An abnormal sound detection method, device, equipment and medium

Abstract

The application relates to the technical field of refrigerator abnormal sound detection, and discloses an abnormal sound detection method, device, equipment and medium, wherein the method comprises the following steps: constructing a space-time sound field data set; filtering and processing acoustic signals in the space-time sound field data set by using an environmental reference signal to obtain pure sound field signals, and when it is monitored that the pure sound field signals meet preset triggering conditions, corresponding target abnormal sound segments are intercepted; the spatial distribution relationship of the target abnormal sound segments at each target step position is determined to reconstruct the spatial sound field energy distribution of a detection space, and the energy peak position of the spatial sound field energy distribution is determined as an abnormal sound source head coordinate generating the target abnormal sound segment; feature extraction and classification processing are performed on the target abnormal sound segment to determine an abnormal sound source component corresponding to the target abnormal sound segment, and a fault diagnosis result is generated in combination with the abnormal sound source head coordinate. The technical scheme provided by the application can realize the technical effects of accurate abnormal sound positioning and component identification.

Description

Technical Field

[0001] This application relates to the field of refrigerator noise detection technology, and in particular to a noise detection method, device, equipment and medium. Background Technology

[0002] In daily refrigerator use, unusual noises are often the earliest sign of potential malfunctions such as wear and tear of mechanical parts, abnormal compressor start-up and shutdown, refrigerant pipe vibration, or loose structural components. Traditional detection methods rely solely on manual listening or single-channel acoustic sensors, which are easily affected by kitchen background noise. The human ear has limited ability to distinguish low-amplitude, short-duration, or high-frequency transient unusual noises and cannot provide spatial distribution information, leading to blind and time-consuming repairs. A single microphone can only collect sound pressure at a single point, making it difficult to use spatial correlation to suppress incoherent noise or perform beamforming. Environmental power frequency interference and internal reverberation can significantly increase the noise level, masking the energy of transient unusual noises. The detection threshold is difficult to balance sensitivity and robustness, and it lacks directional information, thus failing to achieve three-dimensional localization of the source of the unusual noise. It can only determine whether there is an unusual noise but cannot locate the faulty part.

[0003] Therefore, how to overcome environmental noise interference to improve the detection signal-to-noise ratio, and achieve accurate location of abnormal sounds and component identification by combining spatial coordinates, thereby improving the accuracy of abnormal sound capture and the efficiency of fault diagnosis, is a technical problem that urgently needs to be solved in the field of refrigerator abnormal sound detection. Summary of the Invention

[0004] This application provides a method, apparatus, device, and medium for detecting abnormal noise, which achieves the technical effect of overcoming environmental noise interference to improve the detection signal-to-noise ratio, and realizing accurate positioning of abnormal noise and component identification by combining spatial coordinates, thereby improving the accuracy of abnormal noise capture and the efficiency of fault diagnosis.

[0005] To achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, embodiments of this application provide an abnormal sound detection method, applied to a detection system including a mobile acquisition module and a reference acquisition module, the method comprising: The mobile acquisition module is controlled to move within the detection space. For each target step position, acoustic signals and corresponding real-time spatial coordinates are acquired to construct a spatiotemporal sound field dataset. The environmental reference signal collected by the reference acquisition module is obtained, and the acoustic signal in the spatiotemporal sound field dataset is filtered using the environmental reference signal to obtain a pure sound field signal. When the pure sound field signal is detected to meet the preset triggering conditions, the corresponding target abnormal sound segment is extracted. Based on the real-time spatial coordinates corresponding to each target step position in the spatiotemporal sound field dataset, the spatial distribution relationship of the target abnormal sound segment under each target step position is determined to reconstruct the spatial sound field energy distribution of the detection space, and the energy peak position of the spatial sound field energy distribution is determined as the abnormal sound source coordinates that generate the target abnormal sound segment; The target abnormal sound segment is subjected to feature extraction and classification processing to determine the abnormal sound source component corresponding to the target abnormal sound segment, and a fault diagnosis result is generated by combining the coordinates of the abnormal sound source.

[0006] In one embodiment, the step of filtering the acoustic signals in the spatiotemporal sound field dataset using the environmental reference signal to obtain a clean sound field signal includes: Based on the preset bandpass filter coefficients, a recursive differential operation is performed on the acoustic signal to output a preliminary filtered signal; The residual signal between the environmental reference signal and the preliminary filtered signal is determined, and based on the residual signal and the environmental reference signal, the cancellation weight of the environmental reference signal in the preliminary filtered signal is dynamically adjusted to iteratively approximate and remove the external crosstalk components related to the environmental reference signal, and output the preliminary denoised signal. The initial denoised signal is subjected to frequency domain decomposition to obtain the frequency domain amplitude spectrum of the current frame; The global mean square error between the current frame frequency domain amplitude spectrum and the preset background noise model is determined. When the global mean square error is greater than a preset threshold, the current frame frequency domain amplitude spectrum is determined to be in an abnormal sound state. Based on the frequency domain difference between the current frame frequency domain amplitude spectrum and the preset background noise model, the clean sound field signal is reconstructed.

[0007] In one implementation, the preset background noise model is updated in the following way: When the global mean square error is lower than a preset threshold, it is determined that the current frame frequency domain amplitude spectrum is in a noise-free state, and the current frame frequency domain amplitude spectrum is stored as a pure background noise sample in a circular buffer. When the number of pure background noise samples accumulated in the circular buffer reaches a preset number of frames, the spectral amplitude of the preset number of pure background noise samples is averaged to update the preset background noise model.

[0008] In one embodiment, the step of intercepting the corresponding target abnormal sound segment when the pure sound field signal is detected to meet a preset trigger condition includes: The pure sound field signal is processed by frame segmentation to determine the short-time energy value of each frame of the pure sound field signal; The short-term energy value is compared with a preset dynamic trigger threshold, and a trigger command is generated based on the comparison result. In response to the trigger command, the pure sound field signal is truncated based on a preset time window length to obtain the target audio segment.

[0009] In one embodiment, the mobile acquisition module includes multiple acquisition units, and the reconstructing of the spatial sound field energy distribution of the detection space includes: Extract the pure sound field signal collected by each acquisition unit at the same time from the target abnormal sound segment; Determine the cross-power spectrum between the corresponding pure sound field signals of any two acquisition units, and normalize the amplitude of the cross-power spectrum based on the phase transformation weighting function to obtain the generalized cross-correlation function; A peak search is performed on the generalized cross-correlation function, and the time delay variable corresponding to the maximum peak value is determined as the measured arrival time difference between the acquisition unit pairs. A three-dimensional spatial grid covering the detection space is constructed, and each location of the sound source to be estimated in the three-dimensional spatial grid is traversed. The assumed distance difference between the location of the sound source to be estimated and each pair of acquisition units is determined, and the matching degree between the assumed distance difference and the measured physical distance difference corresponding to the measured arrival time difference is determined. Based on the matching degree, the likelihood energy value of the location of the sound source to be estimated is determined, and the likelihood energy values ​​of all locations of the sound source to be estimated in the detection space are aggregated to construct the spatial sound field energy distribution.

[0010] In one embodiment, determining the energy peak position of the spatial sound field energy distribution as the coordinates of the source of the abnormal sound segment includes: The location of the sound source with the largest likelihood energy value in the spatial sound field energy distribution is determined as the energy peak location, and the coordinates corresponding to the energy peak location are determined as the coordinates of the abnormal sound source.

[0011] In one embodiment, the step of performing feature extraction and classification on the target abnormal sound segment to determine the abnormal sound source component corresponding to the target abnormal sound segment, and generating a fault diagnosis result by combining the coordinates of the abnormal sound source, includes: The target abnormal sound segment is subjected to Mel-frequency transform processing to generate a two-dimensional time-frequency feature map; The two-dimensional time-frequency feature map is input into a pre-trained lightweight convolutional neural network classification model. The lightweight convolutional neural network classification model is used to perform feature analysis on the two-dimensional time-frequency feature map and output the classification confidence of the target abnormal sound segment belonging to the preset component category. The component category with the highest classification confidence is identified as the component that causes the abnormal noise. Combined with the coordinates of the source of the abnormal noise, a fault diagnosis result containing the component that causes the abnormal noise and the coordinates of the source of the abnormal noise is generated.

