning models. Summary of the Invention

[0007] The purpose of this invention is to overcome the defects of the prior art and provide an embedded method and device for detecting cavitation acoustic signals of water turbines.

[0008] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a method for detecting cavitation acoustic signals of a hydro turbine based on embedded systems includes: Step S1: Collect the main detection acoustic channel signal and the reference acoustic channel signal during the operation of the water turbine, and obtain the digital signal after analog-to-digital conversion; Step S2: The VMD program is implanted into the DSP and lightweighted. The VMD program is used to perform variational mode decomposition on the digital signal according to the preset number of IMFs K and the penalty parameter α to obtain K intrinsic mode functions. The VMD is variational mode decomposition. Step S3: Based on the reference acoustic channel signal, perform modal reliability evaluation on the intrinsic mode function, and perform weighted enhancement on the intrinsic mode function of the main detection acoustic channel according to the evaluation results; Step S4: Construct multi-channel input samples based on the enhanced intrinsic mode functions, and standardize each channel. Step S5: Input the standardized multi-channel input samples into the pre-trained MSCNN network. The MSCNN network performs multi-scale convolutional feature extraction, feature fusion, and classification mapping on the multi-channel input samples to obtain the turbine cavitation state recognition result. Step S6: Display the results, provide alarm prompts, and transmit data back based on the cavitation status identification results of the water turbine.

[0009] As a preferred technical solution, the selection of the number of IMFs K includes: exist Perform VMD within each round to calculate residual energy. : in, Let K ∈ {3,4,5,6,7,8,9,10}; The minimum value is represented by x(n); x(n) represents the nth sampling point of the input digital acoustic signal. The k-th intrinsic mode function is represented by N; N represents the signal length. After normalizing the three indicators, a comprehensive indicator is constructed. in, The average correlation of the IMF; The average envelope entropy of the IMF; make according to Determination of the inflection point of adjacent improvement .

[0010] As a preferred technical solution, the penalty parameter The specific choices include: exist Perform VMD within each round to calculate the average envelope entropy of the IMF. Average correlation and average kurtosis And construct fitness : Let α ∈ {500, 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000, 5500, 6000}; select smallest As a VMD decomposition parameter; Represents the average envelope entropy; Indicates the average correlation of the IMF; Indicates the effective IMF mean kurtosis; This represents the prior value of α; Indicates the prior penalty weight; This represents the minimum value.

[0011] As a preferred technical solution, the VMD program porting to the DSP end and lightweight processing specifically include: The matrix operations in the VMD decomposition process are converted into one-dimensional array loop operations based on C language, and the input time domain signal, the frequency domain representation of the input signal, the frequency domain representation of each intrinsic mode component, the center frequency, the Lagrange multiplier, the cumulative update amount, and the output results of the intrinsic mode components are stored in pre-allocated fixed-length arrays respectively. The system employs a single-precision floating-point data format and pre-sets the input signal length, the number of modal components, the maximum number of iterations, and the storage space of each array, so that dynamic memory allocation and deallocation are not performed during the VMD decomposition process. During the center frequency update process, the center frequency is calculated only based on the spectral energy of the corresponding intrinsic mode components within the frequency update interval, and the update of the Lagrange multiplier constraint terms in the original VMD iteration process is omitted. The number of modal components K, the penalty factor α, the maximum number of iterations, and the convergence threshold are set to predetermined fixed parameters, and these fixed parameters are written into the DSP program.

[0012] As a preferred technical solution, the modal reliability assessment of the intrinsic mode function specifically includes: The first of the main detection sound channels The eigenmode functions are denoted as And select the reference mode whose center frequency is closest from the set of intrinsic mode functions of the reference acoustic channel, denoted as . ; Based on the intrinsic mode function of the main detection acoustic channel With the reference mode The temporal correlation coefficient is determined by considering the temporal correlation, frequency coherence, and energy overlap between the components. Frequency domain coherence coefficient and energy overlap coefficient ; According to the time-domain correlation coefficient Frequency domain coherence coefficient and energy overlap coefficient Determine the first Modal reliability coefficients of each intrinsic mode function: in, For the first Modal reliability coefficients of each intrinsic mode function. , and These are the weighting coefficients, and .

[0013] As a preferred technical solution, the intrinsic mode function is enhanced based on the modal reliability coefficient, specifically including: Based on the modal reliability coefficient Determine the first Enhancement coefficients of each eigenmode function: in, For the first The enhancement coefficients of each intrinsic mode function, These are the modal retention coefficients; According to the enhancement coefficient The first detection channel of the main detection channel eigenmode functions After weighting, the enhanced IMF modes are obtained. ; The reference acoustic channel signal is used only for reliability assessment and modal enhancement of the intrinsic mode functions of the main detection acoustic channel, and does not directly participate in the cavitation state category determination.

[0014] As a preferred technical solution, the input data processing of the MSCNN includes: obtaining the data after modal reliability evaluation and weighted enhancement. Each intrinsic mode function; normalization is performed on each IMF channel: in, To enhance the k-th IMF point n, This represents the mean of the k-th IMF window before enhancement. This represents the standard deviation of the k-th IMF window before enhancement.

