Device abnormal sound detection method and apparatus, device, and readable storage medium

Abstract

The application discloses a device abnormal sound detection method and device, equipment and readable storage medium, and relates to the technical field of fault diagnosis. The device abnormal sound detection method comprises the following steps: acquiring a to-be-detected audio signal of a to-be-detected device; inputting an acoustic feature of the to-be-detected audio signal into a first detection model to obtain a first detection result; wherein the first detection model is an adaptive model constructed based on machine learning and having an online parameter updating capability; judging whether the first detection result meets a preset condition; if the first detection result meets the preset condition, inputting the acoustic feature into a second detection model to obtain a second detection result; wherein the second detection model is a qualitative model preconfigured and having unchanged model parameters; and outputting a third detection result according to the first detection result and the second detection result. Through the above manner, the problem that the prior art cannot simultaneously consider high sensitivity of micro abnormal sound and low false alarm rate of legal transient bottom noise in an actual production line can be solved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method, apparatus, device, and readable storage medium for detecting abnormal noise in equipment. Background Technology

[0002] In industrial acoustic quality inspection (such as loudspeaker quality testing), when a product plays a sweep frequency signal or a specific test audio, it will not only expose weak abnormal noises (such as scratching sounds or air leakage sounds), but will also generate legitimate transient background noise during normal mechanical start-stop, commutation and other processes.

[0003] Current mainstream detection methods mostly employ whole-segment audio feature calculation combined with global static threshold judgment, or directly apply end-to-end deep neural network models. However, because the characteristics of weak abnormal sound signals and legitimate transient background noise in the time-frequency domain often highly overlap and have blurred boundaries, existing methods cannot simultaneously achieve high sensitivity for weak abnormal sounds and low false alarm rate for legitimate transient background noise in actual production lines. Specifically, if the detection sensitivity is increased to capture minute defect signals, a large number of legitimate transient mechanical sounds will be misjudged as defective products; if the sensitivity is reduced to filter out the aforementioned transient background noise, it is very easy to miss weak abnormal sounds.

[0004] Therefore, how to establish a stable mechanism to distinguish between faint abnormal sounds and legitimate transient background noise in complex temporal audio streams remains a technical challenge that urgently needs to be solved in the field of device abnormal sound detection. Summary of the Invention

[0005] The main objective of this invention is to propose a method, apparatus, device, and readable storage medium for detecting abnormal noise in equipment, aiming to solve the problem that existing solutions cannot simultaneously achieve high sensitivity for minute abnormal noises and low false alarm rate for legitimate transient noise in actual production lines.

[0006] The present invention provides a method for detecting abnormal noise in equipment, comprising: Acquire the audio signal to be tested from the device under test; The acoustic features of the audio signal to be detected are input into the first detection model to obtain the first detection result; wherein, the first detection model is an adaptive model built based on machine learning and has the ability to update parameters online; Determine whether the first detection result meets the preset conditions; If the first detection result meets the preset conditions, the acoustic features are input into the second detection model to obtain the second detection result; wherein, the second detection model is a pre-configured qualitative model with unchanged model parameters; Based on the first detection result and the second detection result, a third detection result is output.

[0007] In one embodiment, prior to the step of inputting the acoustic features of the audio signal to be detected into the first detection model, the method further includes: The audio signal to be detected is divided into multiple consecutive time intervals along the time axis; Extract the acoustic features corresponding to each time interval; The step of inputting the acoustic features of the audio signal to be detected into the first detection model to obtain the first detection result includes: The acoustic features corresponding to the current time interval are input into the first detection model to obtain the first detection result of the current time interval.

[0008] In one embodiment, the step of determining whether the first detection result meets the preset conditions includes: Extract a first probability value from the first detection result indicating that there is an anomaly in the current time interval; The first probability value is compared numerically with the first judgment threshold; If the first probability value is greater than or equal to the first determination threshold, then the first detection result is determined to meet the preset condition.

[0009] In one embodiment, the third detection result includes normal results and abnormal results; the step of outputting the third detection result based on the first detection result and the second detection result includes: Extract a second probability value indicating an anomaly in the current time interval from the second detection result; Determine whether the second probability value is greater than or equal to a pre-configured second determination threshold; If the result is greater than or equal to the second determination threshold, then the abnormal result is generated; If the result is less than the second threshold, then the normal result is generated.

[0010] In one embodiment, when performing the step of inputting the acoustic features into the second detection model, the calculation of signal physical indices for the current time interval is performed simultaneously; wherein, the signal physical indices include at least one of short-time energy fluctuation rate, high-frequency subband energy ratio, zero-crossing rate standard deviation, and spectral centroid shift; the second detection result includes a second probability value indicating that an anomaly exists in the current time interval; the step of outputting a third detection result based on the first detection result and the second detection result includes: Determine whether the second probability value is greater than or equal to a pre-configured second determination threshold, and determine whether the signal physical index has reached a preset physical transition boundary condition; If the second probability value is greater than or equal to the second judgment threshold, and the signal physical index reaches the physical jump boundary condition, then the abnormal result is generated; If the second probability value is less than the second determination threshold, or if the signal physical index does not reach the physical transition boundary condition, then the normal result is generated.

[0011] In one embodiment, the step of constructing the second detection model includes: Obtain the offline acoustic features of pre-collected historical abnormal sound sample sets and historical normal sound sample sets; An initial ensemble model is constructed using multiple decision trees, and a preset upper limit for the maximum depth of the trees and a lower limit for the minimum number of samples required for node splitting are configured for the ensemble model. The ensemble model is trained under supervision using the offline acoustic features. During the growth of each decision tree, splitting stops when the preset upper limit is reached or the number of samples in the current node is lower than the minimum number of samples. Freeze the network weights of the ensemble model after training is completed, and use the frozen ensemble model as a second detection model with the model parameters remaining unchanged.