[0012] Secondly, embodiments of this application provide an abnormal sound detection device, applied to a detection system including a mobile acquisition module and a reference acquisition module, the device comprising: The signal acquisition unit is used to control the movement of the mobile acquisition module within the detection space, and to acquire acoustic signals and corresponding real-time spatial coordinates for each target step position to construct a spatiotemporal sound field dataset. The signal processing unit is used to acquire the environmental reference signal acquired by the reference acquisition module, use the environmental reference signal to filter the acoustic signal in the spatiotemporal sound field dataset to obtain a clean sound field signal, and when the clean sound field signal is detected to meet the preset triggering conditions, the corresponding target abnormal sound segment is extracted. The sound source localization unit is used to determine the spatial distribution relationship of the target abnormal sound segment at each target step position based on the real-time spatial coordinates corresponding to each target step position in the spatiotemporal sound field dataset, so as to reconstruct the spatial sound field energy distribution of the detection space, and determine the energy peak position of the spatial sound field energy distribution as the coordinates of the abnormal sound source that generates the target abnormal sound segment; The fault diagnosis unit is used to extract and classify the target abnormal sound segment, determine the abnormal sound source component corresponding to the target abnormal sound segment, and generate a fault diagnosis result by combining the coordinates of the abnormal sound source.

[0013] Thirdly, embodiments of this application provide a computer device, including: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes these computer instructions to perform the aforementioned abnormal sound detection method.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions, which are used to cause a computer to execute the above-described abnormal sound detection method.

[0015] The technical solution provided by one or more embodiments of this application controls a mobile acquisition module to step through the detection space, synchronously acquiring acoustic signals and corresponding real-time spatial coordinates at each target step position to construct a spatiotemporal sound field dataset. This not only breaks through the spatial coverage limitations of fixed microphone arrays and avoids missing abnormal noises in different areas inside the refrigerator, but also provides a multi-dimensional spatial sampling benchmark for subsequent sound field reconstruction. Subsequently, the environmental reference signal acquired by the reference acquisition module is obtained, and the acoustic signals in the spatiotemporal sound field dataset are filtered to cancel out steady-state background noise interference such as the refrigerator compressor's operating noise and airflow noise in the duct, resulting in a clean sound field signal with a high signal-to-noise ratio. The corresponding signal is intercepted only when the clean sound field signal meets the preset trigger condition. The system can filter out a large amount of invalid data under normal operating conditions by targeting abnormal noise segments, reducing the computational load of subsequent feature extraction and classification, and reducing the probability of misjudgment. Then, based on the real-time spatial coordinates corresponding to each target step position in the spatiotemporal sound field dataset, the spatial distribution relationship of the target abnormal noise segment at each step position is determined to reconstruct the spatial sound field energy distribution of the detection space. The energy peak position is determined as the coordinate of the abnormal noise source, which can effectively eliminate reverberation interference inside the refrigerator and greatly improve the accuracy of abnormal noise localization. Finally, feature extraction and classification processing of the target abnormal noise segment are performed to determine the abnormal noise source component. The component classification result is fused with the abnormal noise source coordinate to generate the fault diagnosis result, which can simultaneously output the fault component type and precise spatial location, greatly improving the efficiency of fault diagnosis. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A flowchart of an abnormal sound detection method provided in an embodiment of this application; Figure 2 Three-view drawings of a refrigerator for detecting space provided in an embodiment of this application; Figure 3 This is a schematic diagram of the guide rail provided in an embodiment of this application; Figure 4 A flowchart for obtaining a pure sound field signal provided in this application embodiment; Figure 5 A comparison diagram of time-domain waveforms and spectrum diagrams provided in the embodiments of this application; Figure 6 A flowchart illustrating the update method of the preset background noise model provided in the embodiments of this application; Figure 7This is a flowchart illustrating the process of extracting the corresponding target abnormal sound segment as provided in an embodiment of this application; Figure 8 Time-frequency analysis diagrams from a microscopic perspective provided in the embodiments of this application; Figure 9 A flowchart illustrating the spatial sound field energy distribution in the reconstruction detection space provided in this application embodiment; Figure 10 A three-dimensional spatial visualization diagram provided for an embodiment of this application; Figure 11 A flowchart for determining the energy peak position of the spatial sound field energy distribution as the coordinates of the source of the target abnormal sound segment, provided for embodiments of this application; Figure 12 A flowchart of step S7 provided in an embodiment of this application; Figure 13 A block diagram of an abnormal noise detection device provided in an embodiment of this application; Figure 14 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] During the daily use of a refrigerator, unusual noises are often the earliest sign of potential malfunctions such as wear and tear on mechanical parts, abnormal compressor start-up and shutdown, refrigerant pipe vibration, or loose structural components. Traditional methods for detecting unusual noises are basically limited to manual listening or single-channel acoustic sensors. This approach has revealed a series of systemic shortcomings in practical applications, making it difficult to guarantee the timeliness and accuracy of fault warnings.

[0020] First, manual listening relies heavily on the operator's subjective experience and auditory sensitivity. In noisy kitchens or commercial environments, background noise (such as fans, air conditioners, and human voices) often masks subtle abnormal noises generated inside the refrigerator, making it difficult for inspectors to identify low-amplitude, short-duration abnormal sound waves. Even in quiet laboratory conditions, the human ear has limited ability to distinguish high-frequency transient impacts (such as the clicking sound of compressor valves or the impact sound of fan blades), making it easy to miss or misdiagnose. Furthermore, manual judgment cannot provide information on the spatial distribution of abnormal sounds; repair personnel can only guess the possible location of the fault based on experience, leading to blind and time-consuming disassembly and inspection.

[0021] Furthermore, single-channel acoustic sensors face severe signal-to-noise ratio (SNR) issues during the signal acquisition phase. The interior of a refrigerator is a closed cavity filled with reverberation and multipath propagation. Multiple vibration sources, such as the compressor, evaporator fan, and defrost heater, generate superimposed and out-of-phase sound waves. A single microphone can only capture pressure changes at a single point and cannot utilize spatial correlation to suppress incoherent noise or perform beamforming. Therefore, environmental noise (such as 50 / 60Hz power line interference and external vibration transmission) and internal reverberation significantly amplify the noise level, masking the energy of transient abnormal sounds and making it difficult to set a detection threshold that is both sensitive and robust. More importantly, single-channel data lacks directional information, thus failing to achieve three-dimensional localization of the noise source; even if it can determine that "abnormal noise exists," it cannot tell repair personnel whether the abnormal sound wave originates from the compressor side or the door seal.

[0022] In summary, overcoming environmental noise interference to improve the signal-to-noise ratio and achieving precise location and component identification of abnormal sounds by combining spatial coordinates, thereby improving the accuracy of abnormal sound capture and the efficiency of fault diagnosis, are the technical problems that urgently need to be solved in the field of refrigerator abnormal sound detection.

[0023] According to an embodiment of this application, an embodiment of an abnormal sound detection method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0024] This embodiment provides a method for detecting abnormal sounds. Figure 1 A flowchart of an abnormal sound detection method provided in an embodiment of this application is shown below. Figure 1 As shown, this process, applied to a detection system that includes a mobile acquisition module and a reference acquisition module, includes the following steps: Step S1: Control the mobile acquisition module to move within the detection space, and for each target step position, acquire acoustic signals and corresponding real-time spatial coordinates to construct a spatiotemporal sound field dataset.

[0025] Specifically, this embodiment adopts an "edge-cloud" collaborative architecture, mainly including: a perception layer, an edge computing layer, and a system decision layer. Preferably, the detection space is a refrigerator. The perception layer is used to monitor the acoustic signals and environmental reference signals during the refrigerator's operation in real time; the edge computing layer is used to preprocess, reduce noise, and perform preliminary feature extraction on the data collected by the perception layer; the system decision layer is used to receive the processed data, perform deep learning inference, and generate and display fault diagnosis results. The perception layer mainly consists of a multi-channel microphone array, UWB positioning tags, and temperature sensors. In order to achieve comprehensive coverage of the complex sound field inside the refrigerator and effective elimination of environmental noise, this embodiment physically divides the microphone array into a reference acquisition module and a mobile acquisition module.

[0026] like Figure 2 As shown, the reference acquisition module includes node C located outside the refrigerator. Node C is equipped with an industrial-grade MEMS PDM microphone, preferably placed on the back of the refrigerator. Figure 2 (c) or the side away from the internal sound source ( Figure 2 (b) in the middle.

[0027] like Figure 2 As shown in (a), the mobile acquisition module includes node A and node B, which are located inside the refrigerator (fresh food compartment and freezer compartment).

[0028] Node A (preservation layer): Equipped with 3 industrial-grade MEMS PDM microphones in a triangular array layout.

[0029] Node B (Frozen Layer): Equipped with 3 industrial-grade MEMS PDM microphones, also arranged in a triangular array.

[0030] To increase the depth and coverage of abnormal sound detection, the microphone arrays of both node A and node B are mounted on specially designed rails. For example... Figure 3 As shown, Figure 3 (a) in the figure is an isometric side view of the guide rail. Figure 3 (b) in the diagram is a side view of the guide rail. Figure 3 (c) in the figure is a front view of the guide rail. By extending or sliding the guide rail, the microphone array can move inside the refrigerator to collect acoustic signals at different target step positions and construct a spatiotemporal sound field dataset.

