Industrial device edge detection method, system and device based on voiceprint recognition

Abstract

The present application relates to the technical field of industrial equipment detection, and provides an industrial equipment edge detection method, system and device based on voiceprint recognition. The detection method comprises: collecting a voiceprint signal of the running industrial equipment; performing adaptive modal decomposition on the voiceprint signal to obtain a plurality of intrinsic modal components; removing noise dominant components based on the sample entropy of each intrinsic modal component; performing wavelet threshold denoising on the retained effective intrinsic modal components; reconstructing to obtain a pure voiceprint signal and extracting an acoustic feature vector; inputting the acoustic feature vector into a lightweight classifier model deployed in an edge computing unit to output a fault diagnosis result; returning the labeled data collected by the edge computing unit to the cloud to update the lightweight classifier model through incremental learning, and distributing the updated model to the edge computing unit. The present application deploys acoustic feature extraction and a lightweight classifier model in an edge computing unit, eliminates cloud transmission delay and bandwidth pressure, and reduces costs.

Description

Technical Field

[0001] This invention relates to the technical field of industrial equipment inspection, and in particular to a method, system and device for edge detection of industrial equipment based on voiceprint recognition. Background Technology

[0002] Industrial rotating machinery (motors, fans, pump sets, compressors, etc.) is prone to failures such as bearing wear, rotor imbalance, component loosening, and uneven air gaps during long-term operation. These failures cause abnormal vibrations and noise, leading to decreased equipment operating efficiency and a significant increase in ineffective energy consumption. Statistics show that ineffective energy consumption caused by abnormal equipment conditions accounts for 15% to 25% of total industrial energy consumption, and the proportion of ineffective energy consumption for large equipment reaches 20% to 25%. This not only results in a large amount of energy waste but also easily leads to equipment failure, production interruptions, and safety accidents.

[0003] Traditional equipment monitoring mainly relies on vibration sensing, which has problems such as intrusive installation, complex wiring, high maintenance costs, and poor adaptability, making it difficult to meet the monitoring needs of small and medium-sized enterprises for low cost, rapid deployment, and wide coverage.

[0004] In recent years, non-invasive monitoring based on acoustic signals has gradually become a research hotspot. However, existing acoustic detection solutions still have significant drawbacks: traditional filtering methods struggle to effectively reduce noise while preserving fault characteristics, leading to feature distortion and high false alarm rates; monitoring solutions largely rely on cloud computing, resulting in high transmission latency, large bandwidth consumption, data privacy risks, and inability to function without network connectivity, making it difficult to meet the millisecond-level real-time early warning requirements of industrial sites; feature extraction depends on human experience and has weak generalization ability; and the large model size makes it impossible to deploy on resource-constrained edge hardware such as ESP32 and STM32 for real-time inference. Furthermore, most systems only support single-device monitoring, failing to achieve universal diagnosis for multiple types of equipment and relying on manually designed features and fixed models, with models lacking incremental updates, making it difficult to adapt to long-term equipment aging and changing operating conditions. Summary of the Invention

[0005] This invention provides an edge detection method for industrial equipment based on voiceprint recognition, which solves the defects of existing technologies such as strong industrial noise interference, poor noise reduction effect, easy false alarm of single signal, high latency due to reliance on cloud, difficulty in deploying models on edge hardware, weak long-term operating condition adaptability, and high cost.

[0006] This invention provides a method for edge detection of industrial equipment based on voiceprint recognition, comprising the following steps:

[0007] S1. An external, non-invasive sensor collects acoustic signature signals from the operation of industrial equipment;

[0008] S2. Perform adaptive mode decomposition on the voiceprint signal to obtain multiple intrinsic mode components. Based on the sample entropy of each intrinsic mode component, remove the noise-dominant component. Perform wavelet threshold denoising on the remaining effective intrinsic mode components to reconstruct a clean voiceprint signal.

[0009] S3. Extract acoustic feature vectors from the pure voiceprint signal;

[0010] S4. The acoustic feature vector is input into a lightweight classifier model deployed in an edge computing unit, and the fault diagnosis result is output.

[0011] S5. The labeled data collected by the edge computing unit is sent back to the cloud, the lightweight classifier model is updated through incremental learning, and the updated model is sent back to the edge computing unit.

[0012] According to the industrial equipment edge detection method based on voiceprint recognition provided by the present invention, step S2 specifically includes the following steps:

[0013] Perform CEEMDAN decomposition on the original voiceprint signal x(t) to obtain IMF1~IMF n With residual component RES;

[0014] Calculate the sample entropy of each IMF component, and identify and remove IMF components whose sample entropy is greater than a preset threshold as noise-dominant components.

[0015] Wavelet decomposition and soft thresholding denoising are performed on the retained effective IMF components.

[0016] The reconstructed IMF components yield a clean voiceprint signal.

[0017] According to the industrial equipment edge detection method based on voiceprint recognition provided by the present invention, step S3 specifically includes the following steps:

[0018] The pure voiceprint signal is subjected to pre-emphasis, frame windowing, short-time Fourier transform, power spectrum calculation, Mel filter bank filtering, logarithmic operation and discrete cosine transform in sequence, and the 13-dimensional MFCC feature vector is extracted as the acoustic feature vector.

[0019] According to the industrial equipment edge detection method based on voiceprint recognition provided by the present invention, the lightweight classifier model is obtained using the following method:

[0020] Using a high-precision neural network in the cloud as the teacher model and a micro neural network at the edge as the student model, the fault discrimination knowledge of the teacher model is transferred to the student model through a double-loss weighted distillation scheme.

[0021] The distilled student model is subjected to parameter quantization, mapping floating-point weights to fixed-point integers;

[0022] The quantized model is then subjected to structured pruning to remove redundant parameters.

[0023] The compressed lightweight classifier model has 10KB-20KB of parameters, a peak memory usage of ≤50KB, and a single inference time of ≤17ms.

[0024] According to the industrial equipment edge detection method based on voiceprint recognition provided by the present invention, step S5 specifically includes:

[0025] When the confidence level of multiple consecutive diagnoses is lower than the first preset threshold, or the decrease in confidence level exceeds the second preset threshold, the voiceprint data collected by the edge computing unit and the corresponding annotation information are sent back to the cloud.

[0026] The cloud uses few-shot augmentation technology to fine-tune the local parameters of the top classification layer and feature mapping layer of the lightweight classifier model without retraining the entire model.

[0027] The fine-tuned model is then sent to the edge computing unit.

[0028] According to the industrial equipment edge detection method based on voiceprint recognition provided by the present invention, step S2 further includes: performing FFT spectrum analysis on the reconstructed pure voiceprint signal to extract the characteristic energy changes in the 8kHz-12kHz frequency band for use in identifying bearing wear or rotor imbalance faults.

[0029] According to the industrial equipment edge detection method based on voiceprint recognition provided by the present invention, the fault diagnosis result in step S4 includes fault type, level and diagnostic confidence score, wherein the level is determined based on the correspondence between fault type and equipment operating conditions.

[0030] Another aspect of the present invention provides a detection system for performing the above-described voiceprint recognition-based edge detection method for industrial equipment, comprising:

[0031] The signal acquisition module is used to acquire acoustic signature signals during the operation of industrial equipment;

[0032] A joint noise reduction module, deployed at the edge, is used to perform adaptive mode decomposition on the voiceprint signal to obtain multiple intrinsic mode components, remove noise-dominant components based on the sample entropy of each intrinsic mode component, perform wavelet threshold noise reduction on the remaining effective intrinsic mode components, and reconstruct a clean voiceprint signal.

[0033] The feature extraction and inference module, deployed on the edge side, is used to extract acoustic feature vectors from the pure voiceprint signal, input them into a lightweight classifier model, and output fault diagnosis results.

[0034] The local interaction module is used to display the fault diagnosis results and trigger an alert when the diagnosis confidence exceeds a preset threshold.

[0035] The cloud-based model update module is used to receive labeled data uploaded from the edge side, update the lightweight classifier model through incremental learning, and then send the updated model down to the edge side.

