A device working state recognition method based on a voiceprint recognition model

Abstract

The application discloses a device working state recognition method based on a voiceprint recognition model and relates to the technical field of industrial device operation state recognition. The device working state recognition method based on the voiceprint recognition model collects running audio waveform data of a target device, extracts acoustic representation data containing parameters such as short-time energy, spectral centroid, spectral flux, MFCC and zero-crossing rate, inputs the acoustic representation data into a pre-trained voiceprint recognition model to extract a voiceprint feature representation vector, and determines a current running state category through a state classification model. The application divides an audio signal into frames, extracts acoustic features such as short-time energy, spectral centroid, spectral flux, MFCC and zero-crossing rate, models interframe evolution relationship in combination with a bidirectional neural network, introduces an attention mechanism to highlight key frame segments, effectively enhances the recognition capability for working conditions such as state boundary ambiguity or unclear transition, and improves the time sequence analysis and state discrimination precision.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment operating status recognition technology, specifically a method for recognizing equipment operating status based on a voiceprint recognition model. Background Technology

[0002] In industrial production, the operating status of equipment directly affects the stability, safety, and production efficiency of the system. Traditional methods of sensing operating status rely heavily on temperature sensors, vibration sensors, and current / voltage monitoring modules to assess equipment status by collecting structural or electrical change signals. However, with the development of artificial intelligence and acoustic signal processing technologies, more and more research and practice are beginning to explore using the audio waveform signals generated by equipment during operation as input for status sensing. By analyzing the acoustic signature changes of equipment under different operating conditions, non-invasive and high-precision operating status identification can be achieved.

[0003] Audio data offers advantages such as convenient acquisition, lack of structural limitations, and strong adaptability. Combined with deep learning models, it can further extract subtle differences in equipment operation, providing a more sensitive perception method for status recognition. Current research trends are gradually evolving from traditional feature engineering to deep neural network-driven voiceprint modeling methods. Through in-depth analysis of the energy distribution, spectral morphology, and temporal dynamic features in the audio of equipment operation, refined identification and classification of equipment operating status can be achieved, thus playing a crucial role in application scenarios such as predictive maintenance, intelligent inspection, and fault early warning.

[0004] The limitations of existing technologies include at least the following problems: Existing solutions often rely solely on manually set acoustic feature extraction rules or fixed feature vectors of limited dimensions, such as making rough judgments based on simple statistical parameters like mean, peak value, and energy. This approach is difficult to reflect the complex temporal changes and operating characteristics of the equipment in its operating sound. Especially under multi-state and multi-category operating conditions, it is difficult to accurately distinguish the different states corresponding to similar sound features, resulting in unstable equipment state recognition results, insufficient recognition granularity, and the risk of missed or misjudged cases. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a device operating status recognition method based on a voiceprint recognition model. This method solves the problem that existing technologies lack structured modeling and intelligent classification mechanisms for device operating audio signals, resulting in insufficient accuracy and weak generalization ability in operating status recognition.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A device operating status identification method based on a voiceprint recognition model includes the following steps: acquiring operating audio waveform data of the target device in its current operating state, wherein the operating audio waveform data includes the operating sound amplitude values ​​corresponding to several consecutive time points; preprocessing the operating audio waveform data to obtain operating acoustic representation data, wherein the operating acoustic representation data consists of operating acoustic parameters of several frames, including short-time energy, spectral centroid, spectral flux, Mel frequency cepstral coefficients, and zero crossover rate; inputting the operating acoustic representation data into a pre-trained voiceprint recognition model to extract the operating voiceprint feature representation vector of the target device; and inputting the voiceprint feature representation vector into a pre-trained state classification model for classification analysis to determine the current operating status category of the target device.

[0008] Further, the specific steps for preprocessing the running audio waveform data are as follows: noise suppression and amplitude normalization are performed on the running audio waveform data; the processed running audio waveform data is divided into frames according to a set frame length and frame shift, and a window function is applied to each frame for weighted processing to obtain several frames of running audio frame data; frequency domain transformation is performed on each frame of running audio frame data to obtain the corresponding running spectrum data; feature extraction and analysis are performed on the running spectrum data to obtain the running acoustic parameter data for each frame; the running acoustic parameter data of each frame are combined in chronological order to form running acoustic representation data.