[0012] In one embodiment, the first determination threshold is obtained through the following steps: Construct a timing parameter scheduling mapping table, and independently configure the initial judgment threshold corresponding to each timing interval in the timing parameter scheduling mapping table; Obtain the product batch identifier of the device under test, and / or collect the environmental reference noise data of the current test environment; Based on the product batch identifier and / or the environmental reference noise data, the initial judgment threshold in the time-series parameter scheduling mapping table is dynamically corrected, and the corrected threshold is synchronously called as the first judgment threshold according to the time flow during the detection process.

[0013] The present invention also provides a device for detecting abnormal noise in equipment, the device comprising: The audio acquisition module is used to acquire the audio signal to be tested from the device under test. The processing module is used to divide the audio signal to be detected into multiple time intervals and extract the acoustic features corresponding to each time interval; The first detection module is used to receive the acoustic features corresponding to each of the time intervals, output the first detection result, and has the ability to update parameters online. The second detection module is used to receive the acoustic features of the corresponding time interval and output the second detection result when the first detection result meets the preset conditions, and the parameters of the second detection module remain fixed. The output module is used to output a third detection result based on the first detection result and the second detection result.

[0014] The present invention also provides a device for detecting abnormal noise, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the device noise detection method described above.

[0015] The present invention also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the device noise detection method described above.

[0016] The device noise detection method provided by this invention solves the problem that existing solutions cannot simultaneously achieve high sensitivity for subtle noises and low false alarm rate for legitimate transient noise in actual production lines by employing a cascaded triggering mechanism that combines adaptive initial screening with deterministic verification. Specifically, in the detection process, the acoustic features of the acquired audio signal to be detected are first input into an adaptive first detection model with online parameter updates to obtain a first detection result. Subsequently, instead of processing all the data, the acoustic features are only input into a pre-configured qualitative second detection model with unchanged model parameters when the first detection result meets preset conditions to obtain a second detection result. Finally, a third detection result is output based on the first and second detection results. Through the above settings, the first detection model, which has the ability to update parameters online, can dynamically adapt to fluctuations in production line data and capture weak abnormal noise signals to ensure detection sensitivity. At the same time, a trigger mechanism that "meets preset conditions" is used to initially screen signals. For suspected signals that enter the review stage, a second detection model with unchanged model parameters provides a fixed judgment boundary, filtering out legitimate transient noise that is misjudged by the adaptive model. This suppresses the drift of the judgment boundary and the tendency of overfitting, thereby reducing the false alarm rate caused by transient noise while ensuring the detection rate of minor abnormal noises. This achieves an effective balance between high sensitivity and low false alarm rate. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from the structures shown in these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the first embodiment of a device noise detection method according to the present invention; Figure 2This is a flowchart illustrating a second embodiment of a device noise detection method according to the present invention; Figure 3 This is a flowchart illustrating a third embodiment of a device noise detection method according to the present invention; Figure 4 This is a flowchart illustrating the fourth embodiment of the device noise detection method of the present invention; Figure 5 This is a flowchart illustrating the fifth embodiment of the device noise detection method of the present invention; Figure 6 This is a flowchart illustrating the sixth embodiment of a device noise detection method according to the present invention; Figure 7 This is a flowchart illustrating the seventh embodiment of a device noise detection method according to the present invention; Figure 8 This is a structural block diagram of the first embodiment of the abnormal noise detection device of the present invention; Figure 9 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention.

[0019] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0021] Reference Figure 9 , Figure 9 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention.

[0022] like Figure 9As shown, the device may include: a processor 1001 (such as an industrial-grade ARM Cortex series microcontroller or FPGA), a communication bus 1002, a user interface 1003, a network interface 1004, a memory 1005, a sensor module, an audio processing module, an actuator module, and a power supply module. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include physical buttons (such as a start test button, an emergency stop button, and a mode switch button) and status indicator lights. The physical buttons are used to receive manual operation commands from the user, and the status indicator lights use different colors or flashing frequencies to reflect the device's power-on / power-off status, test running status, network connection status, and abnormal alarm status. Optionally, the user interface 1003 may also include an industrial communication interface (such as an RS485 interface or an Ethernet port) to accommodate both power supply and production line data interface requirements. All interfaces are equipped with dustproof and surge-proof components to adapt to the complex electromagnetic environment of industrial workshops. The network interface 1004 is a wired Ethernet module or an industrial wireless communication module, used to establish a network connection with terminal devices (host computer, PLC controller, etc.) to achieve low-latency transmission of raw audio signals and detection results. It also supports users to remotely set detection parameters and switch test product models via terminal devices. In some scenarios, an IO control module can be integrated to receive the arrival trigger signal from the production line's photoelectric sensors. The memory 1005 uses a high-reliability solid-state memory (such as eMMC or industrial-grade Flash Memory) to store the operating system, equipment noise detection program, preset detection models (including an adaptive first detection model with online parameter updates and a qualitative second detection model with unchanged model parameters), user-defined parameters (such as time interval division parameters, first judgment threshold, second judgment threshold, and physical transition boundary conditions of signal physical indicators), collected historical audio feature data, and audio cache data, ensuring secure data storage and fast retrieval.

[0023] The sensor module includes a microphone array and a photoelectric sensor. The microphone array features a noise-resistant and windproof design and is connected to the processor 1001 via a signal line. It collects the raw audio signal generated by the device under test when playing test audio, laying the foundation for extracting the acoustic characteristics of the audio signal under test. The photoelectric sensor is connected to the processor 1001 via a signal line and is used to detect whether the device under test is in position, providing a synchronization signal to trigger audio acquisition.

[0024] The audio processing module integrates a digital signal processing (DSP) chip. Its core functions include high-precision analog-to-digital conversion of audio signals, frequency band processing (splitting low-frequency, mid-frequency, and high-frequency sub-band signals through filter banks), and feature extraction operations (such as calculating short-time energy fluctuation rate, high-frequency sub-band energy ratio, zero-crossing rate standard deviation, and spectral centroid shift, etc.). It can perform targeted preprocessing on audio signals according to detection requirements, and also supports signal amplitude detection and amplitude limiting to avoid acquisition distortion or equipment overload.

