Shovel mechanical room anomaly detection method based on audio signal analysis, storage medium and computer device

Abstract

The application discloses a kind of based on audio signal analysis's electric shovel mechanical room abnormality detection method, storage medium and computer equipment, the method includes: through the pickup deployed in electric shovel mechanical room gathers actual audio signal when actually running in electric shovel mechanical room;Actual audio signal is sent to edge computing platform by audio acquisition interface;Edge computing platform carries out feature extraction to actual audio signal, to obtain actual audio feature, and reconstructs actual audio feature by pre-training sound reconstruction model, to obtain actual reconstruction feature, wherein, pre-training sound reconstruction model is based on original audio signal when training obtained in the normal operating state of electric shovel mechanical room;Edge computing platform calculates the reconstruction error of actual reconstruction feature according to actual audio feature, whether equipment abnormality occurs based on reconstruction error and adaptive threshold value is identified, and equipment abnormality identification result is sent to remote monitoring terminal by industrial switch.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring technology for construction machinery, and in particular to an anomaly detection method, storage medium and computer equipment for electric shovel machine room based on audio signal analysis. Background Technology

[0002] Electric shovels are core equipment in large-scale open-pit mining operations. Their machine compartments integrate dozens of key electromechanical devices, including air compressor units, swing motors, cable reel mechanisms, hoist mechanisms, and crowd mechanisms. The machine compartments typically exceed 100 square meters in area, featuring high equipment density and complex operating conditions. Various mechanical and electrical devices operate continuously under harsh conditions of high load, high vibration, and high dust levels.

[0003] Currently, equipment status monitoring in electric shovel machine rooms mainly relies on manual inspections and experience-based judgment. Maintenance personnel need to periodically enter the machine room to assess the equipment's operating status through subjective methods such as auditory perception and tactile judgment. This traditional maintenance model has the following significant drawbacks: poor real-time detection, making 24 / 7 continuous monitoring impossible; high reliance on the professional experience of maintenance personnel, with personnel turnover leading to unstable detection capabilities; the high-noise, high-temperature, and dusty environment of the machine room posing a threat to personnel health; and difficulty in ensuring the consistency and repeatability of subjective judgments, easily leading to missed detections and misjudgments. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method for detecting abnormalities in the mechanical chamber of an electric shovel based on audio signal analysis, a storage medium, and a computer device.

[0005] According to one aspect of this application, a method for detecting anomalies in the mechanical compartment of an electric shovel based on audio signal analysis is provided, the method comprising: The actual audio signal of the electric shovel during actual operation is collected by a microphone deployed inside the shovel's machine room; the actual audio signal is then sent to the edge computing platform via an audio acquisition interface. The edge computing platform extracts features from the actual audio signal to obtain actual audio features, and reconstructs the actual audio features using a pre-trained sound reconstruction model to obtain actual reconstructed features. The pre-trained sound reconstruction model is trained based on the original audio signal during normal operation of the electric shovel machine. The edge computing platform calculates the reconstruction error of the actual reconstruction features based on the actual audio features, identifies whether a device malfunction has occurred based on the reconstruction error and an adaptive threshold, and sends the device malfunction identification result to the remote monitoring terminal through an industrial switch. The remote monitoring terminal displays the results of the device anomaly identification.

[0006] According to another aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for detecting abnormalities in the electric shovel's mechanical chamber based on audio signal analysis.

[0007] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described method for detecting abnormalities in the electric shovel's mechanical chamber based on audio signal analysis.

[0008] By employing the above technical solutions, this application provides an anomaly detection method, storage medium, and computer equipment for electric shovel machinery rooms based on audio signal analysis. Through the localized deployment of microphones and edge computing platforms, it achieves real-time acquisition and on-site processing of acoustic signals from electric shovel machinery rooms, avoiding the bandwidth pressure caused by uploading massive amounts of audio data. By extracting log-Mel spectrum features and calculating the reconstruction error of a pre-trained sound reconstruction model, it enables the establishment of a detection baseline with only normal samples, solving the training bottleneck of scarce abnormal samples in industrial scenarios. Through an adaptive threshold mechanism based on gamma distribution, it achieves automatic adaptation of anomaly judgment boundaries without the need for manual experience intervention, providing a technical solution for monitoring the status of electric shovel machinery rooms that requires no abnormal samples, is highly adaptive, and can be deployed in real time.

[0009] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating an anomaly detection method for the mechanical compartment of an electric shovel based on audio signal analysis, provided in an embodiment of this application, is shown. Figure 2 A flowchart illustrating another method for detecting anomalies in the mechanical chamber of an electric shovel based on audio signal analysis, provided in an embodiment of this application, is shown. Figure 3 A flowchart illustrating another method for detecting anomalies in the mechanical chamber of an electric shovel based on audio signal analysis, provided in an embodiment of this application, is shown. Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0011] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0012] This embodiment provides a method for detecting anomalies in the mechanical compartment of an electric shovel based on audio signal analysis, such as... Figure 1 As shown, the method includes: Step 101: Collect the actual audio signal of the electric shovel during actual operation by using a microphone deployed in the electric shovel's machine room; send the actual audio signal to the edge computing platform through the audio acquisition interface.

[0013] Among them, the microphone is a kind of acoustic-electric transducer that converts air vibrations into analog electrical signals. It contains an electret capacitor sensing unit, which has linear response characteristics to sound pressure changes over a wide frequency range. This characteristic makes it suitable for complex sound field environments with multiple sound sources mixing in the electric shovel machinery room. The edge computing platform is an embedded computer device deployed on-site in the electric shovel machinery room, which has data acquisition and real-time inference capabilities. It includes an ARM architecture processor, a neural network acceleration unit, and an industrial-grade audio acquisition interface. It realizes on-site processing of sensor data and result reporting through local computing and communication functions.

[0014] Specifically, the microphone is first mounted on the top center of the electric shovel's machine room using a magnetic base. This location is chosen based on the spatial distribution characteristics of the equipment inside the machine room, ensuring even reception of radiated sound energy from equipment in all directions. Then, the microphone's analog output is connected to the edge computing platform's audio acquisition interface via a shielded audio cable. This interface includes a preamplifier and an anti-aliasing filter for level adaptation of weak signals. When the microphone acquires the actual audio signal from the machine room during operation, the audio acquisition interface digitizes the signal at a preset sampling rate and transmits the converted digital audio data to the edge computing platform's memory buffer via an internal bus. If the input signal amplitude exceeds the interface's range during signal acquisition (e.g., due to clipping caused by a sudden impact sound), the gain control circuit is automatically activated to attenuate the input to a safe range, preventing errors in subsequent processing units due to data overflow.