[0036] The communication module is used to enable data transmission between the edge and the cloud.

[0037] According to the detection system provided by the present invention, the edge side uses an ESP32 series or STM32 series microcontroller as the core computing unit.

[0038] In another aspect, the present invention provides an industrial equipment inspection device based on voiceprint recognition, comprising:

[0039] The housing has a receiving chamber inside, and a sound receiving channel is provided on one side wall of the housing. The sound receiving channel has a first end near the outside and a second end near the receiving chamber. The second end is a closed end, and a connecting channel communicating with the receiving chamber is provided on the closed end. The sound receiving channel communicates with the outside and the receiving chamber through the connecting channel, and the inner diameter of the connecting channel is smaller than the inner diameter of the sound receiving channel.

[0040] A sound receiving component, located within the accommodating cavity, is used to convert the collected sound waves into electrical signals;

[0041] A guide tube, located within the sound receiving channel, is used to collect external sound signatures;

[0042] A sound guide tube, located within the accommodating cavity, is used for the transmission of sound patterns. One end of the sound guide tube is connected to the connecting channel, and the other end is connected to the sound receiving component.

[0043] A voiceprint analysis module is disposed in the accommodating cavity. The voiceprint analysis module is electrically connected to the sound receiving component. The voiceprint analysis module is used for the industrial equipment edge detection method based on voiceprint recognition provided above, or the voiceprint analysis module is used to deploy the detection system provided above to determine whether the industrial equipment is abnormal.

[0044] The present invention provides an edge detection method for industrial equipment based on voiceprint recognition. This method employs adaptive mode decomposition combined with wavelet thresholding for denoising. First, the original voiceprint signal is adaptively decomposed into multiple intrinsic mode components using CEEMDAN. Then, sample entropy is used to intelligently identify and remove noise-dominant components. Next, wavelet thresholding is applied to the remaining effective components to filter out residual broadband noise, ultimately reconstructing a clean voiceprint signal. Experimental data shows that compared to single denoising schemes using only CEEMDAN or only wavelet thresholding, the combined denoising scheme of this invention improves the signal-to-noise ratio by ≥12dB, achieves a fault feature retention rate of ≥96%, and reduces the system false alarm rate by more than 80%. It achieves high-fidelity purification of weak fault voiceprint signals even in high-noise industrial environments, ensuring the accuracy of subsequent diagnosis.

[0045] Furthermore, acoustic feature extraction and lightweight classifier models are deployed on edge computing units, enabling the entire computation process from signal processing to fault diagnosis to be completed locally in the industrial field, achieving real-time inference without uploading data to the cloud. After multi-level compression through knowledge distillation, parameter quantization, and structured pruning, the model has only 10KB-20KB of parameters, a peak memory usage of ≤50KB, and a single inference time of ≤17ms, allowing it to run smoothly on resource-constrained, low-cost microcontrollers such as ESP32 and STM32. The local closed-loop processing mode not only eliminates cloud transmission latency and bandwidth pressure but also ensures the system's independent operation capability in the event of a network outage, meeting the stringent requirements of millisecond-level real-time early warning in industrial settings and effectively reducing costs.

[0046] Furthermore, this invention uses acoustic signature signals as the monitoring medium. The sensor is externally mounted, requiring no structural modifications to the device itself or any wiring. This results in high deployment efficiency and low cost for a single unit. It is particularly suitable for small and medium-sized enterprises with mixed new and old equipment and difficult wiring scenarios, significantly reducing the deployment threshold and overall cost of industrial equipment condition monitoring. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating the edge detection method for industrial equipment based on voiceprint recognition provided by the present invention.

[0049] Figure 2 This is a flowchart of the industrial equipment edge detection method based on voiceprint recognition provided by the present invention.

[0050] Figure 3 This is a flowchart of the CEEMDAN combined with wavelet threshold denoising process in the industrial equipment edge detection method based on voiceprint recognition provided by this invention.

[0051] Figure 4 This is a time-domain comparison diagram of the noise-reduced sound signal in the industrial equipment edge detection method based on voiceprint recognition provided by the present invention.

[0052] Figure 5 This is a comparative example diagram of different noise reduction methods in the industrial equipment edge detection method based on voiceprint recognition provided by the present invention.

[0053] Figure 6 These are the signal-to-noise ratio and error histograms of different noise reduction methods in the industrial equipment edge detection method based on voiceprint recognition provided by this invention.

[0054] Figure 7 This is one of the cloud server interface diagrams in the industrial equipment edge detection method based on voiceprint recognition provided by this invention.

[0055] Figure 8 This is the second cloud server interface diagram in the industrial equipment edge detection method based on voiceprint recognition provided by this invention.

[0056] Figure 9 This is a schematic diagram of the layout of the industrial equipment edge detection system based on voiceprint recognition provided by the present invention.

[0057] Figure 10 This is a schematic diagram of the overall structure of the industrial equipment edge detection device based on voiceprint recognition provided by the present invention.

[0058] Figure 11 This is a schematic diagram of the internal cross-sectional structure of the industrial equipment edge detection device based on voiceprint recognition provided by the present invention.

[0059] Figure 12 A schematic diagram of the lateral structure of the industrial equipment edge detection device based on voiceprint recognition provided by the present invention.

[0060] Figure label:

[0061] 10. Housing; 11. Main housing; 111. Charging port; 112. Radio channel; 1121. Second end; 113. Connection channel; 12. First connecting housing; 121. Display; 13. Second connecting housing; 131. Energy storage chamber; 14. Receiving chamber; 20. Guide tube; 30. Radio receiving component; 40. Sound guide tube; 50. Voiceprint analysis module. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0063] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0064] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0065] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0066] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0067] like Figure 1 , Figure 2 As shown, to address the problems of strong industrial noise interference, poor noise reduction effect, high latency due to reliance on cloud computing, difficulty in deploying models on edge hardware, and weak long-term adaptive capability in existing industrial equipment abnormal energy consumption detection technologies, this embodiment provides an edge detection method for abnormal energy consumption of industrial equipment based on voiceprint recognition, including the following steps:

[0068] Step S1: An external, non-invasive sensor collects the acoustic signature signal of the industrial equipment during operation;

[0069] When collecting signals, a non-invasive method is employed. This can be done by placing the acquisition module or device externally to the industrial equipment. Alternatively, acoustic sensors (such as high-sensitivity miniature microphones) can be mounted on the surface of the industrial equipment's casing without requiring any structural modifications to the equipment itself. The objects collected include, but are not limited to, the sound signals generated by rotating machinery such as motors, fans, pumps, and compressors. Sensors with high sensitivity (e.g., -26dBFS) and a high signal-to-noise ratio (e.g., 62dB) can be selected to accurately detect subtle abnormal noises such as bearing wear, rotor imbalance, and loose components in complex industrial noise environments.

[0070] Step S2: Perform adaptive mode decomposition on the voiceprint signal to obtain multiple intrinsic mode components. Based on the sample entropy of each intrinsic mode component, remove the noise-dominant component. Perform wavelet threshold denoising on the remaining effective intrinsic mode components to reconstruct a clean voiceprint signal.

[0071] When acquiring acoustic signals in the field, the presence of significant background noise, multi-device coupling noise, and environmental vibration noise in industrial environments can severely obscure weak fault acoustic signature signals. This embodiment employs a dual denoising strategy combining adaptive mode decomposition and wavelet threshold denoising. Details are as follows:

[0072] First, adaptive mode decomposition (IMD) is performed on the acquired voiceprint signal. IMD is a signal processing method that decomposes a nonlinear, non-stationary signal into multiple intrinsic mode functions (IMFs). Algorithms such as Empirical Mode Decomposition (EMD), Ensemble Empirical Mode Decomposition (EEMD), or Adaptive Noise Complete Ensemble Empirical Mode Decomposition (CEEMDAN) can be used. After decomposition, the original signal is represented as the sum of several IMF components and a residual component.