[0015] As a preferred technical solution, the MSCNN network includes an input layer, a first-scale convolutional branch, a second-scale convolutional branch, a third-scale convolutional branch, a feature fusion layer, a fully connected layer, and a classification output layer. The input layer is used to receive standardized multi-channel input samples and construct multi-channel time-series input samples according to the arrangement order of the intrinsic mode components; The first-scale convolution branch is used to sequentially perform first-scale one-dimensional convolution processing, nonlinear activation processing and pooling processing on the multi-channel temporal input samples to extract the local transient impact features generated by the collapse of cavitation bubbles and obtain the first-scale feature vector. The second-scale convolution branch is used to sequentially perform second-scale one-dimensional convolution processing, nonlinear activation processing, and pooling processing on the multi-channel temporal input samples to extract local correlation features and local energy change features between multiple consecutive impact events, and obtain the second-scale feature vector. The third-scale convolution branch is used to sequentially perform third-scale one-dimensional convolution processing, nonlinear activation processing, and pooling processing on the multi-channel time-series input samples to extract the signal envelope change features and the overall evolution features of cavitation impact within a preset time range, and obtain the third-scale feature vector. Among them, the kernel size used in the first-scale one-dimensional convolution processing, the second-scale one-dimensional convolution processing and the third-scale one-dimensional convolution processing increases sequentially, so that each convolution branch obtains the cavitation sound signal characteristics in the receptive field at different times. The feature fusion layer is used to perform dimension unification processing on the first-scale feature vector, the second-scale feature vector, and the third-scale feature vector, and then splice the dimension-unified feature vectors of each scale along the feature channel direction to obtain a multi-scale fused feature vector. The fully connected layer is used to perform feature mapping on the multi-scale fused feature vector to obtain a classification feature vector characterizing the cavitation state of the turbine. The classification output layer is used to output the predicted probability of the turbine being in different cavitation states based on the classification feature vector, and to determine the cavitation state with the highest predicted probability as the cavitation identification result. The MSCNN is trained on a PC using training samples labeled with cavitation states. The weight parameters, bias parameters, and fully connected layer parameters of each convolutional layer are exported and fixed to the DSP. The DSP executes MSCNN forward inference based on the VMD program and outputs the turbine cavitation state recognition result.

[0016] According to another aspect of the present invention, an embedded hydro turbine cavitation acoustic signal detection device includes a sensor module, an ADC module, a DSP chip module, a communication module, and a display module. The sensor module is used to collect acoustic signals during the operation of the water turbine and output the collected acoustic signals to the ADC module. The ADC module is used to amplify, filter, and perform analog-to-digital conversion on the acoustic signal of the turbine operation, and output the converted digital signal to the DSP chip module. The DSP chip module is used to sequentially perform VMD decomposition, intrinsic mode function reliability assessment based on reference acoustic channel, intrinsic mode function weighted enhancement, multi-channel sample construction, and MSCNN forward inference on the digital signal to obtain the cavitation state identification result of the turbine. The communication module is used to transmit the detection results, cavitation state probability, alarm status and corresponding time information back to the host computer. The display module is used to display the cavitation state category and the corresponding cavitation state identification result.

[0017] As a preferred technical solution, the device also includes a storage module, an alarm module, and a power module; The storage module is used to store sampling data, VMD processing intermediate data, and model parameters; The alarm module is used to output an alarm prompt when the preset alarm conditions are met in the cavitation state; The power supply module is used to supply power to the sensor module, ADC module, DSP chip module, storage module, communication module, display module and alarm module.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. Significantly improved the ability to extract and recognize weak features of nascent cavitation. To address the challenge of primary cavitation signals being easily masked by water flow noise, this invention innovatively introduces a reference acoustic channel-assisted modal reliability assessment mechanism. By comparing the temporal correlation, frequency coherence, and energy overlap of the intrinsic mode functions (IMFs) of the main detection channel and the reference channel, the reliability of each IMF component is accurately quantified. Based on the reliability coefficient, the main channel IMFs are adaptively weighted and enhanced, effectively suppressing irrelevant background noise interference while strengthening the true cavitation characteristics. This dual-channel collaborative processing strategy greatly enhances the robustness of the algorithm in complex water environments and solves the problem of high false positive rates in traditional single-channel methods.

[0019] 2. For the first time, the VMD-MSCNN deep learning algorithm was implemented in real-time on a resource-constrained DSP handheld device. To address the bottleneck of deploying deep learning models on embedded systems, this invention employs several key optimizations: Algorithm decoupling and solidification: an architecture is adopted where training is performed on a PC and inference is performed on a DSP. The complex VMD parameter optimization and MSCNN network training are completed on the PC, while only the final model parameters and the simplified VMD algorithm are solidified onto the DSP.

[0020] Lightweight Porting of VMD: The VMD implementation in MATLAB, which relies on matrix operations, is rewritten as a loop operation based on a one-dimensional array in C language. It uses single-precision floating-point numbers (float), pre-defines a fixed array space, and eliminates runtime dynamic memory allocation and parameter iterative search. Furthermore, the iteration process is simplified, retaining only the current and previous iteration data for convergence determination, significantly reducing storage usage and computational overhead.

[0021] MSCNN structure optimization: Design a multi-scale parallel convolutional structure and replace fully connected layers with global average pooling (GAP), which significantly reduces the number of parameters and computation while ensuring multi-scale feature extraction capabilities.

[0022] These measures enable complex AI algorithms to run smoothly on low-power DSP chips, with inference latency meeting the requirements for real-time on-site detection, breaking through the limitation of existing deep learning cavitation detection being limited to offline analysis on PCs.