[0015] Step 102: The edge computing platform extracts features from the actual audio signal to obtain actual audio features, and reconstructs the actual audio features using a pre-trained sound reconstruction model to obtain actual reconstructed features. The pre-trained sound reconstruction model is trained based on the original audio signal during normal operation of the electric shovel machine indoors.

[0016] In one optional implementation, feature extraction of the original audio signal to obtain original audio features includes: normalizing the original audio signal; dividing the normalized original audio signal into frames; applying a Hanning window to each frame; performing a Fourier transform on each frame to convert the time-domain waveform into the frequency domain; taking the square of the amplitude of the spectrum to obtain the power spectrum; and superimposing the power spectra of all frames in time order to obtain a time-frequency graph; extracting the Mel spectrum from the time-frequency graph using a Mel filter bank to obtain a Mel spectrum matrix; and taking the logarithm of each element in the Mel spectrum matrix to obtain a logarithmic Mel spectrum matrix; and, based on the logarithmic Mel spectrum matrix, indexing each time frame and stacking the logarithmic Mel spectrum vectors of itself and the two preceding and following frames into three channels to obtain the original audio features.

[0017] Among them, the pre-trained sound reconstruction model refers to an encoder-decoder symmetric structure built on a deep convolutional neural network. This model is trained only on the original audio signal under normal operating conditions of the electric shovel machinery room. By minimizing the reconstruction error between the input and output, the encoder learns to compress normal acoustic features into a low-dimensional latent space representation, and the decoder learns to recover the original features from this latent representation. Normalization is a preprocessing operation that scales the amplitude range of the original audio signal to a uniform interval. It is achieved by calculating the maximum absolute value in the signal sequence and dividing all sampling points by this maximum value. This processing can eliminate amplitude differences caused by different recording devices or different gain settings. The Hanning window is a window function that is high in the middle and gradually approaches zero at both ends. Its mathematical form is the cosine square function. This window function has smooth edges in the time domain and is used to weight each frame of the signal after framing. Its special performance is that it can effectively suppress spectral leakage and make the frequency components after Fourier transform more concentrated. The Mel filter bank refers to a set of filters that are equally spaced according to the Mel frequency scale. A triangular array of bandpass filters, each covering a continuous linear frequency range, outputs a weighted sum of the power spectra within that range. This filter bank simulates the nonlinear resolution characteristics of the basilar membrane of the human ear for different frequencies. The log-Mel spectrum matrix is ​​a two-dimensional array obtained by applying the natural logarithm to each element of the Mel spectrum matrix. This transformation can compress the dynamic range of signal energy, making the characteristic differences in low-energy regions relatively amplified. Three-channel stacking refers to splicing the log-Mel spectrum vectors of three consecutive time frames along a new dimension to form a tensor structure with three channels, each corresponding to the spectral characteristics of the previous, current, and next frames, respectively.

[0018] In this step, the original audio signal is first read into the memory of the edge computing platform as a digital sequence. Normalization is performed on all sampling points in the sequence, and the maximum absolute value is calculated. Then, the value of each sampling point is divided by this maximum value, ensuring that the amplitude range of the entire sequence falls within a closed interval from -1 to +1. After normalization, the signal sequence is framed according to a preset frame length and frame shift. Each frame contains a fixed number of consecutive sampling points, and adjacent frames overlap. For each frame, it is multiplied point-by-point by a Hanning window function to obtain a windowed frame sequence. Subsequently, a Fast Fourier Transform is performed on each windowed frame to convert the time-domain waveform into a complex frequency spectrum. The square of the modulus of the complex result at each frequency point is calculated to obtain the power spectrum of that frame. The power spectra of all frames are arranged sequentially in time and stacked into a two-dimensional time-frequency graph matrix, where the horizontal axis represents the time frame number and the vertical axis represents the frequency interval. After obtaining the time-frequency graph, it is input into a set of Mel-triangle filters. Each filter covers a linear frequency range, outputting a weighted sum of power spectral values ​​within the corresponding frequency interval. Iterating through all filters yields a Mel spectrum matrix with dimensions equal to the number of Mel bands multiplied by the number of time frames. The natural logarithm (or common logarithm) of each element in this matrix is ​​taken. When an element is zero, a very small constant is added before taking the logarithm again, resulting in a logarithmic Mel spectrum matrix. Finally, based on the logarithmic Mel spectrum matrix, for each internal time frame index (excluding the first and last frames), logarithmic Mel spectrum vectors for the previous, current, and next frames are extracted. Each vector has the same dimension as the number of Mel bands. These three vectors are stacked along a new channel dimension to form a three-dimensional tensor with a shape equal to the number of channels multiplied by the number of Mel bands multiplied by the number of time frames. For the first and last frames on the time axis, since a previous or next frame is missing, adjacent frames are copied to fill the missing channels, ensuring that all time positions output a three-channel tensor with the same dimensions.

[0019] After feature extraction, the feature matrix is ​​fed as an input tensor into a pre-trained sound reconstruction model. For example, the sound reconstruction model includes an encoder and a decoder. The encoder compresses the original audio features into a low-dimensional latent space representation, and the decoder restores the low-dimensional latent space representation output by the encoder to reconstructed sound features with the same dimension as the original audio features. The encoder sequentially includes an input layer, an initial convolutional layer, multiple inverse residual blocks, a bottleneck layer, a global average pooling layer, and a fully connected layer; the decoder sequentially includes an initial fully connected layer, multiple upsampling blocks, a final output convolutional layer, and an output activation layer. During the model's forward propagation, the encoder compresses the spatial size and number of channels of the input tensor layer by layer through multiple layers of depthwise separable convolutions and inverse residual blocks, ultimately outputting a low-dimensional latent space feature vector. The decoder receives this latent vector and restores the spatial size of the feature map layer by layer through transposed convolutions or upsampling interpolation operations, ultimately outputting a reconstructed feature matrix with the same dimension as the actual audio features. When the temperature of the computing unit exceeds the safety threshold during model inference, the operating frequency of the neural network acceleration unit is automatically reduced, and the model quantization inference mode is enabled. This sacrifices a small amount of reconstruction accuracy for continuous operation, preventing the edge computing platform from shutting down due to overheating. Normalization in this step ensures that audio signals from different sources have a consistent amplitude reference, avoiding interference from volume differences in subsequent feature extraction. Hanning windowing and Fourier transform accurately convert the time-domain waveform into a frequency-domain representation, suppressing spurious frequency components caused by spectral leakage. Mel filter banks and logarithmic compression ensure that the feature representation conforms to human auditory perception, and effectively enhances early fault symptoms of low energy. Three-channel stacking allows the model to simultaneously acquire contextual information from the current frame and its preceding and following frames, enhancing sensitivity to transient anomalies. Boundary padding ensures that complete features are output at all time points, avoiding data loss.