[0073] Next, the sample entropy of each IMF component is calculated. Sample entropy is a metric for time series complexity; noisy signals typically have high sample entropy, while signal components containing deterministic fault characteristics have relatively low sample entropy. Based on this characteristic, IMF components with sample entropy greater than a preset threshold are identified as noise-dominant components and removed, while valid IMF components with lower sample entropy are retained.

[0074] Next, wavelet thresholding denoising is performed on the retained effective IMF components. Wavelet transform has multi-resolution analysis capabilities and can effectively separate abrupt changes and stationary noise in a signal. Specifically, wavelet decomposition is performed on each effective IMF component to obtain wavelet coefficients at each level; thresholding (such as soft or hard thresholding) is applied to the wavelet coefficients at each level according to a preset threshold to suppress the wavelet coefficients corresponding to noise; then, inverse wavelet transform reconstruction is performed to obtain the denoised IMF components.

[0075] Finally, the processed effective IMF components are superimposed and reconstructed to obtain a clean acoustic signature signal. This joint noise reduction method combines the adaptive mode decomposition capability of non-stationary signals with the wavelet transform capability of fine characterization of local features, and can effectively suppress strong industrial noise while preserving fault feature information to the maximum extent.

[0076] Understandably, the signal-to-noise ratio of the voiceprint signal is significantly improved through the joint noise reduction process in step S2, providing a high-quality data foundation for subsequent feature extraction and fault diagnosis, and reducing the risk of false alarms and missed alarms from the source.

[0077] like Figure 3 As shown, in step S2, the specific noise reduction method is to use CEEMDAN combined with wavelet threshold noise reduction. The specific steps are as follows:

[0078] Step S21: Perform CEEMDAN decomposition on the original voiceprint signal x(t) to obtain IMF1~IMF. n With residual component RES;

[0079] Specifically, let the original signal be x(t). In the k-th decomposition, adaptive noise is added to the residual signal, and then the k-th IMF component is extracted. This process is repeated until the residual component can no longer be decomposed. Finally, n IMF components arranged from high to low frequency and a residual component RES are obtained.

[0080] Step S22: Calculate the sample entropy of each IMF component, and identify the IMF components with sample entropy greater than a preset threshold as noise-dominant components and remove them.

[0081] When identifying noise components, the calculation of sample entropy involves two parameters: the embedding dimension *m* and the similarity tolerance *r*. For a time series containing N data points, sample entropy is defined as the negative natural logarithm of the conditional probability, i.e., the probability of a new pattern emerging. In this embodiment, the preset threshold can be calibrated experimentally, such as by setting it in the range of 0.2 to 0.5. IMF components with sample entropy greater than the threshold are considered high-frequency components dominated by noise and are discarded.

[0082] Step S23: Perform wavelet decomposition and soft thresholding denoising on the retained effective IMF components;

[0083] During wavelet thresholding denoising, wavelet decomposition is performed on each retained effective IMF component. Optional wavelet basis functions include db4, sym8, etc., and the number of decomposition layers can be 3 to 5.

[0084] Step S24: Reconstruct the processed IMF components to obtain a clean voiceprint signal.

[0085] In this embodiment, CEEMDAN (Adaptive Noise Complete Set Empirical Mode Decomposition) is used as the specific algorithm for adaptive mode decomposition. CEEMDAN effectively solves the mode aliasing problem existing in traditional EMD algorithms by adaptively adding white noise during the decomposition process, and has better decomposition completeness and stability. Compared with single denoising schemes that only use CEEMDAN or only use wavelet thresholding, the joint denoising scheme in this embodiment can achieve better denoising effect.

[0086] Specific simulation data shows that, in a high-noise industrial environment (SNR=5dB), the CEEMDAN combined with wavelet threshold denoising scheme in this embodiment can improve the signal-to-noise ratio of the acoustic signature signal by ≥12dB and the fault feature retention rate by ≥96%. Compared with single schemes using only CEEMDAN denoising (SNR improvement of approximately 6.2dB) or only wavelet threshold denoising (SNR improvement of approximately 7.5dB), it has a significant synergistic effect. The fault identification accuracy is improved from approximately 85% with the single scheme to over 95%, and the system false alarm rate is reduced by over 80%. Figure 4 As shown, the above-mentioned CEEMDAN combined with wavelet threshold denoising scheme can significantly improve the denoising effect.

[0087] It is understandable that this embodiment, through the specific implementation path of CEEMDAN decomposition combined with sample entropy discrimination and the specific processing method of wavelet soft thresholding denoising, makes the joint denoising scheme highly operable and reproducible. In particular, it can effectively suppress strong background noise in industrial sites while preserving weak fault features such as bearing wear and rotor imbalance, providing high-fidelity acoustic signature signal input for subsequent acoustic feature extraction.

[0088] Step S2 also includes performing FFT spectrum analysis on the reconstructed pure acoustic signal to extract characteristic energy changes in the 8kHz-12kHz frequency band, which are used to identify bearing wear or rotor imbalance faults.

[0089] Common fault types in industrial rotating machinery exhibit specific frequency domain characteristics in the acoustic waveform spectrum. Extensive experimental testing and spectral analysis have revealed that the characteristic frequencies of typical faults such as bearing wear and rotor imbalance are mainly concentrated in the 8kHz-12kHz frequency band. Specifically, the high-frequency components of friction generated by bearing wear typically show a significant energy increase in the 8kHz-10kHz range, while the periodic impacts caused by rotor imbalance produce sideband modulation characteristics in the 10kHz-12kHz frequency band.

[0090] Specifically, for the clean speaker signal reconstructed by CEEMDAN combined with wavelet thresholding, the time-domain signal is converted into a frequency-domain signal using FFT (Fast Fourier Transform) to obtain the signal's spectral distribution. Then, the energy integral or average power spectral density in the 8kHz-12kHz frequency band is calculated and used as the characteristic energy index of this frequency band.

[0091] For example: the system calibrates the reference energy value E for this frequency band under normal operating conditions. base During real-time monitoring, the current characteristic energy value E is continuously calculated. curr And compare it with the benchmark value:

[0092] If E curr / E base If the energy level is greater than 1.5 (i.e., the energy increase exceeds 50%), and the energy increase is mainly concentrated in the 8kHz-10kHz sub-frequency band, then it is determined that there may be a bearing wear fault.

[0093] If E curr / E base If the value is greater than 1.8 and a significant modulation sideband is detected in the 10kHz-12kHz sub-band, it is determined that there may be a rotor imbalance fault.

[0094] In this embodiment, by focusing on the characteristic energy changes in a specific frequency band of 8kHz-12kHz, interference from low-frequency mechanical vibration and environmental noise can be effectively eliminated, reducing the false alarm rate at the source. Experimental data shows that after introducing the characteristics of this frequency band, the identification accuracy of bearing wear and rotor imbalance faults is improved to 97.2% and 96.5% respectively, and the overall false alarm rate of the system is further reduced.

[0095] It is understandable that this embodiment extracts the characteristic energy changes in a specific frequency band of 8kHz-12kHz, focusing the fault diagnosis on the spectral region that is strongly correlated with the target fault type, thus avoiding the computational redundancy and noise interference caused by full-band analysis, and achieving accurate identification of common faults such as bearing wear and rotor imbalance.

[0096] Specific examples of noise reduction effects, such as Figure 5 , Figure 6 As shown in the figure, specific experimental comparisons reveal that, for example... Figure 4In the previous example, the original voiceprint signal was overwhelmed by high-frequency noise, with an amplitude range of approximately ±2.5, indicating a large amount of unprocessed noise in the original voiceprint signal. Comparing Example 1, median filtering showed a weak effect, with an amplitude range of approximately ±2.0. Comparing Example 2, spectral subtraction denoising had limited effect, but still left obvious periodic fluctuation noise. Comparing Example 3, Wiener filtering had a similar effect to Comparison Example 2 (spectral subtraction), showing some improvement, but the effect was not significant. Comparison Example 4, using CEEMDAN + sample entropy denoising, resulted in excessive smoothing, losing high-frequency details of the signal, and causing distortion due to an overly rounded waveform. Comparison Example 5, using wavelet soft thresholding denoising, preserved many details, but periodic noise was still clearly visible. The denoising method in this application, however, achieves waveform smoothing through fusion, effectively suppressing noise while preserving the main trend characteristics of the signal, without significant distortion, and effectively improving the quality of denoising.