[0023] 3. The detection parameters are adaptive and highly versatile, lowering the barrier to manual operation and maintenance. This invention provides a scientific method for adaptive selection of VMD parameters. This is achieved by constructing a comprehensive index and improvement measure. Determine the optimal number of IMFs K, and then determine the optimal penalty parameter by minimizing the fitness function.

[0024] This data-driven parameter calibration method avoids the drawbacks of relying on human experience in the traditional threshold method, enabling the detection device to maintain superior detection performance on turbines of different models and operating conditions without the need for frequent on-site adjustments by experts, thus improving the ease of use and versatility of the product.

[0025] 4. Enabled portable on-site detection of cavitation status in water turbines, eliminating reliance on PCs and fixed systems. Existing detection methods suffer from poor real-time performance, high deployment costs, and inconvenience in portability. This invention highly integrates sensors, ADCs, DSP chips, storage, display, alarms, and power modules into a single handheld detector.

[0026] Inspection personnel can carry the equipment directly to the turbine site and independently complete the entire process of acoustic signal acquisition, processing, identification, result display, and data transmission. This not only significantly reduces system deployment costs but also meets the mobility requirements of hydropower station inspections and temporary spot checks, filling the gap in portable high-precision cavitation detection equipment.

[0027] 5. Strong anti-interference capability, effectively avoiding false alarms. In addition to the aforementioned signal processing gains, this invention incorporates continuous frame judgment logic and an initial cavitation probability threshold at the decision-making level. Unlike simple logic that triggers an alarm based on a single sample, this invention comprehensively judges the cavitation state by statistically analyzing the recognition results of multiple consecutive frames. This mechanism effectively filters out occasional misjudgments caused by instantaneous noise or signal jitter, ensuring the seriousness and accuracy of alarm prompts and meeting the high reliability requirements of diagnostic systems in industrial settings. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2This is a schematic diagram of the device structure of the present invention; Figure 3 This is a schematic diagram of the MSCNN network structure; Figure 4 This is a schematic diagram of the MSCNN parameter training and DSP deployment process of the present invention; Figure 5 This is a schematic diagram of the VMD decomposition results of the present invention; Figure 6 This is a statistical analysis of the accuracy of cavitation diagnosis on the PC side of this invention. Figure 7 This is the PC-side cavitation diagnosis confusion matrix of the present invention. Detailed Implementation

[0029] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0030] Example 1 like Figure 1 The diagram shown is a flowchart of the present invention. Step S1: Collect the main detection sound channel signal and the reference sound channel signal during the operation of the water turbine, and obtain the digital sound signal after analog-to-digital conversion.

[0031] MATLAB's VMD primarily relies on matrix operations, complex arrays, and built-in FFT / IFFT. To enable its operation in an embedded environment, this embodiment rewrites VMD as a C language function oriented towards a fixed sampling length. Specifically, the input signal length is set to 4096 points, and the signal is mirrored and extended by 2048 points at each end, forming an 8192-point extended signal. The vectorized operations in MATLAB are rewritten as for loops, and the IMF frequency domain variables, center frequency, accumulator, and final IMF output are all defined as fixed-length arrays. The input signal, center frequency, and IMF output are stored using float type to reduce storage usage. The frequency domain signal is obtained through the FFT function, and the frequency domain array is manually filled in C language according to the frequency arrangement method of MATLAB's fftshift syntax.

[0032] In the VMD iteration process of C language, initialization is performed first. The frequency domain representation and center frequency of each IMF are determined, and then the frequency domain IMFs are updated one by one according to their IMF numbers. When updating each IMF, the accumulated result of other IMFs is subtracted from the original frequency domain signal, and then... Frequency domain filtering is performed with the current center frequency; then... The center frequency is updated for the weights. After each iteration, the difference between the old and new IMF frequency domain results is calculated, and the iteration stops when the difference is below the convergence threshold or the number of iterations reaches a preset upper limit. Finally, position recovery and IFFT are performed on the frequency domain results of each IMF, and the corresponding part of the original 4096 points is truncated to obtain the IMF output in the form of an 8×4096 two-dimensional array. This porting method avoids the dependence on the MATLAB runtime environment and allows the VMD output to be directly used as the input of subsequent CNNs.

[0033] Step S2: Using the VMD program, perform variational mode decomposition on the digital signal according to the preset decomposition parameters K and α to obtain K eigenmode functions; first, read a representative cavitation acoustic signal in MATLAB, and take 4096 points from a single channel as parameter evaluation samples, with the basic parameters set as follows: .for The choice makes And fix first =1000. For each candidate Perform VMD to calculate the energy of the reconstructed residual. Average correlation among IMFs and average envelope entropy After normalizing the three indicators, we get Then As a quality indicator for decomposition. Calculated... In adjacent The improvement amount between these values ​​is determined by using an inflection point threshold to identify where the improvement in decomposition quality slows down, thus selecting the optimal value. .

[0034] for The choice makes , , In the optimal Next, for each candidate Perform VMD and calculate the average envelope entropy. IMF average correlation Effective IMF mean kurtosis as well as Prior penalty term. Fitness function is: ; The smaller the value, the lower the IMF envelope entropy, the smaller the intermodal redundancy, and the more prominent the impact characteristics. The deviation from the prior is not too large, therefore we take... The smallest candidate value is used as the final value. .thus, and It is not arbitrarily given by humans, but is determined by residual energy, correlation, envelope entropy, kurtosis and prior constraints.

[0035] Step S3: Based on the reference acoustic channel signal, perform modal reliability assessment on the intrinsic mode function, and perform weighted enhancement on the intrinsic mode function of the main detection acoustic channel according to the assessment results.