[0031] A UWB tag is placed on each of the guide sliders in the fresh food compartment and the freezer compartment, and a silicone shock-absorbing pad is added between the tag and the slider. The UWB tag is used to calculate the precise physical position (real-time spatial coordinates) of the microphone array in the refrigerator coordinate system in real time, providing a reference for sound source localization.

[0032] Considering the unique environment of high humidity and low temperature inside a refrigerator, all microphones in the sensing layer are wide-temperature industrial-grade products (operating temperature -40℃~85℃). An ePTFE waterproof and breathable membrane is attached to the microphone openings to prevent condensation from entering and causing the membrane to freeze; the PCB board adopts a double-sided conformal coating process to ensure the reliability of the hardware in harsh environments.

[0033] The edge computing layer is deployed on the refrigerator's main control board or in a separate detection box, with an STM32 series microcontroller (such as STM32F407) as the core, and a Wi-Fi transparent transmission module (such as ESP32) as the component.

[0034] The main control MCU (STM32F407) features a built-in multi-channel high-performance 12-bit ADC and an independent DMA controller, supporting high-fidelity synchronous acquisition of multi-channel microphone arrays. Utilizing its 168MHz main frequency and hardware floating-point unit (FPU), it runs the NLMS adaptive filtering algorithm locally, using an external reference signal to perform preliminary noise reduction on the internal signal; simultaneously, it performs IIR filtering and short-time energy detection to determine if the triggering conditions are met.

[0035] Network coprocessor (ESP32): Communicates with STM32 via SPI or serial port. When STM32 detects an abnormal sound trigger, ESP32 is responsible for encrypting and uploading the packaged "abnormal sound coordinate data + temperature data + short audio clip" to the cloud via the MQTT protocol, ensuring the real-time performance and stability of the transmission.

[0036] The system decision-making layer is deployed on cloud servers or local lightweight servers. This layer does not rely on expensive GPU clusters; a CPU with basic computing power is sufficient to meet the inference needs of the miniature convolutional neural network (CNN).

[0037] To avoid data redundancy and ensure positioning accuracy, this embodiment adopts a "step-stop-acquisition" triggering strategy, the specific process of which is as follows: Control the mobile acquisition module to move to the i-th target step position P i The target stepping position interval Δd is preset according to the Nyquist sampling theorem and the sound source localization accuracy requirements, for example, set to 0.5cm or 1cm. When the moving acquisition module reaches P... i When the device is positioned, the motor is paused to eliminate interference from mechanical vibration noise. During the pause, the microphone array is triggered to sample audio, and the UWB tag is simultaneously triggered to calculate coordinates. After acquisition is complete, the mobile acquisition module continues to move to the next position P. i+1 Until the entire detection space is covered (such as full coverage of the depth direction of the refrigerator's crisper compartment).

[0038] For each target step position, multimodal data is acquired synchronously, specifically including: Acoustic signal acquisition: Acoustic signal X acquired by a MEMS microphone array located at this location, such as an array inside the food preservation layer. fresh (n) Acoustic signal X collected by the array inside the cryolayer freeze (n). This signal includes abnormal noise signals generated by internal refrigerator components (such as fans and compressors) and ambient noise. Simultaneously, a UWB tag mounted on the slider emits an ultra-wideband pulse signal, which is received by a UWB base station deployed outside the refrigerator. Using either the Time Difference of Arrival (TDOA) algorithm or the Two-Way Time-of-Flight (TW-TOF) algorithm, the real-time spatial coordinates (x, y, x) of the current microphone array center point in the refrigerator's three-dimensional coordinate system are calculated in real time. i ,y i ,z i Read the temperature sensor value T for the current area. i This is used for subsequent sound speed correction.

[0039] Constructing a spatiotemporal sound field dataset refers to the process of aligning and encapsulating the unstructured acoustic signals acquired above with structured real-time spatial coordinates using timestamps. Specifically, a unique timestamp is assigned to each acquired acoustic signal, and it is bound to the real-time spatial coordinates acquired at the same time to form a multidimensional data tuple.

[0040] Step S3: Obtain the environmental reference signal collected by the reference acquisition module, use the environmental reference signal to filter the acoustic signal in the spatiotemporal sound field dataset to obtain a pure sound field signal, and when the pure sound field signal meets the preset triggering conditions, extract the corresponding target abnormal sound segment.

[0041] Specifically, an environmental reference signal is acquired synchronously by a reference acquisition module (such as a reference microphone). This signal mainly reflects the background noise during equipment operation (such as compressor operation noise, airflow noise, etc.). Then, the acoustic signal in the spatiotemporal sound field dataset is aligned and subtracted from this environmental reference signal to effectively cancel out steady-state background noise interference, resulting in a clean sound field signal with a high signal-to-noise ratio. Based on this, the clean sound field signal is monitored in real time. Once its energy, amplitude, or specific frequency band characteristics meet preset trigger conditions (i.e., a sudden or abnormal noise event is detected), an interception mechanism is immediately activated to accurately extract the target noise segment containing the abnormal event. This processing flow not only significantly improves the purity of the input data but also effectively filters out invalid silent or normal operating condition data, significantly reducing the computational load of subsequent feature extraction and classification, ensuring the accuracy and real-time performance of noise detection and localization.

[0042] Step S5: Based on the real-time spatial coordinates corresponding to the stepping positions of each target in the spatiotemporal sound field dataset, determine the spatial distribution relationship of the target abnormal sound segments at each target stepping position to reconstruct the spatial sound field energy distribution of the detection space, and determine the energy peak position of the spatial sound field energy distribution as the coordinates of the abnormal sound source that generates the target abnormal sound segments.

[0043] Specifically, the real-time spatial coordinates corresponding to the stepping positions of each target in the spatiotemporal sound field dataset are extracted and used as a dynamic reference. Then, combining the acoustic data collected from these different spatial locations of the target abnormal sound fragments, a coherent superposition and reconstruction of the sound field energy within the detection space is performed through multi-view spatial distribution relationship fusion analysis (e.g., based on time difference of arrival or beamforming algorithms) to generate a high-precision spatial sound field energy distribution map. In this reconstructed energy distribution map, the location with the most concentrated energy represents the actual sound source of the abnormal sound. Based on this, the energy peak position of the spatial sound field energy distribution is precisely locked as the coordinates of the abnormal sound source that generated the target abnormal sound fragment. This mechanism not only effectively eliminates the positioning ambiguity caused by reverberation interference but also fully utilizes the spatial diversity gain brought by dynamic topology, significantly improving the resolution and reliability of abnormal sound source localization.

[0044] Step S7: Perform feature extraction and classification on the target abnormal sound segment, determine the abnormal sound source component corresponding to the target abnormal sound segment, and generate fault diagnosis results by combining the coordinates of the abnormal sound source.

[0045] Specifically, the extracted target abnormal noise segments undergo in-depth feature extraction and classification. By capturing the unique characteristics of the abnormal noise signal in the time-frequency domain (such as specific harmonic structures or transient impact characteristics), a pre-trained recognition model is used to accurately determine the component to which the abnormal noise originates (such as a fan, compressor, or solenoid valve). Subsequently, this logical-level classification result is fused with the physical-level abnormal noise source coordinates obtained in the previous steps to generate a multi-dimensional information fusion, ultimately producing a fault diagnosis result that includes the type of faulty component and its precise three-dimensional location. This dual diagnostic mechanism of qualitative and localization not only effectively eliminates false alarm interference in complex environments but also provides maintenance personnel or automated operation and maintenance systems with intuitive and accurate decision-making basis, significantly improving the intelligence level and troubleshooting efficiency of equipment health status assessment.