[0025] The actuator module includes a test fixture and a sorting component. The test fixture is used to fix the equipment to be tested (such as a loudspeaker) and provide a relatively isolated acoustic testing environment; the sorting component (such as a pneumatic push rod or a conveyor belt diversion mechanism) is connected to the processor 1001 and is used to perform physical actions based on the final output third test result to remove defective products judged to be abnormal from the main line, thereby realizing closed-loop control of testing and sorting.

[0026] The power supply module includes an industrial-grade switching power supply, an isolated power supply module, and a power protection circuit: the power supply module is adapted to the 220V AC or 24V DC power supply requirements of industrial workshops, and supports wide voltage input and low ripple output; the isolated power supply module realizes electrical isolation between the analog audio acquisition circuit and the digital control circuit; the power protection circuit can prevent abnormalities such as short circuits and overloads, and together with the whole machine's anti-static grounding design, ensures the safe and stable power supply of industrial production lines.

[0027] Those skilled in the art will understand that the structure shown in the figure does not constitute a limitation on the device and may include more or fewer components than shown, or combine certain components, or arrange different components, such as adding temperature and humidity sensors (for monitoring the temperature and humidity of the test environment for environmental noise compensation), to adapt to the full-scenario needs of industrial acoustic quality inspection.

[0028] like Figure 9 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a device noise detection program.

[0029] exist Figure 9In the device shown, the user interface module is used to interface with the device's physical buttons, status indicator lights, and other interactive components, processing user manual operation commands and providing feedback on the device status. The device noise detection program is the core functional module, integrating logic such as audio acquisition control, time-series interval division, acoustic feature extraction, cascaded inference of adaptive and deterministic models, threshold comparison, and control command generation. It drives the sensor module and audio processing module to complete the data acquisition and feature analysis of the audio signal to be detected, ultimately realizing the device noise detection function based on dual-model triggering. The processor 1001, communication bus 1002, user interface 1003, network interface 1004, memory 1005, sensor module, audio processing module, actuator module, and power supply module in the device of this invention can be set inside an industrial test cabinet or test host. The device calls the device noise detection program stored in the memory 1005 through the processor 1001 and executes the device noise detection method provided in this embodiment of the invention.

[0030] Please refer to Figure 1 , Figure 1 This is a schematic flowchart of a first embodiment of a device noise detection method according to the present invention. In one embodiment, the device noise detection method proposed by the present invention includes the following steps: S100: Acquire the audio signal to be tested from the device under test.

[0031] It should be noted that the equipment under test refers to acoustic devices on the production line that require quality control, such as speakers and receivers; the audio signal under test refers to the sound signal generated when the equipment under test plays a standard sweep frequency signal or a specific test audio at the test station. This signal not only contains effective components reflecting the acoustic quality of the equipment, but may also contain legitimate transient background noise caused by mechanical start-stop and reversal, as well as weak abnormal noise. The core purpose of obtaining this signal is to provide the raw data foundation for subsequent automated abnormal noise detection.

[0032] In this step, there are two implementation methods: In the first implementation method, after the photoelectric sensor detects that the device under test is in place, the processor 1001 directly controls the microphone array to collect the original analog audio signal in the test environment in real time at a sampling rate of 48kHz, and converts it into a digital signal buffer; In the second implementation method, the test fixture directly acquires the digital audio signal (such as I2S signal) generated by the chip inside the device under test through a wired interface, bypassing the microphone acquisition link, and directly transmits it to the processor 1001 as the audio signal to be tested, so as to eliminate the interference of environmental background noise.

[0033] S200. The acoustic features of the audio signal to be detected are input into the first detection model to obtain the first detection result; wherein, the first detection model is an adaptive model built based on machine learning and has the ability to update parameters online.

[0034] It should be noted that acoustic features are data sequences reflecting the time-frequency domain characteristics of audio signals, used to replace the original audio for model analysis. The first detection model is a sensitive model responsible for initial screening. The "adaptive model with online parameter updates" means that during production line operation, this model can dynamically fine-tune its network weight parameters based on newly input labeled data or self-supervised signals, thereby adapting to the slow data drift of the production line and maintaining sensitivity to detect subtle abnormal sounds. Inputting acoustic features into this model aims to utilize its sensitivity to minor defects to output an initial tendency for anomaly judgment.

[0035] In this step, there are two implementation methods: In the first implementation method, the processor 1001 extracts the Mel-frequency cepstral coefficient (MFCC) vector sequence of the audio signal as acoustic features through the audio processing module, and inputs it into the first detection model built on a lightweight neural network to obtain a first detection result containing anomaly scores; In the second implementation method, the processor 1001 converts the audio signal into time-frequency image features based on the Gram matrix, inputs it into the first detection model built on a convolutional neural network (CNN), and the model outputs the probability distribution of the feature map mapping as the first detection result.

[0036] S300. Determine whether the first detection result meets the preset conditions.

[0037] It should be noted that the preset conditions are manually set logical rules or numerical thresholds. Because the first detection model is overly sensitive and easily misclassifies legitimate transient noise as abnormal sound, its results cannot be directly used. The core function of this step is to set up a "filter," allowing the system to proceed to the subsequent rigorous review stage only when the first detection result shows a strong suspicion of anomaly, thereby avoiding unnecessary computational waste.

[0038] In this step, there are two implementation methods: In the first implementation method, the processor 1001 extracts a single anomaly probability value from the first detection result and compares it with a static trigger threshold. If it is greater than the threshold, it is determined that the preset condition is met. In the second implementation method, the processor 1001 combines the first detection results of the current time interval and the previous multiple consecutive time intervals to calculate the moving average trend value of the anomaly probability. If the trend value shows an upward trend and the current value exceeds the threshold, it is determined that the preset condition is met. The temporal trend is introduced to resist instantaneous pulse interference.