[0097] Furthermore, such as Figure 6 As shown, compared with traditional solutions, the denoising solution provided in this application can effectively improve the signal-to-noise ratio (SNR) by 10.6 dB, which is 8.2 dB higher than the baseline of 2.4 dB, significantly improving the SNR). Moreover, the mean square error of the solution in this application is only 0.0432, which is 59.3% higher than the traditional superior method (wavelet soft thresholding) and 40.7% lower error.

[0098] As can be seen from the above, the noise reduction scheme in this application can significantly improve the noise reduction effect, achieve efficient noise reduction, and the overall noise reduction effect is significantly improved compared with the traditional scheme, which can effectively provide better and more accurate data for subsequent voiceprint analysis.

[0099] Step S3: Extract acoustic feature vectors from the pure voiceprint signal.

[0100] Specifically, the acoustic feature extraction process involves the following steps: pre-emphasis to boost high-frequency components and highlight fault acoustic details; frame-by-frame windowing to suppress spectral leakage and ensure short-term signal stability; short-time Fourier transform (STFT) to obtain the time-frequency distribution of the signal; power spectrum calculation to obtain the signal energy distribution in the frequency domain; Mel filter bank filtering to simulate human hearing characteristics and enhance fault-sensitive frequency band features; and logarithmic operations and discrete cosine transform (DCT) for feature dimensionality reduction and decorrelation processing. The final extracted feature vector has high discriminative power and low redundancy.

[0101] Specifically, during acoustic feature extraction, the clean acoustic signature signal undergoes pre-emphasis, frame-by-frame windowing, short-time Fourier transform, power spectrum calculation, Mel filter bank filtering, logarithmic operation, and discrete cosine transform sequentially to extract a 13-dimensional MFCC feature vector as the acoustic feature vector. This embodiment applies MFCC to feature extraction of industrial equipment fault acoustic signatures. Considering the resource constraints of low-computing-power hardware at the edge, the feature dimension is determined to be 13 dimensions, effectively controlling computational complexity and storage consumption while ensuring feature discriminative power.

[0102] Specifically, the MFCC feature extraction process is as follows:

[0103] Pre-emphasis involves passing the pure acoustic signal through a first-order high-pass filter H(z) = 1 - μz⁻¹, where μ is the pre-emphasis coefficient, taken as 0.97. Pre-emphasis can enhance the high-frequency components of the signal, making fault characteristics (such as high-frequency abnormal noises caused by bearing wear) more prominent.

[0104] Framing and windowing involves dividing the pre-emphasized signal into frames with a preset frame length (e.g., 25ms) and frame shift (e.g., 10ms). Each frame is then multiplied by a Hamming window to suppress spectral leakage. The Hamming window function is: w(n) = 0.54 - 0.46·cos(2πn / (N-1)), where N is the frame length.

[0105] Short-Time Fourier Transform (STFT) performs a fast Fourier transform on each windowed signal to obtain the spectrum of that frame.

[0106] Power spectrum calculation: Calculate the square of the spectral amplitude to obtain the power spectrum of the signal frame.

[0107] Mel filter bank filtering filters the power spectrum through a set of triangular bandpass filters (Mel filter bank). The Mel filter bank has dense filters in the low-frequency region and sparse filters in the high-frequency region. The conversion relationship between the Mel frequency and the linear frequency is: Mel(f) = 2595·log 10 (1+f / 700). The number of filter banks is usually set to 24~40.

[0108] Logarithmic operations are performed by taking the logarithm of the energy output of each filter, simulating the logarithmic perception of sound intensity by the human ear.

[0109] Discrete Cosine Transform (DCT) is applied to the logarithmic energy spectrum to achieve feature dimensionality reduction and decorrelation. The first 13 coefficients of the DCT result are used to construct a 13-dimensional MFCC eigenvector.

[0110] The extraction of 13-dimensional MFCC feature vectors balances the full representation of fault features with the computational constraints of edge hardware. Since the aforementioned joint noise reduction steps have already improved the signal-to-noise ratio to a high level, even with the relatively simplified 13-dimensional features, it is still possible to effectively distinguish the differences in voiceprints between normal states and various fault states.

[0111] Understandably, this embodiment uses 13-dimensional MFCC as the acoustic feature vector. Its low feature dimensionality and low computational cost perfectly match the input requirements of lightweight classifier models at the edge. A single feature extraction takes only milliseconds, meeting the real-time diagnostic needs of industrial sites. Furthermore, the standardized MFCC extraction process ensures good robustness and generalization ability of the features, making it adaptable to the acoustic signature representation of various types of rotating machinery such as motors, fans, and pumps.

[0112] Step S4: Input the acoustic feature vector into the lightweight classifier model deployed on the edge computing unit, and output the fault diagnosis result.

[0113] Edge computing units are deployed in industrial equipment sites and can be embedded microcontrollers such as the ESP32 series and STM32 series. The lightweight classifier model is a compressed and optimized neural network model with parameters controlled at the tens of KB level, enabling fast inference on resource-constrained edge hardware. The model outputs fault diagnosis results, including fault type, abnormal power consumption level, and diagnostic confidence score.

[0114] In this embodiment, the lightweight classifier is obtained using the following method:

[0115] Step S41: Using a high-precision neural network in the cloud as the teacher model and a micro neural network at the edge as the student model, the fault discrimination knowledge of the teacher model is transferred to the student model through a double loss weighting scheme distillation.

[0116] In step S41, a dual-model architecture of teacher and student is first constructed. The teacher model is a high-precision neural network deployed in the cloud, which can adopt an 8-layer CNN structure, containing multiple convolutional-pooling units and fully connected layers. It is trained on a large amount of labeled data, and the fault identification accuracy can reach over 98%. The student model is a miniature neural network deployed at the edge, which adopts a 2-layer 1D-CNN structure, containing only two levels of convolutional-pooling units, Dropout regularization layers, and fully connected output layers. The structure is simplified and the computational requirements are low.

[0117] Knowledge distillation employs a double-loss weighted scheme. Let q be the soft label output by the teacher model (a probability distribution softened by a temperature coefficient T), and p be the predicted distribution output by the student model. The temperature coefficient T controls the smoothness of the output distribution; the larger T is, the smoother the distribution, revealing more information about the relationships between categories. In this embodiment, T is set to 7.

[0118] Soft tag q i The calculation formula is:

[0119]

[0120] in, Let C be the logit output of the i-th class of the teacher model, and C be the total number of fault categories.

[0121] Student model predicts distribution p i The calculation formula is:

[0122]

[0123] in, This is the logit output for the i-th class of the student model.

[0124] Total loss function L distill From soft loss L soft and hard loss L hard Composed of weighted combinations:

[0125]

[0126] Among them, soft loss L soft The KL divergence between teacher soft labels and student predicted distributions:

[0127]

[0128] Hard loss L hard The cross-entropy between the student model's prediction and the true label y:

[0129]

[0130] The weighting coefficient α is used to balance the contributions of soft and hard losses, and is set to 0.7~0.9. Through knowledge distillation, the student model can learn soft knowledge with strong generalization ability from the teacher model, and the accuracy loss is controlled within 3% while compressing the model volume by more than 20 times.

[0131] Step S42: Perform parameter quantization on the distilled student model, mapping the floating-point weights to fixed-point integers;

[0132] Affine quantization is performed on the distilled student model to map the 32-bit floating-point (FP32) weights to 8-bit fixed-point integers (INT8), which significantly reduces the model's storage space and computational complexity.

[0133] The quantification formula is:

[0134]

[0135] in, The original floating-point value. Here, S represents the quantized fixed-point integer value, S is the scaling factor, and Z is the zero point. The scaling factor S and the zero point Z are determined using a layer-by-layer calibration method, that is, the range of activation values ​​is statistically analyzed at the output of each layer of the network, and the optimal S and Z are calculated to minimize the quantization error. In this embodiment, affine quantization reduces the number of model parameters by approximately 75% and improves inference speed by 2 to 3 times.