[0046] This embodiment provides an abnormal noise detection method. By controlling a mobile acquisition module to step through the detection space, acoustic signals and corresponding real-time spatial coordinates are synchronously acquired at each target step position to construct a spatiotemporal sound field dataset. This not only overcomes the spatial coverage limitations of fixed microphone arrays, avoiding missed detection of abnormal noises in different areas inside the refrigerator, but also provides a multi-dimensional spatial sampling benchmark for subsequent sound field reconstruction. Subsequently, environmental reference signals acquired by the reference acquisition module are obtained, and the acoustic signals in the spatiotemporal sound field dataset are filtered to cancel out steady-state background noise interference such as refrigerator compressor operating noise and airflow noise, resulting in a clean sound field signal with a high signal-to-noise ratio. The corresponding target is intercepted only when the clean sound field signal meets the preset trigger conditions. Abnormal noise segments can filter out a large amount of invalid data under normal operating conditions, reducing the computational load of subsequent feature extraction and classification, and reducing the probability of misjudgment. Then, based on the real-time spatial coordinates corresponding to the step positions of each target in the spatiotemporal sound field dataset, the spatial distribution relationship of the target abnormal noise segments at each step position is determined to reconstruct the spatial sound field energy distribution of the detection space. The energy peak position is determined as the coordinates of the abnormal noise source, which can effectively eliminate reverberation interference inside the refrigerator and significantly improve the accuracy of abnormal noise localization. Finally, feature extraction and classification processing of the target abnormal noise segments are performed to determine the component from which the abnormal noise originates. The component classification results are fused with the coordinates of the abnormal noise source to generate fault diagnosis results, which can simultaneously output the type of faulty component and the precise spatial location, greatly improving the efficiency of fault diagnosis.

[0047] Figure 4 The flowchart for obtaining a pure sound field signal provided in the embodiments of this application may include the following steps: Step S311: Based on the preset bandpass filter coefficients, perform recursive differential operation on the acoustic signal to output a preliminary filtered signal.

[0048] Specifically, the acoustic signal s1(n) collected by the mobile acquisition module (such as the microphone array in the preservation layer or freezer layer) is acquired, including X fresh (n) and X freeze (n). To separate the valid abnormal noise signal from the irrelevant interference signal, the kernel floating-point unit (FPU) is invoked to accelerate floating-point operations, performing recursive differential operations on the acoustic signal. This embodiment selects a Butterworth filter structure with stable passband characteristics, implementing a fourth-order filter based on a second-order cascade method, strictly limiting the effective detection frequency band of the signal to between 300Hz and 7900Hz. Its time-domain filtering operation form is as follows: Where y(n) is the initial filtered signal; b i With a jThe bandpass filter coefficients are calculated offline and preset based on the abnormal sound frequency distribution characteristics of the refrigerator, and are directly called at runtime; M and N are the order of the filter. To avoid phase distortion during signal processing, a bidirectional recursive processing logic is used internally to effectively suppress energy masking of invalid frequency bands and improve the signal-to-noise ratio of subsequent abnormal sound transient feature extraction.

[0049] Step S313: Determine the residual signal between the environmental reference signal and the preliminary filtered signal, and dynamically adjust the cancellation weight of the environmental reference signal in the preliminary filtered signal based on the residual signal and the environmental reference signal, so as to iteratively approximate and remove the external crosstalk components related to the environmental reference signal, and output the preliminary denoised signal.

[0050] Specifically, in actual operating environments, noise from the refrigerator's external compartment (such as high-frequency compressor vibration and external fan noise) can penetrate and mask the internal acoustic characteristics. To isolate this dynamic interference in real time, the environmental reference signal X is retrieved. ex The signal y(n) and the initial filtered signal y(n) after filtering are used to perform preliminary detection and cancellation of crosstalk in the external environment using the Normalized Least Mean Square (NLMS) algorithm model.

[0051] First, the residual signal e(n) between the environmental reference signal and the preliminary filtered signal is determined, and its calculation process is as follows: Among them, y i (n) represents the initial filtered signal of the i-th array channel.

[0052] Subsequently, based on the residual signal e(n) and the environmental reference signal X ex (n), dynamically adjusting the cancellation weight w(n) of the environmental reference signal in the initial filtered signal. To avoid processing instability caused by drastic fluctuations in audio signal amplitude in the refrigerator scenario, the input signal power is normalized and adapted. The weight update process is as follows: Where w(n) is the dynamically maintained filter weight vector (i.e., cancellation weight) at the current time, μ is the step size factor that adjusts the convergence speed and steady-state error, and δ is the regularization minimization constant to prevent the energy term from being zero and causing computational overflow. After the above iterative approximation, the external crosstalk components related to the external environment are removed, and the output e(n) is the preliminary denoised signal.

[0053] Step S315: Perform frequency domain decomposition on the preliminary denoised signal to obtain the frequency domain amplitude spectrum of the current frame.

[0054] Specifically, the output preliminary denoised signal e(n) is subjected to frequency domain decomposition processing such as Fast Fourier Transform (FFT) to obtain the current frame frequency domain amplitude spectrum of the current audio frame. Where k is the frequency index, representing different frequency points; m is the channel index, representing different microphone acquisition channels in the preservation layer and the freezing layer.

[0055] Step S317: Determine the global mean square error between the current frame frequency domain amplitude spectrum and the preset background noise model. When the global mean square error is greater than the preset threshold, it is determined that the current frame frequency domain amplitude spectrum is in an abnormal sound state. Based on the frequency domain difference between the current frame frequency domain amplitude spectrum and the preset background noise model, a clean sound field signal is reconstructed.

[0056] Specifically, the frequency domain average amplitude spectrum is obtained by averaging multiple frames of pure background samples (i.e., mechanical operating noise of compressors and fans in a noise-free state). This is used as the preset background noise model.

[0057] Subsequently, the frequency domain amplitude spectrum of the current frame is calculated point by point. Compared with the preset background noise model The global mean square error (MSE) is used to assess the degree to which a signal deviates from its steady-state operating condition. The formula for calculating the global mean square error is as follows: Where K is the total number of frequency points in a single frame, and M is the total number of channels acquired.

[0058] When the global mean square error (MSE) exceeds a preset threshold, the current frame's frequency domain amplitude spectrum is determined to be in an abnormal noise state. At this point, the signal contains transient abnormal noise characteristics caused by component collisions or scratches. To eliminate the known steady-state background noise, a clean sound field signal is reconstructed based on the frequency domain difference between the current frame's frequency domain amplitude spectrum and a preset background noise model. To prevent artificial interference distortion caused by excessive reduction, a spectrum lower limit control mechanism is introduced, the frequency domain operation form of which is shown below: in, The signal is a pure sound field signal; α is an over-subtraction factor used to control the noise suppression strength; β is the lower limit coefficient of the spectrum for retaining weak background noise. After this frequency domain difference reconstruction process, the part that remains is the extremely significant abnormal sound energy feature, that is, the pure sound field signal.

[0059] Please see Figure 5 The comparison diagram of the time-domain waveform and spectrum provided for the embodiments of this application intuitively demonstrates the actual processing effect of the embodiments of this application in a complex refrigerator operating environment. Specifically: Time-domain waveform comparison: It can be seen that under normal operating conditions from 0 to 54 seconds, the amplitude of the pure signal after noise reduction is significantly suppressed and approaches zero, indicating that the background steady-state noise has been effectively filtered out; while the sudden abnormal sound pulse that appears at about 55 seconds, the pure signal after noise reduction completely retains its high amplitude characteristics, proving that the transient fault signal is not damaged while powerfully reducing noise.

[0060] Comparison of spectral energy distribution: The panoramic spectrum of the original signal shows dense horizontal stripes, representing strong steady-state harmonic interference generated by the compressor and fan. This interference almost drowns out the details of the entire frequency band. After spectral subtraction and background suppression, the steady-state harmonic energy that originally dominated the panoramic spectrum is greatly weakened, the background becomes cleaner, and the frequency band energy change corresponding to the moment of abnormal sound occurrence is more prominent.

[0061] In summary, the figure strongly demonstrates that the embodiments of this application can effectively separate steady-state background noise from non-steady-state abnormal sounds, significantly improve the signal-to-noise ratio, and provide high-quality data support for subsequent accurate triggering and feature recognition.

[0062] This embodiment utilizes preset bandpass filter coefficients to perform recursive differential operations on the acoustic signal, effectively suppressing electromagnetic / airborne noise and retaining only the effective frequency bands where refrigerator noise may occur, thus obtaining a preliminary filtered signal. Subsequently, by calculating the residual between this preliminary filtered signal and the environmental reference signal, and dynamically updating the cancellation weights based on the residual and the reference signal, iterative approximation and stripping of crosstalk components coupled with the external environment are achieved, outputting a preliminary denoised signal with the influence of external noise removed. A fast Fourier transform is performed on this preliminary denoised signal to achieve the conversion from the time domain to the frequency domain, obtaining the frequency domain amplitude spectrum of the current frame, providing frequency resolution for subsequent feature determination. Finally, the current frame spectrum is compared with the pre-established steady-state background noise model for global mean square error. When the error exceeds a preset threshold, it is determined to be an abnormal noise state. Based on the frequency domain difference between the current frame spectrum and the background model, a spectral subtraction method with over-subtraction factor and lower limit suppression is used to reconstruct the clean sound field signal. The entire process first removes out-of-band interference in the time domain, then uses an adaptive algorithm to eliminate external crosstalk, and then sensitively captures transient abnormal sound energy in the frequency domain by comparing it with the background noise model. While maintaining the integrity of the abnormal sound characteristics, it suppresses the steady-state background noise to the maximum extent, thereby significantly improving the signal-to-noise ratio and achieving reliable detection and localization of abnormal sounds inside the refrigerator.