[0039] S400. If the first detection result meets the preset conditions, the acoustic features are input into the second detection model to obtain the second detection result; wherein, the second detection model is a pre-configured qualitative model with unchanged model parameters.

[0040] It should be noted that the second detection model serves as a review model, acting as the "final judge." The "pre-configured qualitative model with unchanged parameters" means that after offline training, all network weights, node partitioning thresholds, and other parameters are completely fixed and frozen, and absolutely no online modifications are permitted during production line operation. Because legitimate transient noise is morphologically random, while genuine abnormal sounds possess specific physical acoustic laws, the model with fixed parameters can provide an absolutely fixed judgment boundary, unaffected by current environmental fluctuations, based on the extensive abnormal sound characteristics learned during the offline phase.

[0041] In this step, there are two implementation methods: In the first implementation method, if the preset conditions are met, the processor 1001 calls the second detection model built based on the decision tree ensemble learning algorithm (such as random forest), inputs the acoustic features into it, and obtains the second detection result generated by the voting of each decision tree; In the second implementation method, the processor 1001 calls the second detection model built based on the support vector machine (SVM) algorithm, maps the acoustic features to a high-dimensional space, and obtains the second detection result by calculating the distance with the fixed hyperplane.

[0042] S500: Based on the first detection result and the second detection result, output the third detection result.

[0043] It should be noted that the third test result is the final qualitative conclusion of the equipment being tested (e.g., good or defective). Combining the first test result (providing clues to anomalies) and the second test result (providing definitive verification) can effectively filter out legitimate background noise that has been mislabeled by the first test model, while also confirming minor anomalies, thereby outputting a reliable conclusion while ensuring the detection rate.

[0044] In this step, there are two implementation methods: In the first implementation method, the processor 1001 adopts a "one-vote veto" logic, that is, it only uses the second detection result as the final basis. If the second detection result is normal, it directly outputs the normal third detection result, covering the first detection result; In the second implementation method, the processor 1001 performs weighted fusion on the probability values ​​contained in the first and second detection results (e.g., the first model weight is 0.3 and the second model weight is 0.7), compares the fused comprehensive score with the final output threshold, and generates and outputs the third detection result.

[0045] The device noise detection method provided by this invention solves the problem that existing solutions cannot simultaneously achieve high sensitivity for subtle noises and low false alarm rate for legitimate transient noise in actual production lines by employing a cascaded triggering mechanism that combines adaptive initial screening with deterministic verification. Specifically, in the detection process, the acoustic features of the acquired audio signal to be detected are first input into an adaptive first detection model with online parameter updates to obtain a first detection result. Subsequently, instead of processing all the data, the acoustic features are only input into a pre-configured qualitative second detection model with unchanged model parameters when the first detection result meets preset conditions to obtain a second detection result. Finally, a third detection result is output based on the first and second detection results. Through the above settings, the first detection model, which has the ability to update parameters online, can dynamically adapt to fluctuations in production line data and capture weak abnormal noise signals to ensure detection sensitivity. At the same time, a trigger mechanism that "meets preset conditions" is used to initially screen signals. For suspected signals that enter the review stage, a second detection model with unchanged model parameters provides a fixed judgment boundary, filtering out legitimate transient noise that is misjudged by the adaptive model. This suppresses the drift of the judgment boundary and the tendency of overfitting, thereby reducing the false alarm rate caused by transient noise while ensuring the detection rate of minor abnormal noises. This achieves an effective balance between high sensitivity and low false alarm rate.

[0046] Please refer to Figure 2 , Figure 2 This is a schematic flowchart of a second embodiment of a device noise detection method according to the present invention. In this embodiment, before step 200, the method further includes: S110. Divide the audio signal to be detected into multiple consecutive time intervals along the time axis.

[0047] It's important to note that a time-series interval refers to a short segment formed by dividing a continuous audio signal according to a predetermined time length or rule. In industrial acoustic quality inspection, the swept frequency signal is an audio signal that changes continuously over time. Legitimate transient mechanical background noise typically only occurs within a specific, brief time sequence, while abnormal sounds may also lurk at a certain local frequency point. If the entire audio segment is directly treated as a whole for feature extraction, the weak local abnormal sound features will be masked by the averaging effect of the long time sequence. Therefore, time-series slicing is a prerequisite for refined detection.

[0048] In this step, there are two implementation methods: In the first implementation method, the processor 1001 uses a fixed-length sliding window mechanism to mechanically divide the audio signal to be detected into multiple overlapping continuous time intervals according to a window length of 50ms and a step size of 10ms; In the second implementation method, the processor 1001 first calculates the global energy envelope curve of the audio signal, identifies the jump points where the energy changes abruptly, and uses these jump points as boundaries to adaptively divide the audio signal into continuous time intervals of unequal lengths, so that the interval boundaries are aligned with the transient changes of the signal.

[0049] S120. Extract the acoustic features corresponding to each time interval.

[0050] It should be noted that extracting acoustic features independently for each segmented short-time interval can preserve the transient change information of the signal at local time points, which is crucial for distinguishing between "legitimate transient noise floor" and "sudden abnormal sound".

[0051] In this step, there are two implementation methods: In the first implementation method, the audio processing module extracts Mel spectral features reflecting the static spectral distribution as acoustic features for each time interval through short-time Fourier transform (STFT); In the second implementation method, the audio processing module extracts multi-scale time-frequency features that can simultaneously take into account time resolution and frequency resolution as acoustic features for each time interval through continuous wavelet transform (CWT).

[0052] The step of inputting the acoustic features of the audio signal to be detected into the first detection model to obtain the first detection result includes: S130. Input the acoustic features corresponding to the current time interval into the first detection model to obtain the first detection result of the current time interval.