[0136] Step S43: Perform structured pruning on the quantized model to remove redundant parameters; wherein, the compressed lightweight classifier model has a parameter size of 10KB-20KB, a peak memory usage of ≤50KB, and a single inference time of ≤17ms.

[0137] In step S43, the quantized model undergoes structured pruning, which involves removing redundant parameters as a whole, such as convolutional kernels and neurons, rather than pruning individual weights. The structured pruning process maintains a regular network structure, facilitating hardware acceleration.

[0138] The specific pruning strategy is as follows: calculate the L1 norm or L2 norm of each convolutional kernel as an importance score, and remove the convolutional kernel with the lowest score according to a preset pruning rate (e.g., 30%~50%). After pruning, perform a small amount of fine-tuning training on the model to restore accuracy.

[0139] The scheme in this embodiment, after three-level compression, results in a final model with 10KB-20KB parameters (preferably 12.8KB), peak memory usage ≤50KB (preferably 48KB), single inference time ≤17ms, and an F1 score of 0.952. This lightweight model can be directly burned into the flash memory of edge computing units such as ESP32-S3, completing the entire inference process locally without relying on cloud computing power.

[0140] In other words, this embodiment systematically solves the technical challenge of deploying high-precision deep learning models on resource-constrained edge hardware. By transferring the generalization ability of the teacher model through knowledge distillation, reducing storage and computational overhead through parameter quantization, and further simplifying the model structure through structured pruning, it ultimately achieves excellent performance of maintaining an F1 score of 0.952 with only 12.8KB of parameters, providing a feasible technical path for real-time edge diagnosis of abnormal energy consumption in industrial equipment.

[0141] In a specific implementation, the fault diagnosis results in step S4 include fault type, abnormal energy consumption level, and diagnostic confidence score, wherein the abnormal energy consumption level is determined based on the correspondence between fault type and equipment operating conditions.

[0142] In practical industrial applications, maintenance personnel not only need to know if equipment is faulty, but also need to understand key information such as the type of fault, its severity, and its impact on energy consumption. The fault diagnosis results output in this embodiment contain information in three dimensions:

[0143] The lightweight classifier model outputs fault category labels. Identifiable fault types include, but are not limited to: normal, bearing wear, rotor imbalance, loose components, uneven air gap, gear failure, poor lubrication, etc. Fault types are output in text or code format for easy display on OLED screens and storage in cloud databases. Figure 7 , Figure 8 As shown, the results of the detection and diagnosis can be displayed on the screen. Figure 8 The middle section is the login interface for remote testing or the testing room. Login is achieved by scanning the device code, and after logging in, access is granted. Figure 7 The front-end interface shown provides a clear and intuitive display of the currently detected data. For example... Figure 7 In the interface, the overall status of multiple devices is displayed in the upper left corner. Below each device status bar, the device's health level is shown, and below that, the collected acoustic signals are displayed. The right-hand display shows alarm information and corresponding AI fault analysis information. This indicates that technicians can... Figure 7 The displayed information provides a comprehensive understanding of all kinds of information.

[0144] Based on the correspondence between fault types and equipment operating conditions, assess the impact of the current fault state on equipment energy consumption. For example, abnormal energy consumption levels can be divided into three levels:

[0145] Level 1 (Minor Abnormality): The equipment can still operate normally, with an energy loss rate of <5%. It is recommended to carry out maintenance as planned.

[0146] Level 2 (Moderate Abnormality): Equipment performance has declined, with an energy loss rate of 5% to 15%. It is recommended to arrange maintenance as soon as possible.

[0147] Level 3 (Severe Abnormality): The equipment has serious potential for failure, with an energy loss rate >15%. It is recommended to shut down the machine immediately for inspection and repair.

[0148] The determination of abnormal energy consumption levels is based on a preset mapping table. An example mapping table is provided below:

[0149] Bearing wear (early stage) Rated speed 3%~6% Level 1 Bearing wear (intermediate stage) Rated speed 8%~12% Level 2 Bearing wear (late stage) Rated speed 15%~22% Level 3 Rotor imbalance Rated speed 5%~10% Level 2 loose parts Any working condition 4%~8% Level 1 or Level 2 Poor lubrication high speed 6%~12% Level 2

[0150] The radiation relationship provided in the table above is based on a large amount of measured data and equipment mechanism analysis, and can be dynamically calibrated in the cloud.

[0151] The maximum Softmax probability output by the model represents the model's confidence level in the current diagnostic result, ranging from 0 to 1. The confidence score can be used to assess the reliability of the diagnostic result and also provides a basis for triggering incremental learning.

[0152] It is understandable that this embodiment provides maintenance personnel with complete decision support information, from qualitative judgment to quantitative assessment, by outputting three-dimensional diagnostic results of fault type, abnormal energy consumption level, and confidence score. The introduction of abnormal energy consumption level directly links voiceprint diagnosis to energy conservation and carbon reduction goals, upgrading the system from a simple fault detection tool to an effective means of energy health management.

[0153] S5. The labeled data collected by the edge computing unit is sent back to the cloud, the lightweight classifier model is updated through incremental learning, and the updated model is sent back to the edge computing unit.

[0154] When a system operates for an extended period, and equipment experiences natural aging or changes in operating conditions, the diagnostic confidence of the model may decline. In this case, the edge computing unit transmits the collected voiceprint data and corresponding annotation information back to the cloud. The cloud employs few-shot augmentation technology, fine-tuning only the parameters of the model's top-level classification and feature mapping layers, eliminating the need to retrain the entire model and significantly reducing computational overhead and training time. After fine-tuning, the updated model is distributed back to the edge computing unit, completing incremental iterations. Through this closed-loop mechanism, the system can continuously adapt to changes in equipment status and maintain high diagnostic performance over the long term.

[0155] It should be noted that the execution subject of the above method embodiments can be an edge computing device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above steps.

[0156] Understandably, traditional industrial equipment anomaly detection primarily relies on vibration sensing or cloud-based acoustic analysis. Vibration sensing suffers from intrusive installation, complex wiring, and high maintenance costs; cloud-based acoustic analysis, on the other hand, suffers from high transmission latency, large bandwidth consumption, and inability to function when the network is down, making it difficult to meet the millisecond-level real-time early warning requirements of industrial sites. Furthermore, existing noise reduction methods in acoustic solutions struggle to effectively suppress strong industrial noise while preserving fault characteristics, and deep learning models are large and cannot be deployed on resource-constrained edge hardware.

[0157] In this embodiment, a non-invasive method for acquiring voiceprint signals, combined with a joint denoising scheme of adaptive mode decomposition and wavelet threshold denoising, enables accurate purification of voiceprint signals in environments with strong industrial noise. A lightweight classifier model is deployed on an edge computing unit, allowing for millisecond-level real-time inference locally, without relying on the cloud. Through a cloud-based incremental learning mechanism, the model can adapt to equipment aging and changes in operating conditions. This method effectively solves the problems of high cost, poor real-time performance, difficulty in edge deployment, and weak adaptability of traditional solutions, offering advantages such as low cost, high accuracy, and strong robustness.

[0158] The specific implementation of step S5 includes the following steps:

[0159] Step S51: When the confidence level of multiple consecutive diagnoses is lower than the first preset threshold, or the confidence level decreases by more than the second preset threshold, the voiceprint data collected by the edge computing unit and the corresponding annotation information are transmitted back to the cloud.

[0160] Step S52: The cloud uses small sample augmentation technology to fine-tune the local parameters of the top classification layer and feature mapping layer of the lightweight classifier model without retraining the entire model;

[0161] Step S53: Send the fine-tuned model to the edge computing unit.