[0063] Figure 6 A flowchart illustrating the update method of the preset background noise model provided in this application embodiment, the process may include the following steps: Step S3161: When the global mean square error is lower than the preset threshold, it is determined that the current frame frequency domain amplitude spectrum is in a noise-free state, and the current frame frequency domain amplitude spectrum is stored as a pure background noise sample in the circular buffer.

[0064] Specifically, continuous monitoring is performed using a microphone array. When the calculated global mean square error (MSE) is below a preset threshold, the current frame's frequency domain amplitude spectrum is determined to be noise-free. At this point, the initial denoising signal mainly consists of steady-state mechanical noise from components such as the compressor and fan, without transient noise characteristics caused by component collisions or scratches. During this stage, edge computing units (such as STM32) extract the initial denoising signals (i.e., X signals) from each channel of the preservation and freezing layers through a circular buffer. fresh (n), X freeze (n) The processed signal, which does not contain the external reference signal X. ex (n)) and the extracted preliminary denoised signal is used as a pure background noise sample and stored in a circular buffer in memory.

[0065] Step S3163: When the number of pure background noise samples accumulated in the circular buffer reaches a preset number of frames, the spectral amplitude of the preset number of pure background noise samples is averaged to update the preset background noise model.

[0066] Specifically, the system monitors the accumulated number of pure background noise samples in the circular buffer in real time. When the accumulated number of pure background noise samples reaches a preset frame count T, the background noise model update mechanism is triggered. The frequency-domain average amplitude spectrum is calculated by performing an arithmetic mean on a frequency-by-frequency and channel-by-channel basis on the spectral amplitude of the pure background noise samples at the preset frame count T. This updates the preset background noise model. The calculation process is as follows: Where T is the total number of steady-state background noise sample frames participating in the averaging calculation (i.e., the preset number of frames); t is the sample frame index, with a value range of t=1,2,...,T; Let be the spectral amplitude of the pure background noise sample in frame t at frequency k and channel m; This is the frequency domain average amplitude spectrum obtained after averaging multiple frames, which is the updated preset background noise model.

[0067] In this embodiment, if the global mean square error between the current frame's frequency domain amplitude spectrum and the pre-stored background noise model is below a preset threshold, it can be determined that the frame contains no abnormal noise. At this point, the initial denoising signal consists only of steady-state mechanical operating noise from components such as compressors and fans, without transient collision or scraping characteristics. The frequency domain amplitude spectrum of this frame is directly written into the circular buffer as a pure background noise sample, achieving continuous acquisition of steady-state background noise without incurring additional storage overhead. Subsequently, the number of pure background noise samples accumulated in the circular buffer is continuously monitored; when the accumulated number of frames reaches a preset number of frames, the background noise model update mechanism is triggered: the spectral amplitude of the preset number of samples in the buffer is arithmetically averaged point by point and channel by channel to obtain a new frequency domain average amplitude spectrum, which serves as the updated preset background noise model. Therefore, the global mean square error benchmark on which subsequent abnormal sound determination is based remains accurate, effectively reducing false alarms or missed detections caused by background noise drift, improving the robustness and sensitivity of abnormal sound detection. At the same time, all operations only require lightweight circular buffers and mean calculations, which facilitates low-latency, real-time background noise self-learning and model maintenance on embedded processors (such as STM32).

[0068] Figure 7 The flowchart for extracting the corresponding target abnormal sound segment provided in this application embodiment may include the following steps: Step S331: Perform frame segmentation processing on the pure sound field signal to determine the short-time energy value of each frame of the pure sound field signal.

[0069] Specifically, for pure abnormal sound signals Perform an inverse Fourier transform (IFFT) to restore the signal to its time domain. The transformation formula is shown below: Where L is the pre-set fixed number of sampling points per frame; s(p,m) is the purified heterophonic time-domain signal of channel m and the p-th time-domain sampling point after transformation; after completing the time-domain restoration, the STM32 main control board performs frame-by-frame processing on the purified heterophonic time-domain signal and calculates the short-time energy E of each frame of purified heterophonic time-domain signal. m This indicator directly reflects the intensity of transient abnormal sounds and is an important indicator for judging abnormal outbreaks. Quantitative analysis of the energy of pure abnormal sound time-domain signals allows for more accurate capture of extremely faint scraping and "clicking" sounds. Short-time energy E m The calculation formula is as follows: Step S333: Compare the short-time energy value with the preset dynamic trigger threshold, and generate a trigger command based on the comparison result.

[0070] Specifically, the short-time energy value E obtained through real-time monitoring and calculation. m When the short-term energy value Em When the dynamic trigger threshold preset by STM32 is exceeded, it is determined that a high-confidence suspected abnormal sound burst has been captured at the current moment. At this time, a trigger command is immediately generated to prepare for data interception and reporting.

[0071] Step S335: In response to the trigger command, the pure sound field signal is truncated based on the preset time window length to obtain the target audio segment.

[0072] Specifically, in response to a trigger command, the STM32 immediately retrieves an audio segment (i.e., the target audio segment) of a preset time length before and after the trigger moment from its internal storage. Simultaneously, it also extracts the UWB coordinate sequence at the corresponding sampling time. Subsequently, the STM32 sends the extracted target audio segment and its corresponding spatial coordinates to the cloud for detection via the Wi-Fi module, in order to proceed to the subsequent precise positioning and fault diagnosis stages. For example, in practical applications, a 5-second noise-reduced audio segment before and after the trigger noise can be extracted as the target audio segment and uploaded.

[0073] Please see Figure 8 The time-frequency analysis diagram provided for the embodiments of this application from a microscopic perspective demonstrates the ability to preserve signal details and the effect of background suppression at the moment of abnormal sound triggering.

[0074] Time-domain waveform comparison: The gray curve represents the original contaminated signal, and the blue curve represents the clean signal after noise reduction. During the silent period before 54.8 seconds, the blue curve almost completely fits the zero axis, proving that the steady-state noise has been completely filtered out; while at the moment of the abnormal sound outbreak at 54.9 seconds, the blue waveform not only completely preserves the impact amplitude of the original signal, but also clearly presents the subsequent decay oscillation process, indicating that no clipping or distortion was caused to the transient fault characteristics.

[0075] Spectral energy distribution comparison: The middle subplot shows severe local noise in the original signal, with high-energy compressor harmonics filling the entire frequency band, causing the abnormal sound feature (the vertical bright band at 2.5 seconds) to be partially submerged. The bottom subplot shows the final output after noise reduction, where the background harmonic energy is significantly suppressed, making the energy abrupt change in the frequency band containing the abnormal sound extremely bright and isolated. This significant improvement in signal-to-noise ratio directly verifies that this preprocessing step can provide excellent input quality for subsequent feature extraction.

[0076] In this embodiment, the denoised clean sound field signal is first converted back to a time-domain signal by performing an inverse Fourier transform. Then, the short-time energy value of each frame is calculated for the time-domain signal. This indicator can intuitively quantify the burst intensity of transient abnormal sounds. Compared with pure frequency domain judgment logic, it can effectively capture the characteristics of extremely weak abnormal sounds such as scratching and clicking, reducing the abnormal sound missed rate. Subsequently, the real-time calculated short-time energy value is compared with a preset dynamic trigger threshold. Only when the short-time energy exceeds the threshold is a high-confidence suspected abnormal sound burst determined and a trigger command is generated. This avoids small fluctuations in steady-state background noise triggering abnormal sounds. The invalid triggering also ensures that real abnormal sounds are not missed, improving the accuracy of trigger judgment; after responding to the trigger command, the pure sound field signal with a preset duration before and after the trigger time can be directly extracted as the target audio segment, and the spatial coordinates of the corresponding sampling time are extracted at the same time. Both are uploaded to the cloud, which not only ensures that the background noise of the uploaded audio segment has been fully filtered out and that there is no clipping distortion in the impact characteristics and attenuation oscillation process of the abnormal sound, but also provides complete spatiotemporal data support for subsequent accurate location of abnormal sounds and fault diagnosis, adapting to the low power consumption and low latency operation requirements of embedded devices.

[0077] Figure 9 The flowchart for reconstructing the spatial sound field energy distribution of the detection space provided in this application embodiment shows that the mobile acquisition module includes multiple acquisition units, and the process may include the following steps: Step S511: Extract the pure sound field signal collected by each acquisition unit at the same time from the target abnormal sound segment.