[0009] Furthermore, the specific steps for feature extraction and analysis of the running spectrum data are as follows: The amplitude squared and accumulated for each frame of running audio data are used to obtain the corresponding short-time energy parameter; the frequency-weighted average of each frame of running spectrum data is performed to obtain the corresponding spectral centroid parameter; the amplitude difference squared and accumulated between each frame of running spectrum data and the previous frame of spectrum data are used to obtain the corresponding spectral flux parameter; Mel filter bank transform, logarithmic compression, and discrete cosine transform are performed on each frame of running spectrum data to obtain the corresponding Mel frequency cepstral coefficient parameter; the number of polarity changes in each frame of running audio data is counted to obtain the corresponding zero-crossing rate parameter.

[0010] Furthermore, the voiceprint recognition model includes an input layer, a convolutional neural network layer, a bidirectional long short-term memory network layer, an attention mechanism layer, and a feature representation layer.

[0011] Further, the specific steps for extracting the operational voiceprint feature representation vector of the target device are as follows: In the input layer of the voiceprint recognition model, operational acoustic representation data is received; in the convolutional neural network layer of the voiceprint recognition model, multiple one-dimensional convolution operations and pooling processes are performed on the operational acoustic representation data to obtain a local time and frequency joint feature map; in the bidirectional long short-term memory network layer of the voiceprint recognition model, forward and backward state propagation is performed on the local time and frequency joint feature map to obtain a bidirectional temporal feature sequence; in the attention mechanism layer of the voiceprint recognition model, attention-weighted fusion is performed on the bidirectional temporal feature sequence to obtain an aggregated feature vector after frame weight enhancement; in the feature representation layer of the voiceprint recognition model, dimensionality mapping and normalization processing are performed on the aggregated feature vector to output the operational voiceprint feature representation vector of the target device.

[0012] Furthermore, the specific steps for attention-weighted fusion of the bidirectional temporal feature sequence are as follows: similarity calculation is performed between the feature vector of each frame in the bidirectional temporal feature sequence and the global context representation to obtain the attention weight value of each frame; the feature vector of the corresponding frame is weighted based on the attention weight value to obtain the aggregated feature vector after frame weight enhancement.

[0013] Furthermore, the specific steps for dimensional mapping and normalization of the aggregated feature vector are as follows: perform fully connected mapping on the aggregated feature vector after frame weight enhancement to obtain a high-dimensional feature representation; perform L2 normalization on the high-dimensional feature representation to output the target device's operating voiceprint feature representation vector.

[0014] Furthermore, the state classification model includes an input layer, a fully connected neural network layer, and a Softmax classification layer.

[0015] Further, the specific steps for determining the current operating state category of the target device are as follows: In the input layer of the state classification model, the operating voiceprint feature representation vector of the target device is received; in the fully connected neural network layer, the operating voiceprint feature representation vector is subjected to multi-layer nonlinear mapping processing to extract state discrimination features; in the Softmax classification layer, the state discrimination features are subjected to multi-class probability normalization processing to output the current operating state category of the target device.

[0016] Furthermore, the specific steps for performing multi-class probability normalization processing on the state discrimination features are as follows: perform a dot product operation on the state discrimination features and the weight vectors of each operating state category to obtain the activation value of each category; calculate the normalized probability distribution based on the activation values ​​of all categories, and the category corresponding to the one with the highest probability is the current operating state category of the target device.

[0017] The present invention has the following beneficial effects:

[0018] (1) The device working state recognition method based on voiceprint recognition model divides the audio signal collected during the device operation into multiple short time frames in the audio processing stage. Each frame extracts key parameters representing energy distribution and spectrum changes, including short-time energy, spectral centroid, spectral flux, MFCC coefficient and zero crossover rate, which can comprehensively capture the dynamic acoustic features of the device in a local time period. Then, the evolution relationship between the preceding and following time frames is extracted using a bidirectional neural network. Then, the attention mechanism is used to highlight the frames that play a key role in the state transition process, which effectively enhances the model's ability to recognize complex working conditions such as blurred operating state boundaries and unclear transitions. Compared with the average value analysis or static classification method, the overall modeling process has higher time sequence analysis capability and state transition discrimination accuracy.