[0162] During long-term operation, industrial equipment experiences gradual drift in voiceprint characteristics due to factors such as natural wear and tear of mechanical parts, adjustments to operating parameters, and changes in environmental temperature and humidity. This leads to a gradual decrease in the diagnostic confidence of the original model. If the model is not updated, the system will experience an increase in either the false alarm rate or the false negative rate. This embodiment achieves adaptive and continuous optimization of the model by setting a reasonable triggering mechanism and an efficient fine-tuning strategy.

[0163] Specifically, the triggering conditions for incremental learning include the following two scenarios:

[0164] The first type of confidence level remains consistently low. The system outputs a confidence score (a value between 0 and 1) for each diagnostic result. When the confidence level of N consecutive diagnoses (e.g., N=10) is lower than the first preset threshold (e.g., 0.70), it indicates that the model's ability to discriminate the current voiceprint features is insufficient, and there may have been changes in operating conditions or equipment aging. At this point, the above-mentioned incremental learning process is triggered.

[0165] The second type of confidence level shows a significant decrease. Under relatively stable equipment conditions, the system continuously monitors the moving average of the confidence level. When the decrease in the moving average within a preset time window (e.g., 24 hours) exceeds a second preset threshold (e.g., 0.15), it indicates that the voiceprint features are undergoing a trend shift. This also triggers the incremental learning process.

[0166] Once triggered, the edge computing unit will transmit the voiceprint data collected within a certain period before and after the trigger (e.g., 1 hour before and after), along with the annotation information confirmed by the operation and maintenance personnel (e.g., fault type, energy consumption level, etc.), back to the cloud AI database via WiFi and TCP communication protocols.

[0167] After receiving incremental data in the cloud, the model is fine-tuned using few-shot augmentation techniques. Few-shot augmentation can employ SMOTE (Synthetic Minor Oversampling) or data generation methods based on Generative Adversarial Networks (GANs) to expand the model with a small number of samples for newly added fault categories, avoiding overfitting caused by class imbalance. During fine-tuning, only the parameters of the top-level classification layer (fully connected layer) and feature mapping layer (the last 1-2 convolutional layers) are updated, while the parameters of the bottom-level feature extraction layers remain frozen. This local parameter update strategy significantly reduces computational overhead, and the fine-tuning process can be completed within minutes.

[0168] After fine-tuning, the cloud sends the updated model parameters to the edge computing unit via wireless network. Upon receiving the parameters, the edge computing unit performs a hot-swap of the model before the start of the next inference cycle, achieving seamless updates.

[0169] The following is a comparison table of incremental learning effects:

[0170] Average confidence level 0.94 0.78 0.93 Fault identification accuracy 95.2% 86.5% 94.8% False alarm rate 3.2% 12.6% 3.8% Fine-tuning time — — 5 minutes

[0171] As can be seen from the table above, the incremental learning mechanism can effectively recover from the performance degradation caused by equipment aging, enabling the system to maintain high accuracy and high robustness in the long term.

[0172] It is understood that this embodiment constructs a closed-loop mechanism for adaptive model updates by clearly defining the triggering conditions for incremental learning and the local fine-tuning strategy. This mechanism eliminates the need for manual intervention to retrain the entire model, significantly reducing model maintenance costs and enabling the system to continuously adapt to actual conditions such as equipment aging and changes in operating conditions, thereby achieving health management throughout the entire lifecycle of the equipment.

[0173] like Figure 9 As shown, this application also provides a voiceprint recognition-based edge detection system for industrial equipment to perform the above-described method embodiments. The system described below can be explained with reference to the aforementioned voiceprint recognition-based edge detection method for industrial equipment. Specifically, the system adopts an end-edge-cloud collaborative architecture, including: a signal acquisition module, a joint noise reduction module, a feature extraction and inference module, a local interaction module, a cloud update module, and a communication module; the signal acquisition module is used to acquire the voiceprint signal during the operation of the industrial equipment.

[0174] A joint noise reduction module is deployed on the edge side to perform adaptive mode decomposition on the voiceprint signal to obtain multiple intrinsic mode components. Based on the sample entropy of each intrinsic mode component, noise-dominant components are removed, and wavelet threshold noise reduction is performed on the remaining effective intrinsic mode components to reconstruct a clean voiceprint signal. A feature extraction and inference module is deployed on the edge side to extract acoustic feature vectors from the clean voiceprint signal, input them into a lightweight classifier model, and output fault diagnosis results. A local interaction module is used to display the fault diagnosis results and trigger an early warning when the diagnosis confidence exceeds a preset threshold. A cloud model update module is used to receive labeled data uploaded from the edge side, update the lightweight classifier model through incremental learning, and distribute the updated model to the edge side. A communication module is used to realize data transmission between the edge side and the cloud.

[0175] The system provided in this embodiment deploys core functional modules such as signal acquisition, joint noise reduction, feature extraction, and model inference on the edge side to achieve local real-time processing and avoid dependence on the cloud. At the same time, it deploys functions such as model updates, data storage, and visualization analysis on the cloud to achieve intelligent iteration and full lifecycle management of the system as a whole.

[0176] Specifically, the implementation methods of each module are as follows:

[0177] Signal Acquisition Module: The core utilizes a high-sensitivity miniature condenser microphone (such as the INMP441), which is non-invasively and wirelessly mounted on the surface of industrial equipment. It can acquire acoustic signals in real time without modifying the equipment itself. This device features a high sensitivity of -26dBFS and a signal-to-noise ratio of 62dB, enabling it to accurately detect subtle fault noises such as bearing wear, rotor imbalance, and loose components in complex industrial environments.

[0178] Joint noise reduction module: Deployed in the edge computing unit, it performs high-precision noise reduction processing on the acquired raw audio. First, it uses adaptive mode decomposition to decompose the signal into multiple intrinsic mode components, eliminating noise-dominant components and retaining effective fault components. Then, it uses wavelet threshold filtering to further suppress residual broadband noise. This module outputs a clean audio signal with a high signal-to-noise ratio and complete retention of fault characteristics.

[0179] Feature extraction and inference module: Deployed in the edge computing unit, it extracts acoustic feature vectors (such as 13-dimensional MFCC) from the pure voiceprint signal and inputs them into a lightweight classifier model for inference. The model uses a three-level compression technique (knowledge distillation + parameter quantization + structured pruning) to obtain the parameters, keeping the number of parameters between 10KB and 20KB, enabling millisecond-level inference to be completed locally. It outputs the fault type, abnormal energy consumption level, and confidence score.

[0180] Local interaction module: Equipped with a low-power OLED display, it displays diagnostic results in real time; when the confidence level exceeds a preset threshold, it automatically triggers an audible and visual warning.

[0181] The cloud-based model update module utilizes the Flet framework and Uvicorn ASGI server to build the backend service, achieving high-concurrency data reception and real-time push via RESTful API and WebSocket. After receiving labeled data uploaded from the edge devices, incremental model training is performed using few-shot augmentation techniques. The frontend is developed based on Vue3, TypeScript, Three.js, and ECharts, providing an integrated service for device monitoring, 3D visualization, spectrum analysis, and alarm management.

[0182] Communication module: It adopts WiFi wireless communication and TCP network protocol to realize stable, high-speed, bidirectional data transmission between edge nodes and cloud platform.

[0183] In conjunction with the above embodiments, the edge side uses ESP32 series or STM32 series microcontrollers as the core computing unit.

[0184] Specifically, the ESP32-S3 chip integrates an Xtensa LX7 dual-core processor with a clock speed of up to 240MHz, 512KB of built-in SRAM, and supports PSRAM expansion to enhance cache capacity. It also integrates WiFi and Bluetooth wireless communication modules and supports stable operation in a wide industrial temperature range of -30℃ to 75℃. The STM32 series, based on the ARM Cortex-M core, features rich peripheral interfaces and low power consumption, making it suitable for industrial scenarios with stringent power requirements.

[0185] With PSRAM enabled, the edge computing unit provides several MB of additional cache space to meet the temporary storage needs of intermediate data during the CEEMDAN decomposition process. The lightweight classifier model, after compression, requires only tens of KB of storage space and can be directly burned into the chip's built-in Flash memory. A single inference time is ≤17ms, meeting the real-time requirements of industrial environments.