[0078] Step S513: Determine the cross-power spectrum between the corresponding pure sound field signals of any two acquisition units, and normalize the amplitude of the cross-power spectrum based on the phase transformation weighting function to obtain the generalized cross-correlation function.

[0079] Specifically, to address the severe reverberation interference caused by the confined metal cavity inside the refrigerator, a frequency domain cross-correlation combined with phase transform weighting (PHAT) processing method is employed. Preferably, the acquisition unit is a microphone array. Let any two channels within the microphone array be numbered i and j, and the corresponding pure heterophonic spectrum of the i-th and j-th output channels be respectively... and . Equivalent to a pure sound field signal with channel m=i , Equivalent to a pure sound field signal with channel m=j .

[0080] Cross-power spectrum G of the two signals ij (f) The calculation formula is shown below: in, Indicates to Take the conjugate.

[0081] Subsequently, based on the cross-power spectrum G ij (f) Perform PHAT phase-weighted cross-correlation with the time domain, the time domain operation of which is shown below: Where f is the frequency variable; τ is the time delay variable, representing the time difference parameter between the arrival of the sound wave at microphone pair i,j; R ij (τ) is the generalized cross-correlation function (cross-correlation coefficient) after PHAT weighting.

[0082] Step S515: Perform peak search on the generalized cross-correlation function and determine the time delay variable corresponding to the maximum peak as the measured arrival time difference between the acquisition unit pairs.

[0083] Specifically, the calculated generalized cross-correlation function R ij (τ) Perform peak search. By finding R... ij The location of the maximum peak value of (τ) can be used to obtain the microsecond-level time difference τ between the sound waves arriving at the two microphones. ij (t). This value is determined as the measured arrival time difference between the acquisition unit pairs, and finally forms the arrival time difference vector between each channel, providing the core data foundation for subsequent precise positioning.

[0084] Step S517: Construct a three-dimensional spatial grid covering the detection space, traverse each location of the sound source to be estimated in the three-dimensional spatial grid, determine the assumed distance difference between the location of the sound source to be estimated and each acquisition unit pair, and determine the degree of matching between the assumed distance difference and the measured physical distance difference corresponding to the measured arrival time difference.

[0085] Specifically, to address the ghosting issue caused by strong reverberation from the refrigerator's metal walls, a three-dimensional spatial mesh covering the refrigerator cavity detection space is pre-constructed in memory. Each point P in this three-dimensional spatial mesh (with 3D coordinates (X, Y, Z), representing the location of the sound source to be estimated) is traversed. For each point P, the distance difference Δd between the estimated sound source location P and microphones i and j is calculated using the real-time spatial coordinates of the microphone array fed back by UWB. ij (P) (i.e., the assumed distance difference). Then, this assumed distance difference Δd... ij (P) and the measured arrival time difference τ ij The measured physical distance difference c·τ is obtained by multiplying (t) by the speed of sound c. ij (t) Perform matching and comparison to determine the matching degree Δd between the two. ij (P)-c·τ ij (t).

[0086] Step S519: Determine the likelihood energy value of the sound source location to be estimated based on the matching degree, and aggregate the likelihood energy values ​​of all sound source locations to be estimated in the detection space to construct the spatial sound field energy distribution.

[0087] Specifically, the likelihood energy value E(P) of each sound source location P to be estimated is calculated using a likelihood probability mapping to reflect the confidence that the point is a true sound source. The calculation formula is shown below: Wherein, σ is a preset standard deviation coefficient used to control the focusing degree of the energy field. This σ is pre-configured according to the size of the refrigerator cavity: the larger the cavity size, the larger the value of σ, to ensure the coverage range of the energy field; the smaller the cavity size, the smaller the value of σ, to improve the positioning resolution.

[0088] Furthermore, multiple observations along the guide rail's movement path (i.e., time delay matching results and energy field data at different array positions) are used to coherently superimpose the energy distribution field, thereby improving the positioning resolution in the Y-axis direction. After traversing the entire three-dimensional spatial grid, the likelihood energy values ​​E(P) of all sound source locations to be estimated are aggregated to construct a complete spatial sound field energy distribution.

[0089] Please see Figure 10 The three-dimensional spatial visualization provided in the embodiments of this application intuitively demonstrates the actual positioning effect and energy distribution reconstruction capability within a complex cavity.

[0090] The figure constructs a three-dimensional coordinate system with a length, width, and height of approximately 0.8m × 0.8m × 1.8m (corresponding to the internal space of the refrigerator). The black dots represent the physical sampling positions of the microphone array at different times, and the gray plane indicates the internal shelf structure, thus establishing the spatial reference for positioning.

[0091] The colored clouds in the image represent the reconstructed spatial sound field energy distribution, with colors ranging from blue to red indicating increasing energy. Two main high-energy regions (red core) are clearly visible: Bottom region (Z≈0-0.4m): It presents a dispersed and complex energy cluster, which usually corresponds to the low-frequency noise source generated by the compressor or mechanical vibration. Due to its proximity to the boundary, it is greatly affected by reverberation and has a wide energy distribution.

[0092] The upper region (Z≈1.3-1.5m) presents a very concentrated and independent spherical high-energy core area, which precisely corresponds to a specific point sound source (such as abnormal noise from a fan or howling airflow), indicating that it has extremely high focusing ability.

[0093] This figure demonstrates that the embodiments of this application can not only determine the presence of abnormal noise, but also accurately depict the shape and location of the abnormal noise in three-dimensional space, successfully transforming abstract acoustic signals into visualized spatial coordinates, providing a direct basis for the physical troubleshooting of faulty components.

[0094] This embodiment synchronously extracts the clean sound field signals of each acquisition unit at the same time from the target abnormal sound segment, which can effectively eliminate background noise and steady-state interference, providing high-quality input for subsequent time delay estimation. The cross-power spectrum of any two clean signals is calculated and amplitude normalized using phase transformation weighting to obtain a generalized cross-correlation function. This weighting mechanism can suppress the influence of strong reverberation and random noise from the refrigerator's metal cavity on the amplitude of the cross-power spectrum, highlighting the actual arrival time difference characteristics of the sound waves. Subsequently, peak search is performed on the generalized cross-correlation function, and the time delay corresponding to the maximum peak is regarded as the measured arrival time difference of the pair of acquisition units, significantly... This process improves the accuracy of time delay estimation and its anti-reverberation capability. A three-dimensional mesh is constructed within the detection space, traversing the location of each sound source to be estimated. The assumed distance difference is calculated based on the known coordinates of the acquisition unit and matched with the measured physical distance difference obtained from the measured arrival time difference. This quantifies the consistency between the location of each sound source to be estimated and the actual sound source within the spatial range. Finally, the likelihood energy value of each sound source location is calculated based on the matching degree, and all likelihood energy values ​​are aggregated to form a spatial sound field energy distribution. This ensures that the energy peaks at the actual sound source location, thus accurately outputting the three-dimensional spatial coordinates of the abnormal sound. In summary, this process, through clean signal extraction, weighted generalized cross-correlation time delay estimation, three-dimensional mesh matching, and likelihood energy aggregation, successfully suppresses strong reverberation interference from the narrow metal cavity of the refrigerator, significantly improving the accuracy of time delay estimation and spatial resolution. It achieves high-precision, non-disassembly-required three-dimensional localization of abnormal sound sources, providing a reliable basis for rapid fault diagnosis.

[0095] Figure 11 The flowchart provided in this application, which determines the energy peak position of the spatial sound field energy distribution as the coordinates of the source of the target abnormal sound segment, may include the following steps: Step S531: The location of the sound source to be estimated with the largest likelihood energy value in the spatial sound field energy distribution is determined as the energy peak location, and the coordinates corresponding to the energy peak location are determined as the coordinates of the abnormal sound source.

[0096] Specifically, the energy distribution of the superimposed spatial sound field is iteratively optimized, and the location of the sound source with the largest likelihood energy value E(P) is determined as the energy peak location. This energy peak location represents the most likely physical location where the abnormal sound is to occur. Finally, the three-dimensional coordinates corresponding to this energy peak location are extracted as the coordinates of the abnormal sound source. At the same time, the energy values ​​of all spatial grid points inside the refrigerator cavity are mapped as visual attributes to generate a sound source localization point cloud map, which intuitively presents the correspondence between energy distribution and sound source location.

[0097] This embodiment performs iterative optimization calculations on the spatial sound field energy distribution, and determines the location of the sound source with the highest likelihood energy value as the energy peak location. This location has the highest matching degree with the actual arrival time difference measured by each acquisition unit among all the sound source locations to be estimated. It completely corresponds to the real physical location where the abnormal sound is most likely to occur. The accurate coordinates of the abnormal sound source can be obtained by directly extracting the three-dimensional coordinates corresponding to this energy peak location. It can effectively filter out the interference of false energy peaks caused by strong reverberation and reflection signals from the refrigerator's metal cavity, avoid the positioning deviation problem caused by false peaks in traditional positioning methods, and accurately locate the location of the abnormal sound fault without disassembling the refrigerator, greatly reducing the manpower and time costs of abnormal sound troubleshooting.