[0019] (2) The device operating status identification method based on the voiceprint recognition model directly collects audio waveform data during the device operation process and extracts acoustic parameters such as short-time energy, spectral centroid, spectral flux, MFCC and zero crossover rate based on the data. It does not rely on additional vibration sensors, current and voltage acquisition devices or special structural monitoring devices. Compared with the traditional method, which requires the installation of special hardware on the device to sense changes in status, this method can complete the high-precision perception and identification of the operating status by relying only on existing microphones or industrial microphones, which greatly reduces the cost and complexity of system deployment. It is especially suitable for small and medium-sized factories, old equipment, distributed systems and other scenarios where transformation is limited. In addition, the voiceprint features are highly correlated with the device operating status. Even weak but representative voiceprint differences can be identified by the model. Therefore, this method has strong versatility and scalability and can adapt to the operating status identification needs of different types of equipment.

[0020] (3) The device operating status identification method based on the voiceprint recognition model adopts an independently constructed status classification model, which receives the operating voiceprint feature representation vector as input, extracts the discriminative features using a multi-layer fully connected network, and outputs the current operating status category through Softmax. It has the advantages of simple structure and high training efficiency. Since the input is a highly abstract voiceprint feature vector, the classification model can achieve high discrimination accuracy under limited sample conditions and supports extended recognition of multiple status categories. Compared with the traditional method that relies on manual features or single threshold judgment, this classification model strengthens the feature discrimination through nonlinear mapping and provides clear classification confidence by combining probability output mechanism, which helps to improve the recognition stability and support subsequent device management and response decisions. It has good practical application value in multi-state device operation monitoring scenarios.

[0021] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] Figure 1 This is a flowchart of a device operating status recognition method based on a voiceprint recognition model according to the present invention.

[0023] Figure 2 This is a flowchart illustrating the specific steps involved in extracting the operating voiceprint feature representation vector of a target device using a device operating status identification method based on a voiceprint recognition model, as described in this invention.

[0024] Figure 3 This is a flowchart illustrating the specific steps of attention-weighted fusion of bidirectional temporal feature sequences in a device operating status recognition method based on a voiceprint recognition model according to the present invention. Detailed Implementation

[0025] Please see Figure 1 This invention provides a device operating status identification method based on a voiceprint recognition model, comprising the following steps: acquiring operating audio waveform data of the target device in its current operating state, the operating audio waveform data including the operating sound amplitude values ​​corresponding to several consecutive time points; preprocessing the operating audio waveform data to obtain operating acoustic representation data, the operating acoustic representation data consisting of several frames of operating acoustic parameters, including short-time energy, spectral centroid, spectral flux, Mel frequency cepstral coefficients, and zero crossover rate; inputting the operating acoustic representation data into a pre-trained voiceprint recognition model to extract the operating voiceprint feature representation vector of the target device; and inputting the voiceprint feature representation vector into a pre-trained state classification model for classification analysis to determine the current operating status category of the target device.

[0026] Specifically, the preprocessing steps for the running audio waveform data are as follows: noise suppression and amplitude normalization are performed on the running audio waveform data. Specifically, noise suppression algorithms such as spectral subtraction or Wiener filtering are used to process the original audio signal to reduce the interference of background noise on audio feature extraction. Then, the processed audio signal is normalized, that is, by dividing each sampling point in the audio waveform by the maximum amplitude value, the overall amplitude range is limited to [-1, 1] to enhance feature stability.

[0027] The processed audio waveform data is divided into frames according to a set frame length and frame shift. Each frame is then weighted using a window function to obtain several frames of audio data. Specifically, the audio signal data at consecutive time points is segmented into segments with a set frame length (e.g., 25ms), and a frame shift (e.g., 10ms) is set between each segment to form several overlapping frames. Each frame is then weighted using a window function such as a Hamming window or a Hanning window to reduce the impact of frame boundary effects on the spectrum analysis.

[0028] Each frame of running audio frame data is processed by frequency domain transformation to obtain the corresponding running spectrum data. Specifically, Fast Fourier Transform (FFT) is applied to the windowed audio frame data of each frame to transform it from the time domain to the frequency domain, and a complex spectrum representing the frequency composition of the frame is obtained for subsequent feature extraction.

[0029] Feature extraction and analysis are performed on the running spectrum data to obtain the running acoustic parameter data for each frame;

[0030] The acoustic parameters of each frame are combined in chronological order to form the acoustic representation data. Specifically, the acoustic parameters (such as short-time energy, spectral centroid, etc.) calculated for each frame are spliced ​​together in chronological order to form a two-dimensional time series feature matrix, which serves as the acoustic representation data for subsequent model analysis.