[0020] In one optional embodiment, the pre-trained sound reconstruction model is trained as follows: raw audio signals are collected during the normal operation of the indoor equipment of the electric shovel machinery; features are extracted from the raw audio signals to obtain raw audio features; the sound reconstruction model is trained using the raw audio features, wherein the sound reconstruction model includes an encoder and a decoder. The encoder compresses the raw audio features into a low-dimensional latent space representation, and the decoder restores the low-dimensional latent space representation output by the encoder to reconstructed sound features of the same dimension as the raw audio features. Specifically, training the sound reconstruction model using the raw audio features includes: iteratively training the sound reconstruction model using the raw audio features. For each raw audio feature, the following operations are performed: inputting the raw audio feature into the encoder for layer-by-layer processing to obtain a low-dimensional latent space representation; inputting the low-dimensional latent space representation into the decoder for feature recovery to obtain reconstructed sound features of the same dimension as the raw audio features; calculating the loss value between the raw audio features and the reconstructed sound features; and stopping iterative training when the loss value is less than a preset loss value to obtain the pre-trained sound reconstruction model.

[0021] The normal operating state refers to the working state in which all equipment in the electric shovel's mechanical room is within its design operating range, without any faults or performance degradation. The acoustic characteristics of this state exhibit a relatively stable statistical distribution, forming a baseline reference for anomaly detection. The original audio features refer to the log-Mel spectrum three-channel tensor extracted from the normal state audio signal. It is obtained through a series of transformations such as frame windowing, Fourier transform, Mel filtering, logarithmic compression, and three-frame stacking. This feature retains the key time-frequency structure of the soundprint when the equipment is operating normally. The sound reconstruction model refers to a symmetrical encoder-decoder structure based on a deep convolutional neural network. Its encoder part adopts a lightweight backbone network of MobileNetV2, which includes depthwise separable convolutional layers and inverted residual blocks to compress high-dimensional input features into a low-dimensional latent space representation. The decoder part adopts an upsampling network symmetrical to the encoder, which includes transposed convolutional layers or upsampling interpolation layers to restore the latent space representation to the original input dimension. The low-dimensional latent space representation refers to the compact feature vector or feature map output by the encoder. Its dimension is much lower than that of the input features, but it condenses the most essential manifold structure information in the normal soundprint.

[0022] In this step, during a period when there are no known equipment malfunctions in the electric shovel's machinery room, the microphone is activated to continuously collect raw audio signals. The collection duration must cover the typical operating cycle of each piece of equipment in the machinery room. The collected raw audio signals are then sent to an edge computing platform to perform a feature extraction process: amplitude normalization, frame segmentation with Hanning windows, Fourier transform to obtain the power spectrum, Mel filter bank mapping, logarithmic compression, and three-frame stacking. This results in a raw audio feature tensor with a shape equal to the batch size multiplied by the number of channels, the number of Mel frequency bands, and the number of time frames. If a sudden increase in environmental noise is detected during the collection process (e.g., due to the start-up or shutdown of adjacent equipment), the collection window is automatically extended, and disturbed segments are discarded to ensure that the training data contains only clean, normally operating sound. The network structure of the sound reconstruction model is defined on the edge computing platform or an offline training server. The encoder consists of the following layers: an initial ordinary convolutional layer for preliminary extraction of local features from the time-frequency map; multiple inverted residual blocks, each of which first increases the number of channels through pointwise convolution, then performs depthwise convolution to extract spatial features, and finally reduces the number of channels back to the original value through pointwise convolution, with skip connections added when the stride is 1 and the input and output channels are the same; a bottleneck layer as the compressed intermediate feature map; a global average pooling layer to aggregate the spatial dimensions into a one-dimensional vector; and a fully connected layer to further compress the vector to a preset latent vector dimension. The decoder consists of the following layers: an initial fully connected layer to reshape the latent vector into a small-sized feature map; multiple upsampling blocks, each of which first doubles the feature map size through transposed convolution or bilinear interpolation, and then smooths the features through ordinary convolution; a final output convolutional layer to adjust the number of channels to be the same as the input features; and an output activation layer (Sigmoid or ReLU) to constrain the numerical range to a reasonable interval. Through the encoder-decoder architecture design, the model is able to extract low-dimensional manifolds from high-dimensional features and reverse-recover them. At the same time, MobileNetV2's lightweight design ensures that the model can perform real-time inference on edge computing platforms.

[0023] The loss value refers to the mean square error (MSE) between the original audio features and the reconstructed sound features output by the decoder. It is calculated by taking the difference of each element, squaring the result, and then averaging the result. This value quantifies the fidelity of the model's reconstruction of normal samples. The preset loss value is a pre-set training stopping threshold. When the average loss of the model on the validation set is consistently lower than this threshold, it is considered that the model has fully learned the normal voiceprint distribution and iterative training can be stopped.

[0024] During model training, the original audio feature set is first divided into a training set and a validation set. At the start of training, the model parameters are randomly initialized, and the optimizer is set to Adam. In each iteration, a batch-sized original audio feature tensor is randomly sampled from the training set and input into the encoder for forward propagation: the features sequentially pass through an initial convolutional layer, multiple inverted residual blocks, a bottleneck layer, a global average pooling layer, and a fully connected layer, outputting a low-dimensional latent space representation. This latent representation is then input into the decoder for forward propagation: sequentially passing through an initial fully connected layer, multiple upsampling blocks, a final output convolutional layer, and an output activation layer, outputting reconstructed sound features of the same dimension as the input. Subsequently, the mean squared error between the original audio features and the reconstructed sound features is calculated as the loss value for the current batch. Based on the loss value, the gradient of each layer's parameters is calculated using the backpropagation algorithm, and the optimizer updates the parameters according to the gradient. After each training cycle (i.e., traversing the entire training set once), the features from the validation set are input into the model, and the validation loss is calculated. When the validation loss no longer decreases for several consecutive cycles and is stably below a preset loss value, iterative training is stopped, and the current model parameters are saved as a pre-trained sound reconstruction model. If the validation loss oscillates or diverges during training (e.g., due to an excessively high learning rate), the learning rate is automatically reduced to one-tenth of the current value, and training is restarted from the model checkpoint with the lowest most recent validation loss to prevent model degradation.