[0098] Figure 12 The flowchart for step S7 provided in the embodiments of this application may include the following steps: Step S71: Perform Mel spectrum transformation on the target abnormal sound segment to generate a two-dimensional time-frequency feature map.

[0099] Specifically, the target abnormal sound segments are processed using Mel-spectrogram transformation. This process converts the one-dimensional time-domain signal into a two-dimensional time-frequency feature map with a two-dimensional time-frequency distribution. Since different abnormal sounds have different inherent physical characteristics, such as the high-frequency noise details generated by component scratches and the transient pulse characteristics generated by mechanical impacts, the energy distribution of this two-dimensional time-frequency feature map can accurately reflect these physical characteristics, providing a high-quality data foundation for subsequent deep learning classification.

[0100] Step S73: Input the two-dimensional time-frequency feature map into the pre-trained lightweight convolutional neural network classification model, use the lightweight convolutional neural network classification model to perform feature analysis on the two-dimensional time-frequency feature map, and output the classification confidence of the target abnormal sound segment belonging to the preset component category.

[0101] Specifically, the generated two-dimensional time-frequency feature map is input into a pre-trained lightweight convolutional neural network (CNN) classification model. This model performs deep analysis and feature extraction on the input two-dimensional features, and maps them to the confidence distribution of each preset component category using a probability normalization function. The classification decision logic is as follows: Wherein, s(p,m) is the pure abnormal sound time-domain signal; Mel(s(p,m)) is the Mel spectrum processing module, used to convert the pure abnormal sound time-domain signal s(p,m) corresponding to the target abnormal sound segment into a two-dimensional time-frequency feature map; CNN is a pre-trained lightweight convolutional neural network classification model, used to analyze the input two-dimensional time-frequency feature map and output the classification probability of each preset component category; Softmax is the probability normalization function, used to map the classification probability output by the model to the classification confidence of each preset component category; Class Result The final classification result output by the model corresponds to different preset component categories for different sources of abnormal noise (such as fans, solenoid valves, evaporators, etc.). The classification confidence reflects the probability that the target abnormal noise segment belongs to each preset component category.

[0102] Step S75: The preset component category with the highest classification confidence is determined as the component that causes the abnormal noise, and combined with the coordinates of the source of the abnormal noise, a fault diagnosis result containing the component that causes the abnormal noise and the coordinates of the source of the abnormal noise is generated.

[0103] Specifically, after obtaining the Class Result Next, a global optimization process is performed to identify the component category with the highest classification confidence as the source of the abnormal noise. Then, the identified source component is fused with the coordinates of the noise source obtained in the previous steps to generate a fault diagnosis result containing both the source component and its coordinates. This result not only clarifies the type of component emitting the abnormal noise but also indicates its three-dimensional physical location inside the refrigerator, thus achieving automatic identification of the source of refrigerator noise and accurate auxiliary assessment of its health status.

[0104] This embodiment transforms the one-dimensional time-domain noise signal into a two-dimensional time-frequency feature map by performing Mel-frequency transform on the target noise segment. Compared to the original one-dimensional signal, this significantly improves the information density and discriminative power of the features, providing a high-quality feature input foundation for subsequent deep learning classification. The generated two-dimensional time-frequency feature map is then input into a lightweight convolutional neural network classification model pre-trained based on a large number of refrigerator noise samples. The model automatically extracts deep time-domain and frequency-domain correlation features of the noise signal, avoiding the limitations of manual feature design without requiring manually designed feature rules. The output confidence scores for each preset component category intuitively reflect the probability distribution of the target noise belonging to different components such as fans, solenoid valves, and evaporators. The lightweight model... The design significantly reduces computing power consumption while ensuring classification accuracy, adapting to the low computing power deployment requirements of localized refrigerator terminals, and enabling real-time inference for abnormal noise classification. Finally, through global optimization, the preset component category with the highest classification confidence is selected as the component source of the abnormal noise. This category is then fused with the coordinates of the abnormal noise source obtained in the previous steps to generate a fault diagnosis result that includes both the type of the abnormal noise source component and its corresponding spatial location. This not only clarifies the attributes of the sound-emitting component but also locks its specific physical location within the refrigerator cavity. Accurate tracing of abnormal noise faults can be completed without disassembling the refrigerator, greatly reducing the difficulty and time cost of manual troubleshooting. It can also provide structured data support for the intelligent assessment of the refrigerator's health status.

[0105] Accordingly, please refer to Figure 13 A block diagram of an abnormal sound detection device provided in this application embodiment, applied to a detection system including a motion acquisition module and a reference acquisition module, the device includes: The signal acquisition unit 101 is used to control the movement of the mobile acquisition module in the detection space, and to acquire acoustic signals and corresponding real-time spatial coordinates for each target step position to construct a spatiotemporal sound field dataset. The signal processing unit 103 is used to acquire the environmental reference signal acquired by the reference acquisition module, use the environmental reference signal to filter the acoustic signal in the spatiotemporal sound field dataset to obtain a pure sound field signal, and when the pure sound field signal is detected to meet the preset triggering conditions, the corresponding target abnormal sound segment is extracted. The sound source localization unit 105 is used to determine the spatial distribution relationship of the target abnormal sound segment at each target step position based on the real-time spatial coordinates corresponding to each target step position in the spatiotemporal sound field dataset, so as to reconstruct the spatial sound field energy distribution of the detection space, and determine the energy peak position of the spatial sound field energy distribution as the coordinates of the abnormal sound source that generates the target abnormal sound segment. The fault diagnosis unit 107 is used to extract and classify the features of the target abnormal sound segment, determine the abnormal sound source component corresponding to the target abnormal sound segment, and generate fault diagnosis results by combining the coordinates of the abnormal sound source.

[0106] In some alternative implementations, the signal processing unit 103 includes: Based on the preset bandpass filter coefficients, a recursive differential operation is performed on the acoustic signal to output a preliminary filtered signal; The residual signal between the environmental reference signal and the preliminary filtered signal is determined, and the cancellation weight of the environmental reference signal in the preliminary filtered signal is dynamically adjusted based on the residual signal and the environmental reference signal to iteratively approximate and remove the external crosstalk components related to the environmental reference signal, and output the preliminary denoised signal. Perform frequency domain decomposition on the initially denoised signal to obtain the frequency domain amplitude spectrum of the current frame; The global mean square error between the current frame frequency domain amplitude spectrum and the preset background noise model is determined. When the global mean square error is greater than the preset threshold, the current frame frequency domain amplitude spectrum is determined to be in an abnormal sound state. Based on the frequency domain difference between the current frame frequency domain amplitude spectrum and the preset background noise model, a clean sound field signal is reconstructed.

[0107] In some alternative implementations, the signal processing unit 103 includes: When the global mean square error is lower than the preset threshold, the current frame frequency domain amplitude spectrum is determined to be in a noise-free state, and the current frame frequency domain amplitude spectrum is stored as a pure background noise sample in the circular buffer. When the number of pure background noise samples accumulated in the circular buffer reaches a preset number of frames, the spectral amplitude of the preset number of pure background noise samples is averaged to update the preset background noise model.

[0108] In some alternative implementations, the signal processing unit 103 includes: The pure sound field signal is processed by frame segmentation to determine the short-time energy value of each frame of the pure sound field signal; The short-term energy value is compared with the preset dynamic trigger threshold, and a trigger command is generated based on the comparison result. In response to the trigger command, the target audio segment is obtained by truncating the pure sound field signal based on the preset time window length.

[0109] In some optional implementations, the mobile acquisition module includes multiple acquisition units, and the sound source localization unit 105 includes: Extract the pure sound field signal collected by each acquisition unit at the same time from the target abnormal sound segment; The cross-power spectrum between the corresponding pure sound field signals of any two acquisition units is determined, and the amplitude of the cross-power spectrum is normalized based on the phase transformation weighting function to obtain the generalized cross-correlation function. Peak search is performed on the generalized cross-correlation function, and the time delay variable corresponding to the maximum peak value is determined as the measured arrival time difference between the acquisition unit pairs; A three-dimensional spatial grid covering the detection space is constructed, and the location of each sound source to be estimated in the three-dimensional spatial grid is traversed. The assumed distance difference between the location of the sound source to be estimated and each acquisition unit pair is determined, and the matching degree between the assumed distance difference and the measured physical distance difference corresponding to the measured arrival time difference is determined. The likelihood energy value of the sound source location to be estimated is determined based on the matching degree, and the likelihood energy values ​​of all sound source locations to be estimated in the detection space are aggregated to construct the spatial sound field energy distribution.