[0036] To improve the expressive power of effective modal features in the primary cavitation acoustic signal, this embodiment introduces a reference acoustic channel signal for modal reliability assessment after VMD decomposition. Specifically, under the same turbine operating conditions, the main detection acoustic channel signal and the reference acoustic channel signal are collected respectively, and VMD decomposition is performed on each to obtain the corresponding intrinsic mode function set.

[0037] The first of the main detection sound channels The eigenmode functions are denoted as The reference mode closest to its center frequency is selected from the set of eigenmode functions of the reference acoustic channel, denoted as . .according to and The temporal correlation coefficient is determined by considering the temporal correlation, frequency coherence, and energy overlap between the components. Frequency domain coherence coefficient and energy overlap coefficient And calculate the modal reliability coefficient. ,in , and Let be the weighting coefficients, and satisfy: Subsequently, the modal reliability coefficient is used to determine the first... Enhancement coefficients of each eigenmode function .

[0038] in, These are modality preservation coefficients, used to prevent low-reliability modes from being completely eliminated. The enhanced intrinsic modal functions of the main detection channel are weighted according to these enhancement coefficients to obtain the enhanced intrinsic modal functions. The reference acoustic channel signal is used only for reliability assessment and modal enhancement of the intrinsic mode functions of the main detection acoustic channel, and does not directly participate in the cavitation state category determination.

[0039] Step S4: Construct multi-channel input samples based on the enhanced intrinsic mode functions, and standardize each channel.

[0040] Step S5: Input the standardized multi-channel input to the pre-trained MSCNN network. The MSCNN network performs multi-scale convolutional feature extraction, feature fusion, and classification mapping on the multi-channel input samples to obtain the cavitation state probability of the turbine.

[0041] First, VMD was performed on the cavitation sound signals under each operating condition to obtain... Each IMF segment is then divided into 500-point windows, with a construction size of [size missing]. The sequence samples were analyzed, and standardization was performed on each IMF channel. Sample labels were categorized into four types based on cavitation state: no cavitation, nascent cavitation, developing cavitation, and severe cavitation. To avoid class imbalance, random downsampling was performed on categories with a larger number of samples, or the samples were balanced according to the smallest category. The training, validation, and test sets were then divided in an 8:1:1 ratio.

[0042] Step S6: Display the results, provide alarm prompts, and transmit data back based on the cavitation status identification results of the water turbine.

[0043] To verify the effectiveness of the VMD-MSCNN cavitation acoustic signal detection method described in this invention, acoustic signals under different cavitation states were collected on a model turbine experimental platform, and model simulation tests were completed on a PC. The experimental conditions were: guide vane opening a0 = 23 mm, unit rotational speed n11 = 72.65 rpm, and sampling frequency Fs = 40.96 kHz. The main parameters of the model turbine were: model runner diameter D1 = 0.36192 m, D2 = 0.38562 m, 15 model runner blades, 12 fixed guide vanes, and 24 movable guide vanes.

[0044] The experimental data were categorized into four types based on cavitation state: non-cavitation, initial cavitation, cavitation development, and severe cavitation. During the PC simulation, VMD decomposition was first performed on the acquired acoustic signals to obtain IMF multi-channel components; then, MSCNN input samples were constructed according to a fixed window, and each IMF channel was standardized; finally, the samples were input into the trained MSCNN for classification testing.

[0045] Example 2 like Figure 2 The diagram shown is a schematic of the device structure of the present invention, which includes a sensor module, an ADC module, a DSP chip module, a storage module, a communication module, a display module, an alarm module, and a power supply module.

[0046] The sensor module is used to collect acoustic signals during the operation of the water turbine and output the collected acoustic signals to the ADC module. The ADC module is used to amplify, filter, and perform analog-to-digital conversion on the acoustic signal of the turbine operation, and output the converted digital signal to the DSP chip module. The DSP chip module is used to sequentially perform VMD decomposition, intrinsic mode function reliability assessment based on reference acoustic channel, intrinsic mode function weighted enhancement, multi-channel sample construction, and MSCNN forward inference on the digital signal to obtain the cavitation state identification result of the turbine. The storage module is used to store sampling data, VMD processing intermediate data, and model parameters; The communication module is used to transmit the detection results, cavitation state probability, alarm status and corresponding time information back to the host computer. The display module is used to display the cavitation state category and the corresponding cavitation state probability; The alarm module is used to output an alarm prompt when the preset alarm conditions are met in the cavitation state; The power supply module is used to supply power to the sensor module, ADC module, DSP chip module, storage module, communication module, display module and alarm module.

[0047] Figure 3 This is a schematic diagram of the MSCNN network architecture; it includes an input layer, three parallel one-dimensional convolutional branches, a feature fusion layer, and a classification output layer. The input layer receives IMF multi-channel window samples constructed and normalized by VMD decomposition, with a typical input size of [size missing]. Three parallel branches are used to extract cavitation acoustic signal features at different time scales. Each branch includes one-dimensional convolution, batch normalization, ReLU, one-dimensional convolution, batch normalization, and ReLU in sequence. Different branches can use convolution kernels of different lengths to extract local impact features, mesoscale variation features, and trend features over a longer time range, respectively. The outputs of the three branches are deeply concatenated to form a fused feature, which is then sequentially processed through one-dimensional global average pooling, fully connected layers, ReLU, dropout, and Softmax to finally output the cavitation state classification result. This structure enhances the ability to express the weak acoustic features of nascent cavitation through multi-scale parallel convolution, while using one-dimensional global average pooling to reduce the number of parameters in the subsequent fully connected layers, facilitating forward inference deployment on the DSP. During training, the Adam optimizer is used, with an initial learning rate of 1e-3, a MiniBatchSize of 32, and a MaxEpochs of 50. The training process is monitored using a validation set. After training, the classification accuracy is calculated using a test set and a confusion matrix is ​​generated to verify the model's ability to distinguish between different cavitation states.