[0025] In one optional embodiment, after obtaining the pre-trained sound reconstruction model, the method further includes: calculating the reconstruction loss value of each of the original audio features using the pre-trained sound reconstruction model; estimating the gamma distribution of the reconstruction loss value of each of the original audio features using maximum likelihood estimation, and taking the preset percentile of the gamma distribution as the adaptive threshold.

[0026] Among them, the reconstruction loss value refers to the mean squared error (MSE) between the original feature and the reconstructed feature after the pre-trained sound reconstruction model reconstructs each original audio feature. This value quantifies the degree to which normal sound samples deviate from the baseline distribution learned by the model. Ideally, the reconstruction loss values ​​of all normal samples should be concentrated in a small positive interval. Maximum likelihood estimation is a statistical parameter estimation method that fits the probability distribution by finding the parameter value that maximizes the probability of the observed data. This method does not rely on subjective prior settings and can automatically infer the shape parameters of the distribution from the data. The gamma distribution is a continuous probability distribution defined on the positive real number axis. Its probability density function is determined by both shape and scale parameters. This distribution has asymmetric and right-skewed morphological characteristics and can flexibly fit data distributions such as reconstruction loss values ​​that only take positive values ​​and have a long tail on the right. The preset percentile refers to a pre-selected cumulative probability cutoff point. For example, the 90th percentile means that 90% of the normal sample reconstruction loss values ​​are below this percentile, and samples above this percentile are considered statistical outliers.

[0027] Specifically, after the pre-trained sound reconstruction model has completed training and its parameters are fixed, all raw audio features collected under normal operating conditions are input into the model one by one. For each raw audio feature, forward propagation is performed to obtain the corresponding reconstructed sound feature. Then, the element-wise mean square error between the raw feature and the reconstructed feature is calculated, and this value is recorded as the reconstruction loss value of that sample. This process is repeated until all training set samples have obtained corresponding reconstruction loss values, forming a one-dimensional array composed of normal sample reconstruction loss values. Next, a gamma distribution is fitted to this one-dimensional array. Specifically, the maximum likelihood estimation method is used: a likelihood function is constructed with shape and scale parameters as variables, representing the probability of observing all reconstructed loss value samples under the current parameters; the parameter combination that maximizes the likelihood function is solved using a numerical optimization algorithm (e.g., Newton-Raphson iteration or BFGS algorithm), which serves as the shape and scale parameters of the fitted gamma distribution. After obtaining the fitted gamma distribution, the preset percentile of the distribution is calculated. For example, when the preset position is 90%, the value of the inverse function of the cumulative distribution function at the probability value of 0.9 is obtained, and this value is the adaptive threshold. If the threshold needs to be updated due to equipment aging or environmental changes, a new batch of normal operating audio is collected, and the entire process of this step is repeated, replacing the old parameters with the new fitting results. Through the maximum likelihood estimation and gamma distribution fitting in this step, the adaptive threshold can accurately reflect the statistical distribution characteristics of the reconstruction loss value under the current normal equipment condition, without the need for manual experience setting; the positive support and right skewness of the gamma distribution perfectly match the natural shape of the reconstruction loss value, making the fitting results more accurate.

[0028] Step 103: The edge computing platform calculates the reconstruction error of the actual reconstruction feature based on the actual audio feature, identifies whether a device abnormality has occurred based on the reconstruction error and an adaptive threshold, and sends the device abnormality identification result to the remote monitoring terminal through an industrial switch.

[0029] The reconstruction error refers to the difference between the original audio features and the reconstructed features. It is calculated using the L2 norm of element-wise differences, and its magnitude reflects the degree to which the input audio deviates from the normal voiceprint manifold learned by the model. Specifically, the edge computing platform first calculates the sum of squares of the element-wise differences between the reconstructed feature matrix and the original input feature matrix, then takes the square root to obtain the reconstruction error value of the audio segment, which is used as the anomaly score. Subsequently, the platform reads the pre-fitted gamma distribution parameters (shape and scale parameters) from local storage. This distribution is obtained through maximum likelihood estimation based on the anomaly scores of all normal samples during the training phase. The platform calculates a preset percentile (e.g., the 90th percentile) of this gamma distribution as an adaptive threshold for the current device. When the anomaly score is greater than this threshold, an abnormal device event is determined to exist in the current time period; otherwise, it is determined to be normal operation. After the determination is completed, the edge computing platform encapsulates the anomaly identification result into an industrial protocol data frame via an industrial Ethernet interface and sends it to the remote monitoring terminal via an industrial switch.

[0030] Step 104: The remote monitoring terminal displays the device anomaly identification results.

[0031] In this step, the remote monitoring terminal continuously listens for data frames from the industrial switch, and displays the anomaly identification result when it receives the result.

[0032] By applying the technical solution of this embodiment, and through the localized deployment of the microphone and edge computing platform, real-time acquisition and on-site processing of acoustic signals in the electric shovel's mechanical room are achieved, avoiding the bandwidth pressure caused by uploading massive amounts of audio data. Through logarithmic Mel-frequency spectral feature extraction and reconstruction error calculation of the pre-trained sound reconstruction model, a detection baseline can be established with only normal samples, solving the training bottleneck of scarce abnormal samples in industrial scenarios. Through an adaptive threshold mechanism based on gamma distribution, automatic adaptation of the anomaly judgment boundary is achieved without human intervention, providing a technical solution for monitoring the status of electric shovel equipment in the mechanical room that requires no abnormal samples, is highly adaptive, and can be deployed in real time.