[0110] In some alternative implementations, the sound source localization unit 105 includes: The location of the sound source with the highest likelihood energy value in the spatial sound field energy distribution is determined as the energy peak location, and the coordinates corresponding to the energy peak location are determined as the coordinates of the abnormal sound source.

[0111] In some optional implementations, the fault diagnosis unit 107 includes: The target abnormal sound segment is processed by Mel spectrum transformation to generate a two-dimensional time-frequency feature map; The two-dimensional time-frequency feature map is input into a pre-trained lightweight convolutional neural network classification model. The lightweight convolutional neural network classification model is used to perform feature analysis on the two-dimensional time-frequency feature map and output the classification confidence of the target abnormal sound segment belonging to the preset component category. The component category with the highest classification confidence is identified as the component from which the abnormal noise originates. Combined with the coordinates of the source of the abnormal noise, a fault diagnosis result containing both the component from which the abnormal noise originates and its coordinates is generated.

[0112] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0113] In this embodiment, an abnormal noise detection device is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above-mentioned functions.

[0114] Please see Figure 14 , Figure 14 This application provides a schematic diagram of the structure of a computer device, as shown in the embodiment of the present application. Figure 14As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 14 Take a processor 10 as an example.

[0115] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0116] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0117] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0118] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0119] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0120] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the methods shown in the above embodiments are implemented.

[0121] The systems, devices, and units described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0122] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0123] Those skilled in the art will understand that the embodiments of this application can be provided as methods or apparatus. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0124] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, and devices according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0125] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0126] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0127] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0128] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0129] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0130] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for detecting abnormal sounds, characterized in that, A detection system comprising a mobile acquisition module and a reference acquisition module, wherein the acquisition module is disposed inside a detection space and the reference acquisition module is disposed outside the detection space, wherein the detection space is a refrigerator, the method comprising: The mobile acquisition module is controlled to move within the detection space. For each target step position, the motor is controlled to pause its movement. During the pause, acoustic signals and corresponding real-time spatial coordinates are acquired to construct a spatiotemporal sound field dataset. The step interval of the target step position is preset based on the Nyquist sampling theorem and the sound source localization accuracy requirements. The environmental reference signal acquired by the reference acquisition module is obtained, and a recursive differential operation is performed on the acoustic signal based on the preset bandpass filter coefficients to output a preliminary filtered signal. The residual signal between the environmental reference signal and the preliminary filtered signal is determined, and based on the residual signal and the environmental reference signal, the cancellation weight of the environmental reference signal in the preliminary filtered signal is dynamically adjusted to iteratively approximate and remove the external crosstalk components related to the environmental reference signal, and output the preliminary denoised signal. The initial denoised signal is subjected to frequency domain decomposition to obtain the frequency domain amplitude spectrum of the current frame; The global mean square error between the current frame frequency domain amplitude spectrum and the preset background noise model is determined. When the global mean square error is greater than a preset threshold, the current frame frequency domain amplitude spectrum is determined to be in an abnormal sound state. Based on the frequency domain difference between the current frame frequency domain amplitude spectrum and the preset background noise model, a clean sound field signal is reconstructed. When the pure sound field signal is detected to meet the preset triggering conditions, the corresponding target abnormal sound segment is extracted; The preset background noise model is updated in the following way: When the global mean square error is lower than a preset threshold, it is determined that the current frame frequency domain amplitude spectrum is in a noise-free state, and the current frame frequency domain amplitude spectrum is stored as a pure background noise sample in a circular buffer. When the number of pure background noise samples accumulated in the circular buffer reaches a preset number of frames, the spectral amplitude of the preset number of pure background noise samples is averaged to update the preset background noise model. Based on the real-time spatial coordinates corresponding to each target step position in the spatiotemporal sound field dataset, the spatial distribution relationship of the target abnormal sound segment under each target step position is determined to reconstruct the spatial sound field energy distribution of the detection space, and the energy peak position of the spatial sound field energy distribution is determined as the abnormal sound source coordinates that generate the target abnormal sound segment; The target abnormal sound segment is subjected to feature extraction and classification processing to determine the abnormal sound source component corresponding to the target abnormal sound segment, and a fault diagnosis result is generated by combining the coordinates of the abnormal sound source.

2. The method according to claim 1, characterized in that, When the pure sound field signal is detected to meet the preset triggering conditions, the corresponding target abnormal sound segment is extracted, including: The pure sound field signal is processed by frame segmentation to determine the short-time energy value of each frame of the pure sound field signal; The short-term energy value is compared with a preset dynamic trigger threshold, and a trigger command is generated based on the comparison result. In response to the trigger command, the pure sound field signal is truncated based on a preset time window length to obtain the target audio segment.

3. The method according to claim 1, characterized in that, The mobile acquisition module includes multiple acquisition units, and the reconstruction of the spatial sound field energy distribution of the detection space includes: Extract the pure sound field signal collected by each acquisition unit at the same time from the target abnormal sound segment; Determine the cross-power spectrum between the corresponding pure sound field signals of any two acquisition units, and normalize the amplitude of the cross-power spectrum based on the phase transformation weighting function to obtain the generalized cross-correlation function; A peak search is performed on the generalized cross-correlation function, and the time delay variable corresponding to the maximum peak value is determined as the measured arrival time difference between the acquisition unit pairs. A three-dimensional spatial grid covering the detection space is constructed, and each location of the sound source to be estimated in the three-dimensional spatial grid is traversed. The assumed distance difference between the location of the sound source to be estimated and each pair of acquisition units is determined, and the matching degree between the assumed distance difference and the measured physical distance difference corresponding to the measured arrival time difference is determined. Based on the matching degree, the likelihood energy value of the location of the sound source to be estimated is determined, and the likelihood energy values ​​of all locations of the sound source to be estimated in the detection space are aggregated to construct the spatial sound field energy distribution.

4. The method according to claim 3, characterized in that, Determining the energy peak position of the spatial sound field energy distribution as the coordinates of the source of the abnormal sound segment that generates the target abnormal sound segment includes: The location of the sound source with the largest likelihood energy value in the spatial sound field energy distribution is determined as the energy peak location, and the coordinates corresponding to the energy peak location are determined as the coordinates of the abnormal sound source.

5. The method according to claim 1, characterized in that, The process of extracting and classifying features from the target abnormal sound segment, determining the abnormal sound source component corresponding to the target abnormal sound segment, and generating a fault diagnosis result by combining the coordinates of the abnormal sound source, includes: The target abnormal sound segment is subjected to Mel-frequency transform processing to generate a two-dimensional time-frequency feature map; The two-dimensional time-frequency feature map is input into a pre-trained lightweight convolutional neural network classification model. The lightweight convolutional neural network classification model is used to perform feature analysis on the two-dimensional time-frequency feature map and output the classification confidence of the target abnormal sound segment belonging to the preset component category. The component category with the highest classification confidence is identified as the component that causes the abnormal noise. Combined with the coordinates of the source of the abnormal noise, a fault diagnosis result containing the component that causes the abnormal noise and the coordinates of the source of the abnormal noise is generated.

6. An apparatus for implementing the abnormal sound detection method according to any one of claims 1-5, characterized in that, An apparatus for use in a detection system comprising a mobile acquisition module and a reference acquisition module, the apparatus comprising: The signal acquisition unit is used to control the movement of the mobile acquisition module within the detection space, and to acquire acoustic signals and corresponding real-time spatial coordinates for each target step position to construct a spatiotemporal sound field dataset. The signal processing unit is used to acquire the environmental reference signal acquired by the reference acquisition module, use the environmental reference signal to filter the acoustic signal in the spatiotemporal sound field dataset to obtain a clean sound field signal, and when the clean sound field signal is detected to meet the preset triggering conditions, the corresponding target abnormal sound segment is extracted. The sound source localization unit is used to determine the spatial distribution relationship of the target abnormal sound segment at each target step position based on the real-time spatial coordinates corresponding to each target step position in the spatiotemporal sound field dataset, so as to reconstruct the spatial sound field energy distribution of the detection space, and determine the energy peak position of the spatial sound field energy distribution as the coordinates of the abnormal sound source that generates the target abnormal sound segment; The fault diagnosis unit is used to extract and classify the target abnormal sound segment, determine the abnormal sound source component corresponding to the target abnormal sound segment, and generate a fault diagnosis result by combining the coordinates of the abnormal sound source.

7. A computer device, characterized in that, include: The device includes a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the abnormal sound detection method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the abnormal sound detection method according to any one of claims 1 to 5.

Images