[0031] In this implementation scheme, classic noise suppression algorithms such as spectral subtraction or Wiener filtering are used to effectively eliminate the interference of background environmental noise, making the retained audio signal more accurately reflect the operating status of the device itself. Secondly, through amplitude normalization processing, the dynamic range of the audio waveform is standardized to [-1, 1], which enhances the feature stability and model generalization ability across devices and scenarios. Thirdly, reasonable framing of the audio signal is achieved by setting frame length and frame shift, and window function weighting methods such as Hamming window are introduced to avoid spectral distortion caused by frame boundary discontinuities, laying a stable foundation for frequency domain analysis. Subsequently, the frame data is mapped from the time domain to the frequency domain using fast Fourier transform, providing necessary support for the extraction of multi-dimensional acoustic parameters. Finally, the acoustic parameters of each frame are integrated into a two-dimensional time series matrix in chronological order, constructing a complete and clearly structured acoustic feature expression for subsequent neural network input, which helps the system to more accurately identify the device's operating status.

[0032] Specifically, the specific steps for feature extraction and analysis of the operating spectrum data are as follows: the amplitude squared and accumulated for each frame of operating audio frame data to obtain the corresponding short-time energy parameter, specifically: for each sampling point in each frame, its amplitude value is squared and summed to represent the energy intensity of the signal in that frame, reflecting the characteristics of sound intensity change during the operation of the equipment;

[0033] The frequency-weighted average of each frame of running spectrum data is processed to obtain the corresponding spectral centroid parameter. Specifically, the frequency value of each frequency component in the spectrum of the frame is multiplied by its corresponding amplitude value (or power value), the weighted average of all components is calculated, and then divided by the total amplitude or total power to obtain the spectral center position, which reflects the centroid of the frequency distribution.

[0034] The amplitude difference between each frame of the running spectrum data and the previous frame of the spectrum data is squared and accumulated to obtain the corresponding spectral flux parameter. Specifically, the spectrum of the current frame is subtracted from the spectrum of the previous frame one by one according to the frequency components, the square of the difference is calculated and summed to characterize the amplitude of the inter-frame spectrum change and reflect the dynamics of the equipment operation.

[0035] For each frame of running spectrum data, Mel filter bank transformation, logarithmic compression, and discrete cosine transform are performed to obtain the corresponding Mel frequency cepstral coefficient parameters. Specifically, the spectrum signal is filtered through a set of Mel filter banks to obtain the Mel frequency band energy. After logarithmic operation to compress the dynamic range, the first few coefficients are extracted by applying discrete cosine transform (DCT) as the MFCC parameters of the frame, reflecting the auditory perception characteristics of the sound.

[0036] The number of polarity changes for each frame of audio data is counted to obtain the corresponding zero crossover rate parameter. Specifically, the number of times the signal polarity changes from positive to negative or from negative to positive in the audio sampling points of that frame is counted to reflect high-frequency noise or vibration components, which is suitable for distinguishing working noise in different states.

[0037] In this implementation scheme, the short-time energy parameter can directly reflect the fluctuation of the sound intensity of the equipment under different operating conditions, which helps to identify the energy difference caused by load changes or power fluctuations. The spectral centroid parameter provides the centroid information of the sound frequency distribution, which can reveal the offset characteristics of the main frequency band in the operating state. The spectral flux parameter quantifies the degree of spectral change between frames, which can capture the dynamic characteristics during state switching or abnormal operating conditions. The Mel frequency cepstral coefficient (MFCC) extracts the key auditory features of speech by simulating the human ear perception mechanism, which is suitable for characterizing the differences in voiceprints under different equipment states. The zero crossover rate, as a simple measure of high-frequency components, can effectively reflect the changes in equipment vibration or noise levels. These parameters complement each other, covering both energy intensity and frequency distribution, and taking into account dynamic changes and auditory perception characteristics.

[0038] Specifically, the voiceprint recognition model includes an input layer, a convolutional neural network layer, a bidirectional long short-term memory network layer, an attention mechanism layer, and a feature representation layer.