[0186] It is understood that by selecting ESP32 series or STM32 series microcontrollers as the edge computing core, the system significantly reduces hardware costs and power consumption while ensuring diagnostic performance. The cost of a single hardware unit can be controlled within an extremely low range, approximately 1 / 8 of that of traditional vibration monitoring solutions, making it suitable for the low-cost and rapid deployment needs of small and medium-sized enterprises.

[0187] like Figures 10-12As shown, this application also provides an industrial equipment abnormal energy consumption detection device based on voiceprint recognition, including: a housing 10, a sound receiving component 30, a guide tube 20, a sound guiding tube 40, and a voiceprint analysis module 50. The housing 10 has a receiving chamber 14, and a sound receiving channel 112 is provided on one side wall of the housing 10. The sound receiving channel 112 has a first end near the outside and a second end 1121 near the receiving chamber 14. The second end 1121 is a closed end, and a connecting channel 113 connecting the receiving chamber 14 is provided on the closed end. The sound receiving channel 112 is connected to the outside and the receiving chamber 14 through the connecting channel 113, and the inner diameter of the connecting channel 113 is smaller than the inner diameter of the sound receiving channel 112. The sound receiving component 30 is disposed in the receiving chamber 14 and is used to collect the sound waves. The signal is converted into an electrical signal; the guide tube 20 is located in the sound receiving channel 112 to collect external sound patterns; the sound guide tube 40 is located in the accommodating chamber 14, one end of the sound guide tube 40 is connected to the connecting channel 113, and the other end is connected to the sound receiving component 30; the sound pattern analysis module 50 is located in the accommodating chamber 14, and the sound pattern analysis module 50 is electrically connected to the sound receiving component 30. The sound pattern analysis module is used to execute the aforementioned industrial equipment edge detection method based on sound pattern recognition, or the sound pattern analysis module is used to deploy the aforementioned detection system to determine whether the industrial equipment is abnormal.

[0188] To address the issues of complex structure, insufficient sound reception accuracy, and high cost in related technologies for detecting abnormal energy consumption of industrial equipment, this embodiment provides an integrated detection device that is compact, has excellent sound reception, and is low in cost. This embodiment integrates a sound receiving channel 112, a guide tube 20, a sound guide tube 40, and a sound receiving component 30 within the housing 10, forming a highly efficient acoustic transmission path. Simultaneously, the second end 1121 of the sound receiving channel 112 is designed as a closed end with only a small-diameter connecting channel 113, forming a structure similar to an acoustic collector. This effectively filters out high-frequency environmental noise, making the equipment sound signature entering the sound receiving component 30 clearer and purer.

[0189] Specifically, the structure and function of each component of this device are as follows:

[0190] The housing 10 serves as the protective outer shell and mounting base for the entire device. The housing 10 includes a main housing 11, a first connecting housing 12, and a second connecting housing 13. The first connecting housing 12 and the second connecting housing 13 are connected to the main housing 11 by screws. Figure 10 , Figure 11As shown. Of course, the second connecting housing 13 and the host / guest body can be integrally formed, creating an energy storage chamber 131. The accommodating chamber 14 formed inside the housing 10 provides a relatively enclosed and safe working environment for precision electronic components such as the receiver 30 and the acoustic signature analysis module 50, avoiding direct influence from external dust, moisture, and electromagnetic interference. The housing 10 can be made of PETG (polyethylene terephthalate) engineering plastic, which has excellent chemical corrosion resistance, impact strength, heat resistance, and dimensional stability. It can withstand common industrial environments such as dust, oil, moisture, and mechanical vibration, ensuring long-term stable operation of the device under harsh conditions.

[0191] like Figure 11 , Figure 12 As shown, the sound receiving channel 112 is located on one side wall of the housing 10, and its function is to efficiently guide sound waves generated by external devices into the device. Unlike the prior art where the sound receiving hole is directly through, in this embodiment, the sound receiving channel 112 has a closed end 1121 near the accommodating chamber 14, and a relatively small-diameter connecting channel 113 is formed on this closed end. External sound waves enter the sound receiving channel 112 from the first end near the outside, and after the sound waves are converged and initially acoustically impedance matched, they enter the accommodating chamber 14 through the connecting channel 113.

[0192] like Figure 11 As shown, the sound receiving component 30 is disposed within the accommodating chamber 14 and is used to convert the collected sound waves into electrical signals. The sound receiving component 30 may be an INMP441 miniature condenser microphone, which has a high sensitivity of -26dBFS and a high signal-to-noise ratio of 62dB, and can accurately capture subtle fault noises such as bearing wear, rotor imbalance, and loose friction in strong industrial noise environments.

[0193] The guide tube 20 is located within the sound receiving channel 112 and is used to collect external acoustic signatures. The guide tube 20 provides a smooth, low-loss propagation path for sound waves, preventing diffuse reflection within the rough walls of the housing 10 that would cause signal attenuation. The guide tube 20 is preferably made of metal or rigid plastic, with a smooth inner wall. The guide tube 20 not only serves the physical function of collecting external acoustic signatures, but its opening can also be aligned with the main sound-emitting part of the device under test, achieving directional sound reception.

[0194] like Figure 11 As shown, the sound guide tube is located within the accommodating chamber 14, with one end connected to the connecting channel 113 and the other end connected to the receiving component 30. The sound guide tube 40 directly transmits the sound waves entering through the connecting channel 113 to the receiving component 30, effectively reducing sound wave loss. The sound guide tube 40 ensures the airtightness of the sound wave transmission path, preventing sound waves from diffusing within the accommodating chamber 14 and ensuring that the receiving component 30 can receive a strong and clear sound signal.

[0195] like Figure 11 As shown, the voiceprint analysis module is located within the accommodating chamber 14 and is electrically connected to the receiving component 30. It is used to determine whether the energy consumption is abnormal based on the received voiceprint characteristics. The voiceprint analysis module 50 integrates a processor and a storage unit, and can run the voiceprint recognition and abnormal energy consumption detection methods described in the above method embodiments, or deploy the detection system described in the above system embodiments.

[0196] Understandably, by setting a closed end at the second end 1121 of the sound receiving channel 112 and only opening a small-aperture connecting channel 113, a structure similar to an acoustic collector is formed. This structure can effectively filter out some high-frequency environmental noise, making the mid-to-low frequency acoustic components with fault characteristics generated by the device itself more prominent in the sound waves entering the device. Furthermore, the small-aperture connecting channel 113, in conjunction with the subsequent sound guide tube 40, can form an efficient acoustic transmission path, reducing energy loss and reflection of sound waves during transmission and improving the signal-to-noise ratio and sensitivity of the sound reception. The overall device structure is simple and does not require connection to external sensors or detection systems, effectively reducing subsequent operating costs.

[0197] In conjunction with the above embodiments, the industrial equipment abnormal energy consumption detection device based on voiceprint recognition further includes an energy storage chamber 131, in which a battery is installed. The battery is electrically connected to the voiceprint analysis module 50 to supply power to the voiceprint analysis module 50. Further, a charging port 111 is provided on one side of the housing 10 for charging the battery. Preferably, the energy storage chamber 131 is located below the accommodating chamber 14.

[0198] The embodiment, by independently dividing the energy storage chamber 131 within the housing 10 and embedding a battery therein, enables the device to break free from dependence on an external continuous power supply line, achieving true wireless deployment and facilitating flexible arrangement in industrial sites.

[0199] Specifically, the design of the energy storage chamber 131 physically isolates the battery from the precision electronic components within the housing chamber 14. This isolation design has the following effects: it improves the safety of the internal components, ensuring that even if the battery leaks or bulges, it will not directly corrode or damage the microphone component 30 and the voiceprint analysis module 50; it facilitates thermal management, as the heat generated by the battery during charging and discharging is confined within the energy storage chamber 131, reducing the impact on the temperature-sensitive components in the voiceprint analysis module 50 and ensuring the accuracy and stability of voiceprint acquisition and analysis.