[0048] like Figure 5 As shown, VMD decomposes the acquired cavitation acoustic signal into 8 IMF components. Each IMF component exhibits oscillation characteristics at different time scales under normalized amplitude, providing a basis for the subsequent construction of MSCNN multi-channel input samples.

[0049] Figure 6 As shown, the overall diagnostic accuracy of the method of the present invention on the test set is 89.67%, of which the accuracy rates corresponding to the four states of no cavitation, initial cavitation, cavitation development and severe cavitation are 93.83%, 93.08%, 92.75% and 99.67%, respectively.

[0050] like Figure 7 As shown, the test set contains 1200 samples, of which 1076 are correctly classified and 124 are misclassified. The confusion matrix results indicate that all four states—non-cavitation, initial cavitation, cavitation development, and severe cavitation—can be effectively identified, demonstrating that VMD-MSCNN can effectively extract multi-scale features from cavitation acoustic signals and distinguish between different cavitation states. The above PC simulation results verify the effectiveness of the detection algorithm of this invention and can provide a model foundation for subsequent embedded deployment on DSP platforms. Figure 4 As shown, after training, network parameters such as convolutional layer weights, biases, batch normalization parameters, and fully connected layer parameters are exported. On the DSP side, the network parameters exported from the PC are converted into C language arrays or parameter files that can be called by the DSP and then fixed to the DSP project. During on-site detection, the detector collects the turbine acoustic signal, and the DSP performs VMD decomposition on the collected signal, constructs and normalizes the input window, and then calls the fixed MSCNN forward inference program to output the cavitation detection results. This process achieves the separation of PC-side training and DSP-side real-time inference. The DSP side only performs VMD decomposition, sample construction, and MSCNN forward computation, without performing backpropagation training, thus meeting the real-time and portability requirements of the embedded handheld detector.

[0051] like Figure 4 As shown, after training, network parameters such as convolutional layer weights, biases, batch normalization parameters, and fully connected layer parameters are exported. On the DSP side, the network parameters exported from the PC are converted into C language arrays or parameter files that can be called by the DSP and then fixed to the DSP project. During on-site detection, the detector collects the turbine acoustic signal, and the DSP performs VMD decomposition on the collected signal, constructs and normalizes the input window, and then calls the fixed MSCNN forward inference program to output the cavitation detection results. This process achieves the separation of PC-side training and DSP-side real-time inference. The DSP side only performs VMD decomposition, sample construction, and MSCNN forward computation, without performing backpropagation training, thus meeting the real-time and portability requirements of the embedded handheld detector.

[0052] The DSP embedded system can perform the following functions: 1) acquire and cache cavitation acoustic signals from a water turbine; 2) perform VMD decomposition and IMF output locally; 3) construct a CNN input window and perform normalization; 4) call fixed MSCNN parameters to complete cavitation state inference; 5) display the current cavitation state, category probability, or acoustic signal waveform on an LCD; 6) transmit the detection results, raw acoustic signals, or intermediate VMD results back to the host computer via UART, USB, or other communication methods; and 7) output alarm or prompt information when a preset cavitation level is reached. Therefore, this invention can achieve on-site, portable, and local detection of water turbine cavitation acoustic signals without relying on a PC or cloud server.

[0053] Furthermore, in the acoustic signal acquisition stage, the acoustic signal acquisition module is installed at the test location on the turbine or test bench to obtain a mixed acoustic signal containing cavitation impact, background water flow noise, and mechanical operating noise. The signal conditioning and A / D module first performs impedance matching, amplification, and anti-aliasing filtering on the sensor output, and then converts it into a digital signal according to a preset sampling rate. The converted digital signal enters the buffer of the DSP processing module, which is updated continuously at a fixed length. When the buffer data reaches 4096 points, the DSP processing module triggers a complete VMD-MSCNN detection process; when 4096 points have not been reached, the system continues to acquire data and waits for the next batch of sampling points. This processing method enables the system to continuously process the on-site acoustic signals in a segmented manner.

[0054] Furthermore, in the C language VMD porting, to adapt to the limited storage space and real-time requirements of the DSP, this embodiment avoids dynamically allocating large-scale matrices and prioritizes using static arrays to store intermediate results. For variables such as uhat, sumuk, and ω, only the results of the previous iteration are stored.

[0055] Furthermore, the mirror extension steps of VMD in C language include: taking the first 2048 points of the original 4096-point signal and copying them in reverse to the beginning of the extended signal; placing the original 4096-point signal in the middle of the extended signal; and then taking the last 2048 points of the original signal and copying them in reverse to the end of the extended signal. Mirror extension can reduce the endpoint effect in the VMD decomposition process, allowing the IMF to maintain good continuity at both ends of the signal. After the extension is completed, FFT is performed on the 8192-point signal, and the frequency domain results are stored according to the centered frequency coordinates, which facilitates the subsequent calculation of the filter denominator by ω-ωk. The iterative update does not directly call the matrix expression in MATLAB, but is decomposed into a scalar loop suitable for DSP execution.