[0039] The pre-training steps for the voiceprint recognition model are as follows:

[0040] The training samples used for pre-training come from the operating audio data of various types of industrial equipment under different working conditions. The audio data is collected in industrial settings and manually labeled with corresponding equipment identification tags, forming a set of data pairs containing audio data and equipment identification information. The following data processing flow is then performed on the raw operating audio data:

[0041] Audio preprocessing: Noise reduction and amplitude normalization are performed on the original audio waveform data to reduce background noise interference and unify the dynamic range;

[0042] Framing and Window Function Weighting: The audio waveform is divided into frames according to a set frame length (e.g., 25ms) and frame shift (e.g., 10ms), and Hamming window weighting is applied to each frame;

[0043] Frequency domain transformation and acoustic parameter extraction: Perform Fast Fourier Transform (FFT) on each frame of audio data and extract the acoustic parameters of each frame, including short-time energy, spectral centroid, spectral flux, Mel frequency cepstral coefficients (MFCC) and zero crossover rate, etc.

[0044] Sample construction: Acoustic parameters from several consecutive frames are combined in chronological order to form running acoustic representation data, which is used as input to the model. The labels are the corresponding device numbers.

[0045] The model structure includes an input layer, a convolutional neural network (CNN) layer, a bidirectional long short-term memory (BiLSTM) network layer, an attention mechanism layer, and a feature representation layer. The training objective is to learn to distinguish voiceprint features between different devices.

[0046] The training adopts a supervised learning approach, using the cross-entropy loss function to measure the difference between the model's predicted device category and the true label. Backpropagation and gradient descent are performed through the Adam optimizer to optimize parameters. During training, the number of training epochs (e.g., 100 epochs), batch size (e.g., 64), and initial learning rate (e.g., 0.001) are set. A validation set is used for performance evaluation and an early stopping mechanism is implemented to prevent overfitting.

[0047] After training, the model can output a distinctive voiceprint feature vector based on the input acoustic representation data, which serves as the input basis for subsequent state recognition.

[0048] like Figure 2 As shown, the specific steps for extracting the operating voiceprint feature representation vector of the target device are as follows: In the input layer of the voiceprint recognition model, the operating acoustic representation data is received;

[0049] In the convolutional neural network layer of the voiceprint recognition model, multiple one-dimensional convolution operations and pooling processes are performed on the acoustic representation data to obtain a local time and frequency joint feature map. Specifically, the acoustic representation data is used as a one-dimensional input tensor, and one-dimensional convolution operations are performed sequentially through multiple convolution kernels to extract local time patterns. Each convolution is followed by a pooling operation (such as max pooling or average pooling) to reduce the feature dimension and enhance translation invariance. The output feature map contains local time-frequency joint representations of multiple channels.

[0050] In the bidirectional long short-term memory network layer of the voiceprint recognition model, forward and backward state propagation is performed on the local time and frequency joint feature map to obtain a bidirectional temporal feature sequence. Specifically, forward and backward LSTM networks are used to propagate the state of the convolutional features at each time step. The forward propagation captures past temporal dependencies, and the backward propagation models future dependencies. Finally, the outputs of the two are concatenated according to time steps to form an enhanced temporal context feature sequence.

[0051] In the attention mechanism layer of the voiceprint recognition model, attention-weighted fusion of bidirectional temporal feature sequences is performed to obtain an aggregated feature vector with enhanced frame weights.

[0052] In the feature representation layer of the voiceprint recognition model, the aggregated feature vector is subjected to dimensional mapping and normalization processing to output the operating voiceprint feature representation vector of the target device.

[0053] In this implementation, by sequentially inputting the acoustic representation data into a convolutional neural network, a bidirectional long short-term memory network, an attention mechanism layer, and a feature representation layer, this method can effectively extract the voiceprint feature vectors corresponding to the target device under different operating states. During processing, the convolutional neural network can extract local time-frequency feature patterns in audio frames and capture short-term acoustic changes related to the state. The bidirectional long short-term memory network further models the dependencies between different parts of the voiceprint sequence, enabling the model to identify state change trends over a long time scale. The attention mechanism performs weighted fusion of key frames, highlighting the acoustic segments most closely related to the state, thereby improving the overall recognition accuracy. The final output voiceprint feature representation vector has strong stability and high discriminative power. The entire processing structure, while maintaining the compactness and efficiency of the recognition model, significantly enhances the ability to perceive subtle acoustic differences under different device states, which helps to achieve accurate identification of operating states. It is especially suitable for scenarios with high requirements for accuracy and robustness in industrial equipment condition detection.