[0033] In one optional embodiment, reconstructing the actual audio features using a pre-trained sound reconstruction model includes: reconstructing the actual audio features using a pre-trained sound reconstruction model that matches the current operating condition corresponding to the actual audio signal, wherein the pre-trained sound reconstruction model that matches the current operating condition is trained based on the original audio signal during normal operation under the current operating condition; correspondingly, identifying whether a device malfunction has occurred based on the reconstruction error and an adaptive threshold includes: identifying whether a device malfunction has occurred based on the reconstruction error and an adaptive threshold corresponding to the current operating condition, wherein the adaptive threshold corresponding to the current operating condition is determined by estimating the reconstruction loss value of the original audio features under the current operating condition using a gamma distribution and taking a preset percentile of the gamma distribution.

[0034] The current operating condition refers to the set of specific working states of the electric shovel's mechanical chamber during operation. This set includes, but is not limited to, the start-stop status of equipment combinations (e.g., the pushing mechanism is working while the air compressor is on standby), load levels (light load, heavy load), and environmental conditions (temperature, humidity). The acoustic characteristic distribution of the same equipment varies under different operating conditions. The pre-trained sound reconstruction model matched with the current operating condition refers to a reconstruction model trained separately for a specific operating condition. Its training data comes only from the original audio signal of the electric shovel's mechanical chamber in normal operation under that operating condition. This model specifically learns the manifold structure of normal sound signatures under that operating condition. Different operating conditions correspond to different model parameters. The adaptive threshold corresponding to the current operating condition refers to the judgment boundary obtained by fitting the reconstruction loss value of normal samples under the same operating condition through a gamma distribution. This threshold corresponds one-to-one with the acoustic characteristics under that operating condition. Different operating conditions have different threshold values.

[0035] In this step, a working condition identification module is first maintained within the edge computing platform. This module analyzes the operating information of each device in the electric shovel's machine compartment to determine the current working condition category. The working condition category is predefined into several typical working states, such as "push mechanism operating alone," "lifting mechanism operating alone," "multi-mechanism joint operation," and "idle standby." The working condition identification module outputs a working condition label. After obtaining the working condition label, the edge computing platform loads a pre-trained sound reconstruction model matching the label from local storage. This model has been trained offline using normal operating audio signals collected under this working condition, and its encoder and decoder parameters are specifically adapted to the acoustic feature distribution of this working condition. The platform inputs the actual audio features extracted in step 102 into the model, performs forward propagation, and obtains the actual reconstructed features.

[0036] Subsequently, the platform retrieves the adaptive threshold corresponding to the operating condition label from local storage. This threshold is generated as follows: During the offline phase, normal audio features under the same operating condition are input into the pre-trained model for that condition. The reconstruction loss value of all normal samples is calculated, and then a gamma distribution is fitted using maximum likelihood estimation. A preset percentile (e.g., the 90th percentile) is taken as the specific threshold for that operating condition. The platform calculates the reconstruction error between the actual audio features and the reconstructed features, and compares this error with the threshold for that operating condition. If the reconstruction error is greater than the threshold, it is determined that an equipment malfunction has occurred under the current operating condition; otherwise, it is determined to be operating normally.

[0037] In addition, when the working condition recognition module cannot clearly determine the current working condition (e.g., the feature is in the boundary region of two predefined working conditions), a conservative strategy is automatically adopted: the models and thresholds of two adjacent working conditions are loaded at the same time, the reconstruction error is calculated separately, and if either working condition is determined to be abnormal, an alarm is triggered; at the same time, the audio segment is marked as a "boundary working condition sample" and stored in the cache for subsequent working condition definition expansion.

[0038] In this embodiment, by establishing a working condition recognition module and training a separate sound reconstruction model and calculating a dedicated adaptive threshold for each working condition, the acoustic anomaly detection system for the electric shovel's mechanical chamber can adapt to changes in acoustic characteristics under different working conditions of the equipment, thus solving the problem of increased false alarm rate of a single model in multiple working condition scenarios.

[0039] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, another method for detecting anomalies in the mechanical chamber of an electric shovel based on audio signal analysis is provided, such as... Figure 2 As shown, the method includes: Step 201: The edge computing platform places the actual audio features identified as device malfunctions and manually confirmed into an abnormal sample buffer queue, and places the actual audio features identified as device normal and manually confirmed into a normal sample buffer queue.

[0040] In this step, two fixed-length circular buffers are first allocated in the memory of the edge computing platform, serving as an abnormal sample buffer queue and a normal sample buffer queue, respectively. When an equipment malfunction is detected and the malfunction identification result is sent to the remote monitoring terminal, the maintenance personnel play back the audio segment corresponding to the time period through the terminal's interactive interface and comprehensively observe the parameters such as current and temperature displayed by the electric shovel's main control system. If the maintenance personnel confirm that the malfunction corresponds to a real equipment failure (e.g., hearing a distinct periodic impact sound), they click the "Confirm Anomaly" button, triggering the terminal to send a confirmation command to the edge computing platform. Upon receiving the confirmation command, the edge computing platform pushes the actual audio features corresponding to the audio segment and its reconstruction loss value to the tail of the abnormal sample buffer queue. If the maintenance personnel confirm that the malfunction is a false alarm (e.g., a misjudgment due to external knocking sounds), they click the "Mark False Alarm" button and do not add the sample to any queue. For audio segments determined to be normal, the maintenance personnel can randomly sample and confirm them; when confirmed to be normal, the sample is added to the tail of the normal sample buffer queue. When any queue reaches its capacity limit, the oldest sample is automatically removed from the head of the queue, ensuring that the queue always stores the most recently verified samples. This double-buffered queue design continuously accumulates high-quality, manually validated samples, providing reliable training data for subsequent model fine-tuning.

[0041] Step 202: Perform multiple rounds of sampling on the abnormal sample cache queue and the normal sample cache queue respectively, and use the abnormal samples and normal samples obtained in each round of sampling to iteratively train the pre-trained sound reconstruction model. The training process for each round is as follows: The pre-trained sound reconstruction model is used to reconstruct each abnormal sample and each normal sample to obtain abnormal reconstruction features and normal reconstruction features. The reconstruction loss of each abnormal sample and each normal sample is calculated, and the total loss is calculated based on the reconstruction losses of each abnormal sample and each normal sample. The parameters of the pre-trained sound reconstruction model are then fine-tuned based on the total loss. The expression for the total loss L is: , The number of normal samples. Let i be the loss value for the i-th normal sample. The number of abnormal samples. Let j be the loss value of the j-th outlier sample. The loss value for the i-th normal sample is the reconstruction loss of the i-th normal sample, and the loss value for the j-th anomalous sample is the maximum value between the difference between the positive boundary and the reconstruction loss of the j-th anomalous sample and 0.