[0053] It should be noted that the current time interval refers to the specific short segment that the system is currently focusing on and performing logical judgments during streaming processing. By feeding the features of each time interval into the model one by one in a streaming manner, not only is it consistent with the real-time generation characteristics of audio data on industrial production lines, but it can also achieve precise location of the time of abnormal sound occurrence.

[0054] In this step, there are two implementation methods: In the first implementation method, the processor 1001, in the order of the time flow, only takes the acoustic features of the current time interval as isolated samples and inputs them separately into the first detection model to obtain the first detection result of the interval; In the second implementation method, the processor 1001 not only extracts the acoustic features of the current time interval, but also combines the acoustic features of the previous time interval and the next time interval to form a three-dimensional feature matrix containing contextual information, and inputs it into the first detection model to obtain the first detection result of the current time interval.

[0055] In this embodiment, by dividing the audio signal to be detected into multiple consecutive time intervals along the time axis and extracting acoustic features from each interval, and then inputting each interval into the first detection model for streaming detection, the problem of local transient information being masked by averaging during the extraction of features from the entire audio segment can be effectively avoided. The division of time intervals provides a temporal slicing basis for subsequent refined analysis, allowing legitimate transient noise to be evaluated within specific local intervals. Extracting features separately and performing streaming detection interval by interval ensures real-time matching between the detection process and the production line rhythm, enabling the first detection model to accurately capture and respond to weak abnormal fluctuations occurring at any brief time point, providing high temporal resolution clues for subsequent trigger verification.

[0056] Please refer to Figure 3 , Figure 3 This is a flowchart illustrating a third embodiment of a device noise detection method according to the present invention. In this embodiment, step S300 includes: S310. Extract the first probability value of the existence of an anomaly in the current time interval from the first detection result.

[0057] It should be noted that the first probability value is a quantitative indicator output by the first detection model, used to characterize the confidence level of the model in considering the current time interval to be an abnormal state. In order to achieve an objective and quantitative judgment standard, this probability value needs to be extracted as the direct basis for subsequent logical judgments. This value directly reflects the adaptive model's quantitative assessment of the degree of suspicion of abnormal sounds in the current interval.

[0058] In this step, there are two implementation methods: In the first implementation method, the processor 1001 directly reads the probability value corresponding to the anomaly category obtained after the output layer of the first detection model is processed by the Softmax activation function, and uses it as the first probability value; In the second implementation method, the processor 1001 extracts the high-dimensional feature vector output from the penultimate layer of the first detection model, calculates the Euclidean distance between the vector and the pre-stored "standard anomaly cluster center", and normalizes the distance through a negative exponential function, and uses the generated value as the first probability value.

[0059] S320. Compare the first probability value with the first judgment threshold.

[0060] It should be noted that the first judgment threshold is a manually set critical value standard, which is essentially a knob for adjusting the sensitivity of the initial screening. By comparing continuous probability values ​​with discrete thresholds, the model's fuzzy bias judgment can be transformed into a clear Boolean logic trigger signal.

[0061] In this step, there are two implementation methods: In the first implementation method, the processor 1001 reads a fixed first determination threshold from the memory 1005 and directly compares it with the first probability value; In the second implementation method, the processor 1001 adopts a dual threshold comparison logic with hysteresis effect, sets a higher trigger threshold and a lower cancellation threshold, and determines it as a suspected anomaly when the first probability value rises above the trigger threshold. The suspected state is only lifted when the probability value subsequently falls below the cancellation threshold, thereby preventing frequent false triggers caused by the signal fluctuating at the threshold edge.

[0062] S330. If the first probability value is greater than or equal to the first determination threshold, then the first detection result is determined to meet the preset condition.

[0063] It should be noted that using the "greater than or equal to" logic means that the initial screening stage adheres to the high-sensitivity principle of "better to kill the innocent than let the guilty go free." As long as the suspicion reaches the threshold, it must be reviewed by the subsequent qualitative model to ensure that no slight potential abnormality is missed.

[0064] In this step, there are two implementation methods: In the first implementation method, if the comparison result is that the first probability value is greater than or equal to the first judgment threshold, the processor 1001 directly generates a high-level trigger flag bit and determines that the preset condition is met; In the second implementation method, if the comparison result is that the first probability value is greater than or equal to the first judgment threshold, the processor 1001 not only generates the trigger flag bit, but also encapsulates the absolute value of the first probability value in the trigger instruction and outputs it together for subsequent review models to use as reference weights.

[0065] In this embodiment, by extracting a quantified first probability value from the first detection result and rigorously comparing it with a first judgment threshold to determine whether to trigger a review, precise control of the initial screening triggering condition is achieved. Extracting the first probability value transforms the model's complex output into an intuitive and calculable quantifiable indicator, eliminating the ambiguity of subjective judgment. The comparison with the first judgment threshold establishes a clear screening boundary, giving the system the ability to flexibly adjust the rigor of the initial screening. The strategy of triggering when the value is greater than or equal to the threshold maximizes sensitivity to minor anomalies during the initial screening stage, providing ample and necessary candidate samples for subsequent deterministic model filtering of false alarms, thus strengthening the first line of defense against missed detections while controlling computational overhead.

[0066] Please refer to Figure 4 , Figure 4 This is a flowchart illustrating a fourth embodiment of a device noise detection method according to the present invention. In this embodiment, the third detection result includes normal results and abnormal results; step S500 includes: S510. Extract the second probability value of the current time interval having an anomaly from the second detection result.

[0067] It should be noted that the second probability value is the quantified anomaly confidence level output by the second detection model for the current time series interval. Extracting the second probability value is to further quantify the degree to which the current signal passes the "rigorous test" within the framework of the deterministic model, providing data support for the final judgment.

[0068] In this step, there are two implementation methods: In the first implementation method, if the second detection model is an ensemble tree model, the processor 1001 counts the proportion of the number of trees that are voted as "abnormal" by all decision trees to the total number of trees, and uses this proportion as the second probability value; In the second implementation method, if the second detection model is a neural network model, the processor 1001 extracts the continuous values ​​of its output layer after being mapped by the Sigmoid function, and uses these values ​​as the second probability value.