[0056] Furthermore, in the VMD result recovery stage, the centered frequency-domain IMF is repositioned to the frequency arrangement required by the IFFT, and negative frequency components are supplemented using conjugate symmetry. Subsequently, an IFFT is performed on each IMF to obtain the time-domain IMF at the extended length. To align with the original segmented signal, this embodiment extracts 4096 points from the middle of the extended result corresponding to the original signal, forming a K×4096 IMF matrix. This matrix can be saved as text or cached data for debugging, or the process can directly proceed to the 500-point window construction step.

[0057] Furthermore, in the MATLAB parameter selection stage, the choice of K focuses on addressing the issues of over-decomposition and under-decomposition. If K is too small, cavitation impact components from different frequency bands may overlap in the same IMF, resulting in insufficient resolvability of the CNN input channel; if K is too large, noise is excessively split, increasing the correlation between IMFs and increasing the computational load on the DSP. Therefore, this embodiment simultaneously considers residual energy, average IMF correlation, and average envelope entropy. Residual energy reflects the ability of VMD to reconstruct the original signal, average IMF correlation reflects the degree of redundancy between modes, and average envelope entropy reflects the complexity of the IMF envelope distribution. Through normalized comprehensive evaluation and inflection point judgment, a balance can be achieved between decomposition quality and computational complexity.

[0058] Furthermore, the selection of α focuses on addressing the IMF bandwidth constraint problem. Too small an α will result in an excessively wide IMF bandwidth, which is detrimental to separating cavitation impact features; too large an α may make the IMF too narrow, causing useful features to be dispersed or weakened. This embodiment uses average envelope entropy, average IMF correlation, effective IMF average kurtosis, and α prior penalty to jointly construct the fitness. A lower average envelope entropy indicates a more concentrated IMF envelope, a lower average correlation indicates less modal redundancy, a higher average kurtosis indicates more pronounced impact features, and the α prior penalty is used to prevent candidate α from deviating excessively from the empirically stable region. By minimizing this fitness, an α more suitable for cavitation acoustic signals can be obtained.

[0059] Furthermore, when constructing training samples in MATLAB, each set of operating condition data is first organized into VMD output units according to channel and segment numbering, and then each VMD output unit is windowed. The window length is 500 points, and the step size can be 500 points, or it can be set to an overlap step size of less than 500 points according to the real-time detection and sample amplification requirements. Each window retains K IMF channels, forming a K×500 two-dimensional input; for each IMF channel... The channel is used to calculate the mean within that channel window. and standard deviation And according to the following formula: Standardization is performed. This standardization step can reduce the impact of amplitude differences between different sampling segments on CNN training.

[0060] Furthermore, training labels can be established based on experimental conditions, cavitation coefficients, or manual calibration results. For four-class classification detection, samples can be labeled as no cavitation, nascent cavitation, developing cavitation, and severe cavitation. Nascent cavitation, as an independent category, participates in training, which can improve the system's sensitivity to early cavitation states. If the number of samples in different categories is inconsistent, this embodiment can adopt a class balancing strategy: randomly downsample the category with a larger sample size, or draw the same number of samples from each category according to the minimum category size; then, all samples are randomly shuffled and divided into training set, validation set, and test set.

[0061] Furthermore, the small-scale, medium-scale, and large-scale branches of MSCNN operate in parallel. The small-scale branch uses shorter convolutional kernels to primarily capture short-duration spikes and local high-frequency perturbations in the early stages of cavitation; the medium-scale branch uses medium-length convolutional kernels to primarily capture impact clusters and local energy changes within a series of consecutive sampling points; and the large-scale branch uses longer convolutional kernels to primarily capture envelope changes and low-frequency modulation features of the acoustic signal over a longer time range. The outputs of the three branches are complementary in both the temporal and channel dimensions, and their fusion can simultaneously express the differences between IMFs and multi-scale temporal features.

[0062] Furthermore, during MSCNN training, a cross-entropy loss function is used, with the difference between the class probabilities output by Softmax and the true labels as the optimization objective. The Adam optimizer adaptively adjusts the parameter update magnitude based on the first and second moments. The validation set is used to observe whether the model is overfitting; training can be stopped or the network parameters with the best validation performance can be selected when the validation accuracy stabilizes or the validation loss no longer decreases. After training, the overall accuracy is calculated using the test set, and the confusion matrix is ​​used to observe the confusion between nascent cavitation and other classes. If the nascent cavitation recognition is insufficient, optimization can be achieved by increasing the number of nascent cavitation samples, adjusting the window stride, or adjusting the number of branch convolutional kernels.

[0063] Furthermore, during parameter export, the network parameters in MATLAB are rearranged according to the storage order of the DSP inference program. Convolutional kernel parameters are arranged by output channel, input channel, and kernel position; BN parameters include S-mean, variance, scale, and bias; FC parameters are arranged by output category and input feature order. After export, an equivalent header file is generated and linked with VMD, normalization, and inference functions during DSP project compilation. Since the DSP only performs forward inference, there is no need to save the training optimizer state or backpropagation-related parameters.

[0064] Furthermore, the DSP-side inference function is executed layer by layer. First, the normalized K×500 input window is read; then, three 1DConv branches are executed, and BN and ReLU are performed on each convolution output; subsequently, the outputs of the three branches are written to the fusion buffer; then, GAP is performed on the fusion buffer to obtain a fixed-length feature vector; finally, FC and Softmax are executed to obtain the probabilities of each class. To facilitate embedded implementation, GAP can be accumulated and averaged after the convolution outputs are generated, reducing additional buffering; Softmax can be calculated after numerical stabilization to avoid exponential overflow.