[0042] Among them, multi-round sampling refers to randomly selecting a certain number of samples from the abnormal sample buffer queue and the normal sample buffer queue to form a training batch. This process is repeated multiple times until the model converges. The total loss is the weighted sum of the average normal sample reconstruction loss and the average abnormal sample hinge loss. The abnormal sample hinge loss is defined as the boundary value (positive boundary) minus the maximum value between the abnormal sample reconstruction loss and zero. The balance coefficient is a pre-set positive real number hyperparameter used to adjust the weight of the abnormal sample loss relative to the normal sample loss in the total loss.

[0043] In this step, the fine-tuning trigger condition is first set: when the number of samples in the abnormal sample buffer queue reaches a preset threshold (e.g., 50) and the current edge computing platform is in an idle period (e.g., during nighttime electric shovel maintenance), the fine-tuning process is automatically initiated. Before fine-tuning begins, normal samples proportional to the number of abnormal samples are randomly selected from the normal sample buffer queue (e.g., the number of samples is four times the number of abnormal samples) to form the training dataset for the current batch. Subsequently, the following operations are performed in each round of training: Each normal and abnormal sample in the current batch is sequentially input into the pre-trained sound reconstruction model, and forward propagation is performed to obtain their respective reconstruction features. For normal samples, their reconstruction loss value (mean squared error) is calculated; for abnormal samples, after calculating their reconstruction loss value, the hinge loss is further calculated. The hinge loss is the maximum value between the difference between the boundary value (i.e., the positive boundary) and the reconstruction loss value of the abnormal sample and zero. The boundary value is pre-set based on the 95th percentile of the reconstruction loss value of normal samples in the initial training set multiplied by a safety factor (e.g., 1.2) and stored in the edge computing platform. Then, the total loss is calculated: the average reconstruction loss of normal samples plus the balance coefficient multiplied by the average hinge loss of abnormal samples. The initial value of the balance coefficient is set to 1.0 and is dynamically adjusted according to the gradient magnitude of the two types of losses during fine-tuning: when the gradient norm of the abnormal sample loss is significantly smaller than the gradient norm of the normal sample loss, the balance coefficient is automatically increased; otherwise, the balance coefficient is decreased to keep the contributions of the two types of gradients balanced.

[0044] After obtaining the total loss, the gradients of the parameters of each layer of the model are calculated through backpropagation, and the parameters are updated using the Adam optimizer with a preset learning rate (e.g., one-tenth of the initial training learning rate), completing one round of parameter fine-tuning. Multiple rounds of sampling and parameter updates are repeated until a stopping condition is met. For example, in five consecutive training rounds, the average reconstruction loss of normal samples on the validation set increases by no more than 1%, and the average reconstruction loss of abnormal samples remains consistently higher than the boundary value. If, during fine-tuning, the average reconstruction loss of normal samples increases sharply (e.g., exceeding 10% of the initial value), fine-tuning is immediately stopped, the model parameters are restored to their state before fine-tuning began, the balancing coefficient is halved, and fine-tuning is restarted to prevent the model from forgetting the normal baseline. After fine-tuning, the updated model parameters are saved to the edge computing platform, replacing the original pre-trained model, and the used samples in the abnormal sample buffer queue are cleared, while the samples in the normal sample buffer queue are retained for subsequent fine-tuning. Through the multi-round sampling and hinged loss design in this embodiment, the model can use manually confirmed abnormal samples for incremental learning, and the learning direction is to further increase the reconstruction error of abnormal samples rather than attempt to reconstruct abnormalities, thereby enhancing the model's sensitivity to the types of abnormalities that have appeared.

[0045] Step 203: When the number of normal samples in the normal sample buffer queue exceeds a preset threshold, the reconstruction loss value of each normal sample is calculated using the current sound reconstruction model, and the reconstruction loss value of each normal sample is estimated by gamma distribution using maximum likelihood estimation. The preset percentile of the gamma distribution is then used to update the adaptive threshold.

[0046] In this step, a normal sample quantity monitoring module is first set up in the edge computing platform to continuously count the number of samples in the normal sample buffer queue. When the number of normal samples in the queue exceeds a preset threshold (e.g., 100), the threshold update process is automatically triggered. After triggering, the platform reads the reconstruction loss values ​​of all normal samples from the normal sample buffer queue, forming a one-dimensional array. Gamma distribution fitting is performed on this array: the maximum likelihood estimation method is used to construct a likelihood function with shape and scale parameters as variables, and a numerical optimization algorithm is used to solve for the parameter combination that maximizes the likelihood function. After obtaining the fitted gamma distribution, the preset percentile of the distribution is calculated. This preset percentile is consistent with the percentile set in the initial threshold (e.g., both use the 90th percentile) to ensure that the judgment scale is consistent before and after the threshold update. The calculated percentile value is the updated adaptive threshold, replacing the original adaptive threshold, for subsequent anomaly judgment calls. This allows the adaptive threshold to be automatically updated as the normal state of the device slowly evolves, solving the problem of increased false alarm rate caused by equipment wear and environmental changes during long-term operation of static thresholds.

[0047] The anomaly detection system used in the audio signal analysis-based anomaly detection method for electric shovel machine compartments in this application consists of the following core hardware components:

[0048] The microphone array is optimized according to the spatial distribution of key equipment in the machine room, with each microphone covering a radius of approximately 3-5 meters. The microphones connect to the NVIDIA Jetson Orin edge computing platform via digital audio interfaces, and the computing platform is interconnected with the remote monitoring terminal via industrial Ethernet. The overall hardware architecture follows an "Edge-Cloud Collaborative Computing Paradigm," where acoustic data undergoes real-time inference at the edge, with only detection results and key metadata uploaded to the remote terminal, effectively reducing network bandwidth requirements.

[0049] The software architecture of this system is built on the ROS (Robot Operating System) framework, adopting a modular, loosely coupled node communication architecture. The software stack is divided into the following layers from bottom to top: Operating system layer: Ubuntu Linux operating system, equipped with NVIDIA JetPack SDK, providing GPU-accelerated runtime libraries such as CUDA, cuDNN, and TensorRT.