[0069] S520. Determine whether the second probability value is greater than or equal to a pre-configured second determination threshold.

[0070] It should be noted that the second judgment threshold is a high standard threshold set at the deterministic model level. Since the second model itself is not easily disturbed, this threshold can be set relatively high, specifically to filter out background noise interference that only behaves abnormally in the adaptive model but does not conform to the laws of physical anomalies.

[0071] In this step, there are two implementation methods: In the first implementation method, the processor 1001 calls a globally unified fixed value as the second judgment threshold and uses the same high standard to truncate and judge all time intervals; In the second implementation method, the processor 1001 dynamically calls the segmented second judgment threshold corresponding to the current time interval in the frequency band position (such as low frequency band or high frequency band) from the memory 1005 to make a judgment, so as to adapt to the difference in the difficulty of abnormal sound manifestation under different frequency bands.

[0072] S530. If the result is greater than or equal to the second determination threshold, the abnormal result is generated; if the result is less than the second determination threshold, the normal result is generated.

[0073] It should be noted that this step essentially grants the second detection model "final confirmation authority." Even if the first model identifies an abnormal sound, if the qualitative probability given by the second model is insufficient, it is judged as a false alarm caused by legitimate background noise; only when the probability of the second model meets the standard will a definitive conclusion be reached.

[0074] In this step, there are two implementation methods: In the first implementation method, the processor 1001 generates a binary label based on the judgment result. If the label is greater than or equal to the threshold, a number "1" is generated to represent an abnormal result, and if the label is less than the threshold, a number "0" is generated to represent a normal result. In the second implementation method, the processor 1001 generates a structured text log as the third detection result. The log not only contains the conclusion of "normal" or "abnormal", but also includes the sequence number of the current time interval and the specific value of the second probability value, which is convenient for production line personnel to trace and analyze.

[0075] In this embodiment, the final output is directly determined by extracting the second probability value of the second detection model and comparing it with a high-standard second judgment threshold, thus constructing a final review mechanism based on a deterministic model. By using the second detection model with unchanged parameters to output the second probability value, the interference of production line data drift on the final review result is completely blocked; the comparison with the pre-configured second judgment threshold establishes an insurmountable objective physical judgment boundary; and the normal or abnormal result is directly generated based on the comparison result. This approach is logically clear and highly efficient, effectively filtering out legitimate transient noise and false alarms introduced by the oversensitivity of the adaptive model, ensuring the absolute reliability and low false alarm rate of the production line sorting operation.

[0076] Please refer to Figure 5 , Figure 5This is a flowchart illustrating a fifth embodiment of a device noise detection method according to the present invention. In this embodiment, during the step of inputting the acoustic features into the second detection model, the calculation of signal physical indicators for the current time interval is performed simultaneously; wherein, the signal physical indicators include at least one of short-time energy fluctuation rate, high-frequency subband energy ratio, zero-crossing rate standard deviation, and spectral centroid shift. It should be noted that signal physical indicators refer to underlying physical quantities derived directly from audio waveforms or spectra through mathematical statistics, without relying on deep learning models. These constitute a third physical defense independent of neural network features. Among them, short-time energy fluctuation rate refers to the degree of sudden change in signal amplitude within a very short time, used to capture sudden shocks; high-frequency subband energy ratio refers to the ratio of the energy of a specific high-frequency subband to the total energy of the entire frequency band, reflecting the proportion of abnormal high-frequency noise; zero-crossing rate standard deviation refers to the severity of fluctuations in the number of zero-crossing points of the signal, reflecting abnormal changes in signal zero-crossing; and spectral centroid offset refers to the offset distance of the spectral energy centroid on the frequency axis, reflecting the distortion of the spectral distribution. "Synchronous computation" refers to extracting physical indicators in parallel using the processor's idle computing power or independent hardware resources while the second detection model performs matrix operations, obtaining verification basis with zero additional time delay.

[0077] In this step, there are two implementation methods: In the first implementation method, the processor 1001 uses a heterogeneous hardware architecture to hand over the acoustic features to the AI ​​coprocessor (NPU) to run the second detection model. At the same time, it sends the original digital signal of the current time interval to an independent DSP chip through the bus. The DSP calculates the "short-time energy fluctuation rate" and "high-frequency subband energy ratio" as signal physical indicators, realizing true synchronous parallelism at the hardware level. In the second implementation method, the processor 1001 uses the multi-threading mechanism of the real-time operating system (RTOS). During the blocking period when one thread is suspended and waiting for the second detection model (such as decision tree inference) to return the result, another thread synchronously calculates the "zero-crossing rate standard deviation" and "spectral centroid offset" of the current time interval as signal physical indicators, realizing time multiplexing synchronization at the software thread level.

[0078] The second probability value is a quantitative representation of the output of the second detection model in the preceding steps. It is explicitly included in the second detection result to serve as the direct data source for the subsequent dual-judgment logic, ensuring clear data direction in the judgment process.

[0079] Step S500 includes: S511. Determine whether the second probability value is greater than or equal to a pre-configured second determination threshold, and determine whether the signal physical index reaches a preset physical transition boundary condition.

[0080] It should be noted that the second judgment threshold refers to a high standard threshold set at the deterministic model level, used to filter out background noise interference that does not conform to the physical abnormal sound pattern; the physical jump boundary condition refers to the extreme boundary set for the physical indicators of the signal. When the indicator breaks through the limit of normal fluctuation and exhibits a physical jump that conforms to the abnormal sound mechanism, it is considered to have met the condition. Here, a dual judgment logic of model probability and physical indicators is adopted.