[0054] Specifically, such as Figure 3 As shown, the specific steps for attention-weighted fusion of bidirectional temporal feature sequences are as follows: The similarity between the feature vector of each frame in the bidirectional temporal feature sequence and the global context representation is calculated to obtain the attention weight value for each frame. Specifically, by defining an attention scoring function (such as dot product, additive attention, etc.), the similarity between the feature vector of each frame and the global feature mean or hidden state is calculated as the importance score of that frame; the scores of all frames are normalized and used as attention weights for weighted aggregation.

[0055] The feature vectors of the frames corresponding to the attention weight values ​​are weighted to obtain the aggregated feature vector after frame weight enhancement. Specifically, the feature vector of each frame is multiplied by its corresponding attention weight, and then a weighted summation operation is performed on all frames to generate the aggregated feature vector. This preserves the significant features of key frames in the time series and improves the state discrimination effect.

[0056] In this implementation scheme, by introducing an attention-weighted fusion mechanism into the bidirectional temporal feature sequence, the system can effectively improve its ability to express key acoustic features. Specifically, by calculating the similarity between the feature vector of each frame and the global context representation, the attention weight of each frame is obtained. Thus, in the feature aggregation process, frames with richer state discrimination information are given higher weights. This weighting strategy avoids the dilution of key information by traditional average pooling or last-moment extraction methods, and helps the model focus on those frames that contain significant changes in equipment state. By multiplying the feature vector of each frame with its weight value and summing them to generate an aggregated feature vector, the model's ability to respond to instantaneous acoustic anomalies or changes in the operating state is further enhanced. This aggregated feature takes into account both local details and global structure, enabling the subsequent state recognition model to obtain higher classification accuracy with low redundancy input. It is especially suitable for processing complex fields in industrial equipment audio where there is noise interference, sudden state changes, or differences in operating stability.

[0057] Specifically, the steps for dimensional mapping and normalization of the aggregated feature vector are as follows: Perform fully connected mapping on the aggregated feature vector after frame weight enhancement to obtain a high-dimensional feature representation. Specifically, the aggregated vector is input into a fully connected layer (i.e., a fully connected neural network), and the dimension is increased by transforming the weight matrix. At the same time, a non-linear activation function (such as ReLU) is introduced to enhance the representation capability and extract richer hidden features.

[0058] The high-dimensional feature representation is subjected to L2 normalization to output the voiceprint feature representation vector of the target device. Specifically, the norm is obtained by summing the squares of all elements in the high-dimensional vector and taking the square root. Then, each element of the original vector is divided by the norm to achieve unity norm normalization.

[0059] In this implementation scheme, by performing dimensional mapping and normalization on the aggregated feature vector after frame weight enhancement, the discriminative ability and stability of the subsequent state classification model can be significantly improved. Specifically, by using a fully connected layer to perform high-dimensional mapping on the aggregated vector and introducing a nonlinear activation function (such as ReLU), it is helpful to capture complex nonlinear relationships, further explore potential discriminative features in the voiceprint data, and enhance the model's ability to express differences between devices. Subsequently, L2 normalization is performed to limit the distribution of different samples in the feature space to a uniform scale range, avoid the interference of feature amplitude differences on the classification results, and improve the robustness and convergence speed of the model in tasks such as similarity calculation. This feature representation not only retains key speech structure information but also has good geometric distribution characteristics, which can be used as a high-quality input for the state classification module to support the accurate identification of complex industrial equipment states.

[0060] Specifically, the state classification model includes an input layer, a fully connected neural network layer, and a Softmax classification layer.

[0061] The pre-training steps for the state classification model are as follows:

[0062] The training samples used in the pre-training are audio data records of multiple devices in known working conditions. After the audio data is processed by the voiceprint recognition model, it is extracted into a voiceprint feature representation vector of the device operation. The corresponding device operating status category (such as "idling", "load", "fault" etc.) is labeled by the manual or existing working condition system, forming a set of data pairs containing voiceprint features and status labels.