[0065] Furthermore, the device can be configured with multiple operating modes during field operation. The first is a real-time detection mode, where the device continuously acquires acoustic signals and periodically outputs cavitation status. The second is a data feedback mode, where the device sends the raw acoustic signals, VMD decomposition results, or CNN classification probabilities to a host computer for debugging and archiving. The third is a parameter verification mode, where the device reads pre-stored test data and outputs intermediate VMD results and classification results to verify the consistency between the DSP-side implementation and the MATLAB-side results. The fourth is an alarm mode, where when several consecutive windows are determined to indicate developing or severe cavitation, the device outputs alarm information through the display module or communication module.

[0066] Furthermore, the LCD display interface can include the current detection status, four types of probabilities, cavitation level, sampling channel, running time, and communication status. For portable detection terminals, the display module can also display the original acoustic signal waveform or the selected IMF waveform, allowing on-site personnel to observe signal changes. The communication module can send data frames according to a preset frame format and store historical data. Through the above design, the device can not only provide cavitation status results but also retain intermediate data for subsequent analysis.

[0067] Furthermore, the embedded deployment method of this invention differs from simple PC-based offline recognition. The PC primarily handles model training, parameter selection, and algorithm verification; the DSP primarily handles on-site data acquisition, VMD decomposition, sample construction, MSCNN forward inference, result display, and communication feedback. With the separation of training and inference, the DSP's computational flow is fixed, parameters are fixed, and output is stable, thus avoiding the latency and wiring complexity caused by transmitting large amounts of on-site acoustic signals to the PC for analysis.

[0068] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for detecting cavitation acoustic signals in a hydro turbine based on embedded systems, characterized in that, include: Step S1: Collect the main detection acoustic channel signal and the reference acoustic channel signal during the operation of the water turbine, and obtain the digital signal after analog-to-digital conversion; Step S2: The VMD program is implanted into the DSP and lightweighted. The VMD program is used to perform variational mode decomposition on the digital signal according to the preset number of IMFs K and the penalty parameter α to obtain K intrinsic mode functions. The VMD is variational mode decomposition. Step S3: Based on the reference acoustic channel signal, perform modal reliability evaluation on the intrinsic mode function, and perform weighted enhancement on the intrinsic mode function of the main detection acoustic channel according to the evaluation results; Step S4: Construct multi-channel input samples based on the enhanced intrinsic mode functions, and standardize each channel. Step S5: Input the standardized multi-channel input samples into the pre-trained MSCNN network. The MSCNN network performs multi-scale convolutional feature extraction, feature fusion, and classification mapping on the multi-channel input samples to obtain the turbine cavitation state recognition result. Step S6: Display the results, provide alarm prompts, and transmit data back based on the cavitation status identification results of the water turbine.

2. The method for detecting cavitation acoustic signals of a hydro turbine based on embedded systems according to claim 1, characterized in that, The selection of the number of IMFs K includes: exist Perform VMD within each round to calculate residual energy. : in, Let K ∈ {3,4,5,6,7,8,9,10}; The minimum value is represented by x(n); x(n) represents the nth sampling point of the input digital acoustic signal. The k-th intrinsic mode function is represented by N; N represents the signal length. After normalizing the three indicators, a comprehensive indicator is constructed. in, The average correlation of the IMF; The average envelope entropy of the IMF; make according to Determination of the inflection point of adjacent improvement .

3. The method for detecting cavitation acoustic signals of a hydro turbine based on embedded systems according to claim 1, characterized in that, The penalty parameter The specific choices include: exist Perform VMD within each round to calculate the average envelope entropy of the IMF. Average correlation and average kurtosis And construct fitness : Let α ∈ {500, 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000, 5500, 6000}; select smallest As a VMD decomposition parameter; Represents the average envelope entropy; Indicates the average correlation of the IMF; Indicates the effective IMF mean kurtosis; Indicates the prior value of α; Indicates the prior penalty weight; This represents the minimum value.

4. The method for detecting cavitation acoustic signals of a hydro turbine based on embedded systems according to claim 1, characterized in that, The VMD program porting to the DSP and lightweighting process specifically includes: The matrix operations in the VMD decomposition process are converted into one-dimensional array loop operations based on C language, and the input time domain signal, the frequency domain representation of the input signal, the frequency domain representation of each intrinsic mode component, the center frequency, the Lagrange multiplier, the cumulative update amount, and the output results of the intrinsic mode components are stored in pre-allocated fixed-length arrays respectively. The system employs a single-precision floating-point data format and pre-sets the input signal length, the number of modal components, the maximum number of iterations, and the storage space of each array, so that dynamic memory allocation and deallocation are not performed during the VMD decomposition process. During the center frequency update process, the center frequency is calculated only based on the spectral energy of the corresponding intrinsic mode components within the frequency update interval, and the update of the Lagrange multiplier constraint terms in the original VMD iteration process is omitted. The number of modal components K, the penalty factor α, the maximum number of iterations, and the convergence threshold are set to predetermined fixed parameters, and these fixed parameters are written into the DSP program.