[0081] In this step, there are two implementation methods: In the first implementation method, the processor 1001 adopts a strict "full cross-boundary" logic to determine whether all selected indicators in the signal physical indicators have simultaneously reached their respective physical transition boundary conditions. Only when all indicators cross the boundary is it considered to meet the standard. In the second implementation method, the processor 1001 adopts a "single veto" logic to determine whether any of the selected indicators in the signal physical indicators have reached its physical transition boundary conditions. As long as one indicator crosses the boundary, it is considered to meet the standard.

[0082] S521. If the second probability value is greater than or equal to the second judgment threshold, and the signal physical index reaches the physical jump boundary condition, then the abnormal result is generated; if the second probability value is less than the second judgment threshold, or the signal physical index does not reach the physical jump boundary condition, then the normal result is generated.

[0083] It should be noted that an abnormal result refers to the final conclusion label that the tested equipment has an abnormal sound defect; a normal result refers to the final conclusion label that the tested equipment only has legitimate fluctuations or background noise and is a qualified product. By combining "AND" and "OR" logic, as long as one of them does not meet the strict definition of abnormal sound (either the model is uncertain or the physical characteristics do not meet the requirements), it is judged as normal. This is an extremely conservative final review strategy.

[0084] In this step, there are two implementation methods: In the first implementation method, the processor 1001 uses pure Boolean logic gate operations to strictly generate an abnormal result label of "1" according to "model probability meets standard AND physical index meets standard", otherwise it generates a normal result label of "0"; In the second implementation method, the processor 1001 introduces a weighted scoring card mechanism, assigns a base score to the second probability value, assigns additional points to the physical index that meets the standard, calculates a comprehensive score, and generates an abnormal result containing the score value if the comprehensive score exceeds the preset final passing line, otherwise it generates a normal result, thereby further reducing the risk of misjudgment in extreme boundary cases through soft decision.

[0085] In this embodiment, a three-dimensional defense system against complex false alarms is constructed by simultaneously calculating physical indicators during the execution of the second model inference step and using dual conditions and logical output results. Simultaneously calculating signal physical indicators during the second model operation introduces an independent dimension of identification criteria without increasing the overall detection time. Pure signal physical indicators independent of model fitting are introduced, utilizing the essential difference between legitimate transient noise and abnormal sounds in the underlying physical acoustic laws as the identification criterion. The final judgment is made through a double "AND" logic where the second probability value meets the standard and the physical indicator reaches the jump boundary condition. If either dimension does not support the abnormal sound conclusion, the system is allowed to proceed, fundamentally preventing the adaptive model from misinterpreting legitimate mechanical sounds as abnormal sounds and causing uncontrolled false alarms, thus raising the false alarm rate to a stringent industrial standard.

[0086] Please refer to Figure 6 , Figure 6 This is a flowchart illustrating the process of constructing the second detection model. In this embodiment, the steps for constructing the second detection model include: S10. Obtain the offline acoustic features of the pre-collected historical abnormal sound sample set and historical normal sound sample set.

[0087] It should be noted that the second detection model, serving as a "benchmark" on the production line, must be calibrated offline using massive amounts of manually calibrated historical data. The historical abnormal sound sample set covers various typical defects, while the historical normal sample set includes normal fluctuations within the allowable range for various qualified products (including legitimate mechanical noise). This forms the basis for establishing the cognitive boundaries of the model.

[0088] In this step, there are two implementation methods: In the first implementation method, the processor 1001 extracts the historical audio of good and defective products accumulated in the past six months from the local production line database server by batch via a wired network, and extracts acoustic features offline to construct a feature matrix; In the second implementation method, the processor 1001 pulls the feature library of abnormal and normal samples from multiple different factory production lines from the application programming interface (API) of the cloud big data platform through an encryption protocol to obtain offline acoustic features with wider coverage and stronger robustness.

[0089] S20. Construct an initial ensemble model using multiple decision trees, and configure the ensemble model with a preset upper limit for the maximum depth of the trees and a lower limit for the minimum number of samples required for node splitting.

[0090] It should be noted that the preset upper limit of the maximum depth of the tree refers to the maximum number of splits that a single decision tree can go through from the root node to the leaf node. This upper limit is set to prevent the decision tree from growing overly complex branches that lack generalization ability in order to memorize a specific piece of interference noise. The lower limit of the minimum number of samples required for node splitting refers to the forced stop of splitting when the number of samples contained in the node is lower than this value. This is to prevent the model from overfitting to noise outliers containing very few samples.

[0091] In this step, there are two implementation methods: In the first implementation method, the processor 1001 adopts a more lenient pre-pruning configuration, setting the preset upper limit of the maximum tree depth to 10 layers and the lower limit of the minimum number of samples to 20, allowing the model to learn relatively complex classification boundaries; In the second implementation method, the processor 1001 adopts an extremely strong pre-pruning configuration, limiting the preset upper limit of the maximum tree depth to 5 layers and increasing the lower limit of the minimum number of samples to 50, forcing the model to learn only the coarsest-grained and most universal abnormal sound physical features.

[0092] S30. Supervised training of the ensemble model is performed using the offline acoustic features. During the growth of each decision tree, splitting is stopped when the preset upper limit is reached or the sample size of the current node is lower than the minimum sample size limit.

[0093] It should be noted that strictly enforcing the stopping condition during training is to force the model not to learn those occasional, non-universal local interference noise features (such as a specific airflow sound), but to extract the universal abnormal sound features that are stable in different batches and at different time points.

[0094] In this step, there are two implementation methods: In the first implementation method, when splitting a decision tree node, the processor 1001 uses Gini impurity as an indicator to evaluate the quality of the split and finds the feature threshold that causes the Gini coefficient of the child node to decrease the fastest for splitting; In the second implementation method, when splitting a decision tree node, the processor 1001 uses information gain ratio as an evaluation indicator and penalizes features with too many values ​​based on information gain, thereby constructing a decision tree that is more inclined to select features with clear physical meaning for splitting.

[0095] S40. Freeze the network weights of the ensemble model after training is completed, and use the frozen ensemble model as a second detection model with the model parameters remaining unchanged.