[0200] Heat dissipation holes can be provided on the side walls of the housing 10 containing the energy storage chamber 131 and the accommodating chamber 14 to prevent excessive internal temperature. Furthermore, cooling fans can be connected to the side walls of the housing 10 containing the energy storage chamber 131 and the accommodating chamber 14 to achieve active heat dissipation of the internal space.

[0201] The charging port 111 can be a Micro-USB, Type-C, or a dedicated circular charging interface, facilitating the charging of the internal battery using an external power adapter. The charging port 111, in conjunction with the rechargeable battery, constitutes a complete recyclable power supply system, greatly improving the device's ease of maintenance and long-term cost-effectiveness.

[0202] Arranging the energy storage chamber 131 below the housing chamber 14, i.e., the battery is located at the bottom of the whole unit, has the following effects: The weight of the battery accounts for a large proportion of the overall weight of the equipment. By placing the energy storage chamber 131 at the bottom, the center of gravity of the whole unit can be effectively lowered, making the device more stable when placed or installed, and less likely to tip over due to slight collisions or vibrations; at the same time, it makes the design space of the radio channel more flexible and avoids the battery itself affecting the purity of the sound waves in the housing space.

[0203] It is understood that the detection system provided in this embodiment, through a reasonable division of labor in the end-edge-cloud three-layer architecture, achieves complementary advantages between real-time edge diagnosis and intelligent iteration in the cloud. The system features low cost, low power consumption, high accuracy, and strong versatility, and can be widely used in various industrial scenarios such as factory production lines, energy stations, and municipal pipelines.

[0204] In conjunction with the above embodiments, the industrial equipment abnormal energy consumption detection device based on voiceprint recognition also includes a display 121, which is located on the top of the housing 10 and is electrically connected to the voiceprint analysis module 50 for displaying the detected information.

[0205] In a further embodiment, the inner diameter of the sound receiving channel 112 gradually increases in the direction towards the outside to form a frustum-shaped channel structure. In this embodiment, the internal space of the sound receiving channel 112 is not a cylindrical hole of equal diameter, but a frustum-shaped channel in the shape of a horn or trumpet, whose inner diameter gradually increases from the connecting channel 113 near the accommodating chamber 14 towards the outside.

[0206] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for edge detection of industrial equipment based on voiceprint recognition, characterized in that, Includes the following steps: S1. An external, non-invasive sensor collects acoustic signature signals from the operation of industrial equipment; S2. Perform adaptive mode decomposition on the voiceprint signal to obtain multiple intrinsic mode components. Based on the sample entropy of each intrinsic mode component, remove the noise-dominant component. Perform wavelet threshold denoising on the remaining effective intrinsic mode components to reconstruct a clean voiceprint signal. S3. Extract acoustic feature vectors from the pure voiceprint signal; S4. The acoustic feature vector is input into a lightweight classifier model deployed in an edge computing unit, and the fault diagnosis result is output. S5. The labeled data collected by the edge computing unit is sent back to the cloud, the lightweight classifier model is updated through incremental learning, and the updated model is sent back to the edge computing unit.

2. The industrial equipment edge detection method based on voiceprint recognition according to claim 1, characterized in that, Step S2 specifically includes the following steps: Perform CEEMDAN decomposition on the original voiceprint signal x(t) to obtain IMF1~IMF n With residual component RES; Calculate the sample entropy of each IMF component, and identify and remove IMF components whose sample entropy is greater than a preset threshold as noise-dominant components. Wavelet decomposition and soft thresholding denoising are performed on the retained effective IMF components. The reconstructed IMF components yield a clean voiceprint signal.

3. The industrial equipment edge detection method based on voiceprint recognition according to claim 2, characterized in that, Step S3 specifically includes the following steps: The pure voiceprint signal is subjected to pre-emphasis, frame windowing, short-time Fourier transform, power spectrum calculation, Mel filter bank filtering, logarithmic operation and discrete cosine transform in sequence, and the 13-dimensional MFCC feature vector is extracted as the acoustic feature vector.

4. The industrial equipment edge detection method based on voiceprint recognition according to claim 1, characterized in that, The lightweight classifier model was obtained using the following method: Using a high-precision neural network in the cloud as the teacher model and a micro neural network at the edge as the student model, the fault discrimination knowledge of the teacher model is transferred to the student model through a double-loss weighted distillation scheme. The distilled student model is subjected to parameter quantization, mapping floating-point weights to fixed-point integers; The quantized model is then subjected to structured pruning to remove redundant parameters. The compressed lightweight classifier model has 10KB-20KB of parameters, a peak memory usage of ≤50KB, and a single inference time of ≤17ms.

5. The industrial equipment edge detection method based on voiceprint recognition according to claim 1, characterized in that, Step S5 specifically includes: When the confidence level of multiple consecutive diagnoses is lower than the first preset threshold, or the decrease in confidence level exceeds the second preset threshold, the voiceprint data collected by the edge computing unit and the corresponding annotation information are sent back to the cloud. The cloud uses few-shot augmentation technology to fine-tune the local parameters of the top classification layer and feature mapping layer of the lightweight classifier model without retraining the entire model. The fine-tuned model is then sent to the edge computing unit.

6. The industrial equipment edge detection method based on voiceprint recognition according to claim 1, characterized in that, Step S2 also includes: performing FFT spectrum analysis on the reconstructed pure acoustic signal to extract characteristic energy changes in the 8kHz-12kHz frequency band for use in identifying bearing wear or rotor imbalance faults.

7. The industrial equipment edge detection method based on voiceprint recognition according to claim 1, characterized in that, The fault diagnosis results in step S4 include fault type, level, and diagnostic confidence score, wherein the level is determined based on the correspondence between fault type and equipment operating conditions.

8. A detection system for performing the voiceprint recognition-based edge detection method for industrial equipment according to any one of claims 1 to 7, characterized in that, include: The signal acquisition module is used to acquire acoustic signals from the operation of industrial equipment. A joint noise reduction module, deployed at the edge, is used to perform adaptive mode decomposition on the voiceprint signal to obtain multiple intrinsic mode components, remove noise-dominant components based on the sample entropy of each intrinsic mode component, perform wavelet threshold noise reduction on the remaining effective intrinsic mode components, and reconstruct a clean voiceprint signal. The feature extraction and inference module, deployed on the edge side, is used to extract acoustic feature vectors from the pure voiceprint signal, input them into a lightweight classifier model, and output fault diagnosis results. The local interaction module is used to display the fault diagnosis results and trigger an alert when the diagnosis confidence exceeds a preset threshold. The cloud-based model update module is used to receive labeled data uploaded from the edge side, update the lightweight classifier model through incremental learning, and then send the updated model down to the edge side. The communication module is used to enable data transmission between the edge and the cloud.

9. The detection system according to claim 8, characterized in that, The edge side uses ESP32 series or STM32 series microcontrollers as the core computing unit.

10. An industrial equipment inspection device based on voiceprint recognition, characterized in that, include: The housing has a receiving chamber inside, and a sound receiving channel is provided on one side wall of the housing. The sound receiving channel has a first end near the outside and a second end near the receiving chamber. The second end is a closed end, and a connecting channel communicating with the receiving chamber is provided on the closed end. The sound receiving channel communicates with the outside and the receiving chamber through the connecting channel, and the inner diameter of the connecting channel is smaller than the inner diameter of the sound receiving channel. A sound receiving component, located within the accommodating cavity, is used to convert the collected sound waves into electrical signals; A guide tube, located within the sound receiving channel, is used to collect external sound signatures; A sound guide tube, located within the accommodating cavity, is used for the transmission of sound patterns. One end of the sound guide tube is connected to the connecting channel, and the other end is connected to the sound receiving component. A voiceprint analysis module is disposed in the accommodating cavity. The voiceprint analysis module is electrically connected to the sound receiving component. The voiceprint analysis module is used to execute the industrial equipment edge detection method based on voiceprint recognition as described in any one of claims 1 to 7, or the voiceprint analysis module is used to deploy the detection system as described in claim 8 to determine whether the industrial equipment is abnormal.

Images