[0050] Middleware layer: ROS communication framework, providing inter-process communication mechanisms such as Topic publish / subscribe and Service request / response.

[0051] Algorithm Engine Layer: A high-performance inference engine implemented in C++, integrating the TensorRT inference backend, and containing two core components: the audio preprocessing module (AudioBufferManager) and the acoustic classification module (AudioClassifier).

[0052] Application Interface Layer: Provides the ROS Topic interface to publish anomaly detection results for upper-layer monitoring systems to subscribe to and consume.

[0053] The core software modules include: AudioBufferManager: This module is responsible for subscribing to real-time audio stream data from the ROS audio data Topic ( / audio_main / data), implementing audio frame buffering, splicing, and segmentation management. Key parameters include: buffer size (buffer_size) set to 88200 sample points (corresponding to 2 seconds of audio at a 44.1kHz sampling rate), and the number of data packets read per session (packets_to_read) set to 86. The module employs a circular buffer design to ensure the continuity and real-time performance of the audio data.

[0054] The AudioClassifier module encapsulates a complete acoustic anomaly detection inference pipeline, including: audio signal preprocessing (resampling, normalization), log-Mel spectrum feature extraction (based on a self-developed C++ implementation of the librosa equivalent library), TensorRT inference engine call (loading pre-trained .engine model files), anomaly score calculation, and threshold determination. The module utilizes the CUDA Runtime API for GPU-accelerated computation, and employs FP32 floating-point format for inference precision to ensure detection accuracy.

[0055] Model Conversion and Deployment Pipeline: The deployment of deep neural network models follows a standardized pipeline: First, the model is trained on a GPU server using the PyTorch framework, and the trained model parameters are saved as a .pth format weight file. Then, a custom conversion script converts the .pth file to a .wts intermediate format. Finally, on the NVIDIA JetsonOrin platform, TensorRT Builder is used to compile and optimize the .wts file into an .engine inference engine file (model_excavator_44.1k.engine). During the TensorRT compilation process, optimization strategies such as Layer Fusion and Kernel Auto-Tuning are implemented, significantly improving inference throughput.

[0056] In a specific application scenario, such as Figure 3 This paper demonstrates the complete operation flow of the electric shovel mechanical compartment anomaly detection method based on audio signal analysis described in this application, covering both training and inference modes. The operation flow is divided into the following key stages: (1) Acoustic signal acquisition stage: The industrial-grade microphone array deployed in the machine room continuously acquires multi-channel audio signals, and the audio data is transmitted in real time to the AudioBufferManager module of the edge computing platform through ROS Topic.

[0057] (2) Acoustic feature extraction stage: AudioBufferManager segments the buffered audio data and feeds it into the feature extraction pipeline, which then performs operations such as resampling (44.1kHz), amplitude normalization, short-time Fourier transform (STFT), Mel filter bank mapping, and logarithmic compression in sequence to generate a standardized three-channel log-Mel spectrum tensor.

[0058] (3) Model Training Phase (Offline): In training mode, the system uses a large amount of audio data collected during normal operation to train the deep encoder-decoder network. The network learns a low-dimensional manifold representation of normal voiceprints by minimizing the MSE reconstruction loss of normal audio. After training, a gamma distribution is fitted based on the abnormal score distribution of normal samples, and the 90th percentile is calculated as the abnormality judgment threshold.

[0059] (4) Real-time inference stage (online): In inference mode, the TensorRT inference engine performs forward inference with FP32 precision, calculates the reconstructed features of the audio to be detected, and calculates the anomaly score using the L2 norm. The anomaly score is compared with a preset threshold. If the threshold is exceeded, an anomaly alarm signal is issued through ROS Topic; otherwise, a normal operation signal is issued.

[0060] (5) Remote monitoring and alarm stage: The detection results are transmitted to the remote monitoring terminal in real time via industrial Ethernet. The device operating status, abnormal event timeline, abnormal score trend chart and other information are displayed in a visual interface to support remote decision-making by operation and maintenance personnel.

[0061] Based on comprehensive verification using actual field data from electric shovels and laboratory simulation data, the method described in this application achieves the following performance indicators:

[0062] This system has been verified to effectively detect the following typical abnormalities in the mechanical chamber of an electric shovel: Electric shovel air compressor startup failure: Abnormal acoustic characteristics during the air compressor startup process (such as the starter motor stalling sound, abnormal valve leakage sound) are significantly different from the normal startup sound, and the system can accurately identify them.

[0063] Abnormal sound from the rotary motor: Abnormal frequency components generated by faults such as worn bearings and abnormal gear meshing in the rotary motor deviate from the normal operating sound waveform baseline, triggering an abnormal alarm.

[0064] Abnormal sounds from the cable winding mechanism: Abnormal acoustic events such as impact sounds and friction sounds caused by mechanical failures such as loose chains and worn pulleys in the cable winding mechanism can be effectively captured by the system.

[0065] Furthermore, based on the open set detection capability of the unsupervised detection paradigm, the system also has the potential to detect unknown anomaly types that did not appear during the training phase.

[0066] By applying the technical solutions of the embodiments of this application, at least the following beneficial effects can be achieved: A breakthrough in unsupervised anomaly detection paradigm: Model training can be completed using only normal audio data, eliminating the need to collect and label scarce anomalous audio samples, fundamentally solving the problem of scarce anomalous samples in industrial scenarios. Based on the idea of ​​"normal voiceprint baseline modeling," the system identifies audio signals that deviate from normal voiceprint patterns as anomalous, and possesses open-set detection capability for unknown anomaly types.

[0067] Non-contact multi-device parallel monitoring: By deploying an industrial-grade microphone array in the machine room, the operating sounds of multiple devices can be collected simultaneously in a non-contact manner. A single hardware system can cover the acoustic monitoring needs of the entire machine room, significantly reducing sensor deployment costs and system complexity.

[0068] Edge computing real-time inference architecture: It adopts the NVIDIA Jetson Orin embedded AI computing platform as the edge inference node, combined with the TensorRT high-performance inference engine to accelerate the model. The inference latency of a single 10-second audio segment is less than 100 milliseconds, which meets the real-time requirements of industrial scenarios.