[0063] The model structure includes an input layer, several fully connected neural network layers, and a softmax classification layer. The pre-training steps are as follows:

[0064] Data preparation: The running voiceprint feature representation vector output by the voiceprint recognition model is used as the input feature and paired with the labeled running status category;

[0065] Model input: The feature vector is input into the input layer, and after passing through a fully connected network with multiple nonlinear mappings, the state discrimination features are extracted;

[0066] Classification output: The state discrimination features are mapped to a multi-class probability distribution through a Softmax layer;

[0067] Training objective and loss function: The multi-class cross-entropy loss function is used to measure the difference between the output state probability distribution of the model and the true state label;

[0068] Optimization method: Use the Adam optimizer for backpropagation and parameter update, set the initial learning rate (e.g., 0.001), training batch (e.g., 128), and number of iterations (e.g., 80-150), and introduce the Dropout mechanism (e.g., 0.5) to suppress overfitting;

[0069] Training process control: Use a training set to validation set ratio (e.g., 8:2) to track training performance. Stop training early and save the optimal model weights once the validation set accuracy reaches a stable level.

[0070] The final model has the ability to identify the operating status of any voiceprint feature representation vector and can be deployed in industrial control equipment or edge computing modules for real-time working status monitoring.

[0071] The specific steps to determine the current operating state category of the target device are as follows: In the input layer of the state classification model, the operating voiceprint feature representation vector of the target device is received;

[0072] In the fully connected neural network layer, the voiceprint feature representation vector is subjected to multi-layer nonlinear mapping to extract state discrimination features. Specifically, the input vector is abstracted layer by layer through a multi-layer fully connected network, each layer consisting of linear transformation and nonlinear activation function, and finally outputs a high-dimensional feature representation for state classification.

[0073] In the Softmax classification layer, multi-class probability normalization is performed on the state discrimination features to output the current operating state category of the target device.

[0074] In this implementation scheme, by inputting the voiceprint feature representation vector into the state classification model and performing hierarchical discriminant calculation, the current operating state of the equipment can be accurately identified. Specifically, the state classification model first receives the voiceprint feature vector at the input layer to ensure that subsequent processing has a complete acoustic state information foundation. Then, in the fully connected neural network layer, a deep feature abstraction channel is constructed through multi-layer linear mapping and non-linear activation functions, which effectively improves the model's ability to express complex state patterns. It is particularly suitable for situations where state differences are weak but continuous in industrial equipment. Finally, the Softmax classification layer achieves multi-class normalized output, mapping the model output to the probability distribution of each possible state, thereby determining the category of the current operating state of the equipment.

[0075] Specifically, the steps for performing multi-class probability normalization on the state discrimination features are as follows:

[0076] The dot product operation is performed on the state discrimination feature and the weight vector of each running state category to obtain the activation value of each category. Specifically, the dot product operation is performed on the state discrimination feature vector and the preset classification weight matrix to obtain the linear activation value corresponding to each category, which represents the similarity between the current input and the center of each category.

[0077] A normalized probability distribution is calculated based on the activation values ​​of all categories. The category corresponding to the highest probability is the current operating state category of the target device. Specifically, the activation values ​​are exponentially normalized using the Softmax function to convert them into a probability distribution of all categories. The category corresponding to the highest probability is the recognition result, representing the current operating state of the device.

[0078] In this implementation scheme, by performing multi-class probability normalization processing on the state discrimination features, clear division and high-confidence identification of device operating states can be achieved. This processing first calculates the linear activation value of each category by performing a dot product operation between the state discrimination features and the weight vectors of each state category, quantifying the closeness of the current voiceprint features to each state pattern, providing a basis for multi-class discrimination. Then, the Softmax function is used to normalize all activation values, transforming them into an interpretable probability distribution, so that the model output has clear statistical significance. Finally, the category corresponding to the highest probability is taken as the current operating state, which significantly improves the robustness and practicality of state discrimination. This method can not only adapt to the differences in state distribution between different devices, but also effectively handle complex situations such as boundary states and slight state drift.