5. The method for detecting cavitation acoustic signals of a hydro turbine based on embedded systems according to claim 1, characterized in that, The modal reliability assessment of the intrinsic mode functions specifically includes: The first of the main detection sound channels The intrinsic mode functions are denoted as And select the reference mode whose center frequency is closest from the set of intrinsic mode functions of the reference acoustic channel, denoted as . ; Based on the intrinsic mode function of the main detection acoustic channel With the reference mode The temporal correlation coefficient is determined by considering the temporal correlation, frequency coherence, and energy overlap between the components. Frequency domain coherence coefficient and energy overlap coefficient ; Based on the time-domain correlation coefficient Frequency domain coherence coefficient and energy overlap coefficient Determine the first Modal reliability coefficients of each intrinsic mode function: in, For the first Modal reliability coefficients of each intrinsic mode function. , and These are the weighting coefficients, and .

6. The method for detecting cavitation acoustic signals of a hydro turbine based on embedded systems according to claim 5, characterized in that, The intrinsic mode function is enhanced based on the modal reliability coefficient, specifically including: Based on the modal reliability coefficient Determine the first Enhancement coefficients of each eigenmode function: in, For the first The enhancement coefficients of each intrinsic mode function, These are the modal retention coefficients; According to the enhancement coefficient The first detection channel of the main detection channel Individual eigenmode functions After weighting, the enhanced IMF modes are obtained. ; The reference acoustic channel signal is used only for reliability assessment and modal enhancement of the intrinsic mode functions of the main detection acoustic channel, and does not directly participate in the cavitation state category determination.

7. The method for detecting cavitation acoustic signals of a hydro turbine based on embedded systems according to claim 1, characterized in that, The input data processing for the MSCNN includes: obtaining the data after modal reliability evaluation and weighted augmentation. Each intrinsic mode function; normalization is performed on each IMF channel: in, To enhance the k-th IMF point n, This represents the mean of the k-th IMF window before enhancement. This represents the standard deviation of the k-th IMF window before enhancement.

8. The method for detecting cavitation acoustic signals of a hydro turbine based on embedded systems according to claim 1, characterized in that, The MSCNN network includes an input layer, a first-scale convolutional branch, a second-scale convolutional branch, a third-scale convolutional branch, a feature fusion layer, a fully connected layer, and a classification output layer. The input layer is used to receive standardized multi-channel input samples and construct multi-channel time-series input samples according to the arrangement order of the intrinsic mode components; The first-scale convolution branch is used to sequentially perform first-scale one-dimensional convolution processing, nonlinear activation processing and pooling processing on the multi-channel temporal input samples to extract the local transient impact features generated by the collapse of cavitation bubbles and obtain the first-scale feature vector. The second-scale convolution branch is used to sequentially perform second-scale one-dimensional convolution processing, nonlinear activation processing, and pooling processing on the multi-channel temporal input samples to extract local correlation features and local energy change features between multiple consecutive impact events, and obtain the second-scale feature vector. The third-scale convolution branch is used to sequentially perform third-scale one-dimensional convolution processing, nonlinear activation processing, and pooling processing on the multi-channel time-series input samples to extract the signal envelope change features and the overall evolution features of cavitation impact within a preset time range, and obtain the third-scale feature vector. Among them, the kernel size used in the first-scale one-dimensional convolution processing, the second-scale one-dimensional convolution processing and the third-scale one-dimensional convolution processing increases sequentially, so that each convolution branch obtains the cavitation sound signal characteristics in the receptive field at different times. The feature fusion layer is used to perform dimension unification processing on the first-scale feature vector, the second-scale feature vector, and the third-scale feature vector, and then splice the dimension-unified feature vectors of each scale along the feature channel direction to obtain a multi-scale fused feature vector. The fully connected layer is used to perform feature mapping on the multi-scale fused feature vector to obtain a classification feature vector characterizing the cavitation state of the turbine. The classification output layer is used to output the predicted probability of the turbine being in different cavitation states based on the classification feature vector, and to determine the cavitation state with the highest predicted probability as the cavitation identification result. The MSCNN is trained on a PC using training samples labeled with cavitation states. The weight parameters, bias parameters, and fully connected layer parameters of each convolutional layer are exported and fixed to the DSP. The DSP executes MSCNN forward inference based on the VMD program and outputs the turbine cavitation state recognition result.

9. A device for detecting cavitation acoustic signals in a hydro turbine based on an embedded system, characterized in that, It includes a sensor module, an ADC module, a DSP chip module, a communication module, and a display module; The sensor module is used to collect acoustic signals during the operation of the water turbine and output the collected acoustic signals to the ADC module. The ADC module is used to amplify, filter, and perform analog-to-digital conversion on the acoustic signal of the turbine operation, and output the converted digital signal to the DSP chip module. The DSP chip module is used to sequentially perform VMD decomposition, intrinsic mode function reliability assessment based on reference acoustic channel, intrinsic mode function weighted enhancement, multi-channel sample construction, and MSCNN forward inference on the digital signal to obtain the cavitation state identification result of the turbine. The communication module is used to transmit the detection results, cavitation state probability, alarm status and corresponding time information back to the host computer. The display module is used to display the cavitation state category and the corresponding cavitation state identification result.

10. The embedded-based cavitation acoustic signal detection device for water turbines according to claim 9, characterized in that, The device also includes a storage module, an alarm module, and a power module; The storage module is used to store sampling data, VMD processing intermediate data, and model parameters; The alarm module is used to output an alarm prompt when the preset alarm conditions are met in the cavitation state; The power supply module is used to supply power to the sensor module, ADC module, DSP chip module, storage module, communication module, display module and alarm module.

Images