[0069] Empowering unmanned remote control operations in mines: It can be deeply integrated with the remote control operating system of electric shovels, providing remote operators with real-time acoustic perception capabilities of the equipment status in the machine room. This compensates for the lack of perception that operators cannot be physically present to listen to and identify obstacles in the remote control operation mode, and is an important component of the intelligent and unmanned operation system in mines.

[0070] This application also provides a computer device, which may specifically be a personal computer, a server, a network device, etc. Figure 4 As shown, the computer device includes a bus, a processor, memory, and a communication interface, and may also include an input / output interface and a display device. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores location information. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0071] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0072] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, having stored thereon a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0073] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0074] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0075] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0076] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting anomalies in the mechanical compartment of an electric shovel based on audio signal analysis, characterized in that, The method includes: The actual audio signal of the electric shovel during actual operation is collected by a microphone deployed inside the shovel's machine room; the actual audio signal is then sent to the edge computing platform via an audio acquisition interface. The edge computing platform extracts features from the actual audio signal to obtain actual audio features, and reconstructs the actual audio features using a pre-trained sound reconstruction model to obtain actual reconstructed features. The pre-trained sound reconstruction model is trained based on the original audio signal during normal operation of the electric shovel machine. The edge computing platform calculates the reconstruction error of the actual reconstruction features based on the actual audio features, identifies whether a device malfunction has occurred based on the reconstruction error and an adaptive threshold, and sends the device malfunction identification result to the remote monitoring terminal through an industrial switch. The remote monitoring terminal displays the results of the device anomaly identification.

2. The method according to claim 1, characterized in that, The pre-trained sound reconstruction model is trained in the following way: The original audio signal of the electric shovel's indoor equipment during normal operation is collected, and the original audio signal is used to extract features to obtain the original audio features; A sound reconstruction model is trained using the original audio features. The sound reconstruction model includes an encoder and a decoder. The encoder is used to compress the original audio features into a low-dimensional latent space representation, and the decoder is used to restore the low-dimensional latent space representation output by the encoder into reconstructed sound features of the same dimension as the original audio features.

3. The method according to claim 2, characterized in that, Feature extraction is performed on the original audio signal to obtain original audio features, including: The original audio signal is normalized, the normalized original audio signal is divided into frames, a Hanning window is applied to each frame, and a Fourier transform is performed on each frame to convert the time-domain waveform into the frequency domain. The power spectrum is obtained by taking the square of the amplitude of the spectrum, and the power spectra of all frames are superimposed in time order to obtain the time-frequency diagram. Mel spectrum extraction is performed on the time-frequency graph using a Mel filter bank to obtain a Mel spectrum matrix. The logarithm of each element in the Mel spectrum matrix is ​​then taken to obtain a logarithmic Mel spectrum matrix. Based on the log-Mel spectrum matrix, for each time frame index, the log-Mel spectrum vectors of itself and the two frames before and after it are stacked into three channels to obtain the original audio features.

4. The method according to claim 2, characterized in that, The encoder sequentially comprises an input layer, an initial convolutional layer, multiple inverted residual blocks, a bottleneck layer, a global average pooling layer, and a fully connected layer; the decoder sequentially comprises an initial fully connected layer, multiple upsampling blocks, a final output convolutional layer, and an output activation layer. Training a sound reconstruction model using the original audio features includes: The sound reconstruction model is iteratively trained using the original audio features, with each original audio feature performing the following operations: The original audio features are input into the encoder for layer-by-layer processing to obtain a low-dimensional latent space representation. The low-dimensional latent space representation is then input into the decoder for feature recovery to obtain reconstructed sound features with the same dimension as the original audio features. Calculate the loss value between the original audio features and the reconstructed sound features, and stop iterative training when the loss value is less than a preset loss value to obtain the pre-trained sound reconstruction model.

5. The method according to claim 4, characterized in that, After obtaining the pre-trained sound reconstruction model, the method further includes: The reconstruction loss value for each of the original audio features is calculated using the pre-trained sound reconstruction model; The reconstruction loss value of each of the original audio features is estimated by gamma distribution using maximum likelihood estimation, and the preset percentile of the gamma distribution is taken as the adaptive threshold.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: The edge computing platform places the actual audio features identified as device malfunctions and manually confirmed into an abnormal sample buffer queue, and places the actual audio features identified as device normal and manually confirmed into a normal sample buffer queue. Multiple rounds of sampling are performed on the abnormal sample cache queue and the normal sample cache queue, respectively. The abnormal samples and normal samples obtained in each round of sampling are used to iteratively train the pre-trained sound reconstruction model. The training process in each round is as follows: The pre-trained sound reconstruction model is used to reconstruct each anomalous sample and each normal sample to obtain anomalous reconstruction features and normal reconstruction features. The reconstruction loss for each anomalous sample and each normal sample is calculated, and the total loss is calculated based on these losses. The pre-trained sound reconstruction model is then updated and its parameters are fine-tuned based on the total loss. The expression for the total loss L is as follows: , The number of normal samples. Let i be the loss value for the i-th normal sample. The number of abnormal samples. Let j be the loss value of the j-th outlier sample. The loss value for the i-th normal sample is the reconstruction loss of the i-th normal sample, and the loss value for the j-th anomalous sample is the maximum value between the difference between the positive boundary and the reconstruction loss of the j-th anomalous sample and 0.

7. The method according to claim 6, characterized in that, The method further includes: When the number of normal samples in the normal sample buffer queue exceeds a preset threshold, the reconstruction loss value of each normal sample is calculated using the current sound reconstruction model, and the reconstruction loss value of each normal sample is estimated by gamma distribution using maximum likelihood estimation. The preset percentile of the gamma distribution is then used to update the adaptive threshold.

8. The method according to claim 6, characterized in that, The actual audio features are reconstructed using a pre-trained sound reconstruction model, including: The actual audio features are reconstructed using a pre-trained sound reconstruction model that matches the current operating condition corresponding to the actual audio signal. The pre-trained sound reconstruction model that matches the current operating condition is trained based on the original audio signal under the normal operating condition. Based on the reconstruction error and adaptive threshold, it is determined whether a device malfunction has occurred, including: Whether a device malfunction has occurred is identified based on the reconstruction error and the adaptive threshold corresponding to the current operating condition. The adaptive threshold corresponding to the current operating condition is determined by estimating the reconstruction loss value of the original audio features under the current operating condition using a gamma distribution and taking a preset percentile of the gamma distribution.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.

Images