[0079] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0080] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for identifying the working status of a device based on a voiceprint recognition model, characterized in that, Includes the following steps: Collect the operating audio waveform data of the target device in its current operating state. The operating audio waveform data includes the operating sound amplitude values ​​corresponding to several consecutive time points. The running audio waveform data is preprocessed to obtain running acoustic representation data, which consists of several frames of running acoustic parameters, including short-time energy, spectral centroid, spectral flux, Mel frequency cepstral coefficients, and zero crossover rate. The operational acoustic representation data is input into a pre-trained voiceprint recognition model to extract the operational voiceprint feature representation vector of the target device; The voiceprint feature representation vector is input into a pre-trained state classification model for classification analysis to determine the current operating state category of the target device. The voiceprint recognition model includes an input layer, a convolutional neural network layer, a bidirectional long short-term memory network layer, an attention mechanism layer, and a feature representation layer; The specific steps for extracting the operational acoustic signature feature vector of the target device are as follows: In the input layer of the voiceprint recognition model, acoustic representation data is received. In the convolutional neural network layer of the voiceprint recognition model, multiple one-dimensional convolution operations and pooling processes are performed on the acoustic representation data to obtain a local temporal and frequency joint feature map. In the bidirectional long short-term memory network layer of the voiceprint recognition model, forward and backward state propagation is performed on the local time and frequency joint feature map to obtain a bidirectional temporal feature sequence. In the attention mechanism layer of the voiceprint recognition model, attention-weighted fusion of bidirectional temporal feature sequences is performed to obtain an aggregated feature vector with enhanced frame weights. In the feature representation layer of the voiceprint recognition model, the aggregated feature vector is subjected to dimensional mapping and normalization to output the operating voiceprint feature representation vector of the target device. The specific steps for attention-weighted fusion of bidirectional temporal feature sequences are as follows: The similarity between the feature vector of each frame in the bidirectional temporal feature sequence and the global context representation is calculated to obtain the attention weight value of each frame; The feature vectors of the frames corresponding to the attention weight values ​​are weighted to obtain the aggregated feature vectors after frame weight enhancement. The specific steps for performing dimension mapping and normalization on the aggregated feature vector are as follows: A fully connected mapping is performed on the aggregated feature vector after frame weight enhancement to obtain a high-dimensional feature representation; The high-dimensional feature representation is subjected to L2 normalization, and the output is the voiceprint feature representation vector of the target device.

2. The device operating status identification method based on voiceprint recognition model according to claim 1, characterized in that, The specific steps for preprocessing the running audio waveform data are as follows: Noise suppression and amplitude normalization are performed on the running audio waveform data; The processed audio waveform data is divided into frames according to the set frame length and frame shift, and a window function is applied to each frame for weighted processing to obtain several frames of audio frame data. Each frame of audio data is processed by frequency domain transformation to obtain the corresponding runtime spectrum data; Feature extraction and analysis are performed on the running spectrum data to obtain the running acoustic parameter data for each frame; The running acoustic parameter data of each frame are combined in chronological order to form running acoustic representation data.

3. The device operating status identification method based on a voiceprint recognition model according to claim 2, characterized in that, The specific steps for feature extraction and analysis of the running spectrum data are as follows: The amplitude squares of each running audio frame data are accumulated to obtain the corresponding short-time energy parameters; The frequency-weighted average of each frame of running spectrum data is processed to obtain the corresponding spectral centroid parameters; The amplitude difference between each frame of running spectrum data and the previous frame of spectrum data is squared and accumulated to obtain the corresponding spectral flux parameter; Perform Mel filter bank transform, logarithmic compression and discrete cosine transform on each frame of running spectrum data to obtain the corresponding Mel frequency cepstral coefficient parameters; The number of polarity changes is counted for each frame of audio data to obtain the corresponding zero crossover rate parameter.

4. The device operating status identification method based on a voiceprint recognition model according to claim 1, characterized in that, The state classification model includes an input layer, a fully connected neural network layer, and a Softmax classification layer.

5. The device operating status identification method based on a voiceprint recognition model according to claim 4, characterized in that, The specific steps to determine the current operating status category of the target device are as follows: In the input layer of the state classification model, the target device's operating voiceprint feature representation vector is received; In the fully connected neural network layer, the voiceprint feature representation vector is subjected to multi-layer nonlinear mapping to extract state discrimination features; In the Softmax classification layer, multi-class probability normalization is performed on the state discrimination features to output the current operating state category of the target device.

6. The device operating status identification method based on a voiceprint recognition model according to claim 5, characterized in that, The specific steps for performing multi-class probability normalization on the state discrimination features are as follows: The activation value for each category is obtained by performing a dot product operation between the state discrimination features and the weight vectors of each running state category; A normalized probability distribution is calculated based on the activation values ​​of all categories, and the category corresponding to the highest probability is the current operating state category of the target device.

Images