[0096] It should be noted that freezing refers to solidifying all parameters determined during the model training phase (for tree models, this means the splitting feature index and splitting threshold of each node) into a static data structure, depriving it of any self-modification rights during the runtime phase. This is a crucial step in giving the second detection model its "deterministic" essence.

[0097] In this step, there are two implementation methods: In the first implementation method, the processor 1001 serializes all node parameters of the trained ensemble model into a standard binary file, burns it into the read-only partition of the memory 1005, and instantiates the model by loading the file in the running state, thereby achieving parameter freezing at the software level; In the second implementation method, the processor 1001 directly converts the logic of the trained ensemble model into a dedicated digital logic circuit through a hardware description language, and burns it into a field-programmable gate array (FPGA), thereby achieving absolute parameter freezing and high-speed inference at the physical hardware level.

[0098] In this embodiment, an integrated model is constructed by acquiring sufficient offline calibration data, and strong pruning constraints are applied during training. Finally, the model weights are completely frozen, creating a highly robust qualitative benchmark. Supervised training based on historical normal and abnormal sound sample sets endows the model with the initial ability to distinguish between defects and legitimate fluctuations. By configuring a maximum depth upper limit and a minimum sample number lower limit and strictly implementing stop splitting during training, strong pruning is used to cut off the path of overfitting the model to local specific interference noise, forcing it to learn more generalized general abnormal sound physical features. Finally, the network weights after training are completely frozen, fundamentally preventing the possibility of the model's judgment boundary shift due to environmental drift or abnormal background noise interference during production line operation, ensuring the absolute certainty and high reliability of the second detection model as a final review mechanism.

[0099] Please refer to Figure 7 , Figure 7 This is a schematic diagram illustrating the process of obtaining the first determination threshold according to the present invention. In this embodiment, the first determination threshold is obtained through the following steps: S50. Construct a timing parameter scheduling mapping table, and independently configure the initial judgment threshold corresponding to each timing interval in the timing parameter scheduling mapping table.

[0100] It should be noted that the timing parameter scheduling mapping table is a data structure used to establish the correspondence between the time dimension (time interval) and the judgment parameters; the initial judgment threshold refers to the basic threshold value preset for each time interval without considering external interference. When playing a frequency sweep signal, the frequency corresponding to different time points is different, and the noise floor performance and the difficulty of detecting abnormal sounds in different frequency bands are drastically different (for example, the noise floor energy is large in the low frequency band, while the noise floor is small in the high frequency band but abnormal sounds are easy to appear). Therefore, the threshold for initial screening cannot be "one-size-fits-all" and must be bound to the timing (i.e., the corresponding frequency band).

[0101] In this step, there are two implementation methods: In the first implementation method, the processor 1001 divides the entire detection period into multiple levels according to a fixed time interval (such as every 100ms), and configures an initial judgment threshold with a step-decreasing step in the mapping table for each time level; In the second implementation method, the processor 1001 obtains the frequency change curve (such as a logarithmic frequency sweep curve) of the frequency sweep signal to be played, divides the detection period into low frequency band, mid frequency band, and high frequency band according to the frequency coordinate of the curve, and configures differentiated initial judgment thresholds for the time intervals corresponding to different frequency bands in the mapping table.

[0102] S60. Obtain the product batch identifier of the device under test, and / or collect the environmental reference noise data of the current test environment.

[0103] It should be noted that the product batch identifier is used to distinguish between batches of equipment acoustic characteristics due to process fine-tuning or material changes; the environmental reference noise data refers to the background noise level at the test station when no test audio is played. Introducing these two dimensions of real-time data is to give this "initial screening benchmark" basic scene adaptability.

[0104] In this step, there are two implementation methods: In the first implementation method, the processor 1001 only obtains the product batch identifier of the device under test from the host MES system through the network interface, without considering changes in environmental noise; In the second implementation method, the processor 1001 not only obtains the product batch identifier, but also controls the microphone array to perform 200ms of silent sampling before each test, and obtains the environmental reference noise data of the current test environment by calculating the average energy of the spectrum.

[0105] S70. Based on the product batch identifier and / or the environmental reference noise data, dynamically correct the initial judgment threshold in the time sequence parameter scheduling mapping table, and synchronously call the corrected threshold as the first judgment threshold according to the time flow during the detection process.

[0106] It's important to note that dynamic correction is the final step in transforming static segmented thresholds into dynamically adapted thresholds. For example, if the yield of the current batch is unstable or the environment suddenly becomes noisy, the threshold is raised accordingly to prevent false alarms. The corrected threshold is strictly synchronized with the time stream and audio data to ensure the accuracy of the initial screening conditions.

[0107] In this step, there are two implementation methods: In the first implementation method, the processor 1001 uses a simple offset superposition algorithm to query the corresponding compensation offset according to the batch identifier, and directly adds the offset to the initial judgment threshold of each timing interval in the timing parameter scheduling mapping table to achieve dynamic correction; In the second implementation method, the processor 1001 uses a linear lookup table interpolation algorithm to construct a two-dimensional mapping surface with the batch identifier and environmental reference noise data as independent variables and the threshold correction coefficient as dependent variable, calculates the accurate correction coefficient in the current state through interpolation, and then multiplies it by the initial judgment threshold to complete the dynamic correction.

[0108] In this embodiment, by constructing a time-series parameter scheduling mapping table and dynamically correcting it by integrating product batch and environmental noise data, the intelligent and scenario-based adaptation of the initial screening trigger conditions is achieved. Constructing a time-series scheduling mapping table and independently configuring the initial judgment threshold based on the characteristics of the swept frequency band breaks the limitation of a single threshold across the entire domain, enabling the initial screening sensitivity to accurately follow the time-frequency characteristics changes of the audio signal. Furthermore, product batch identification and environmental baseline noise data are acquired, introducing two key variables: production line process status and physical environmental interference. Based on these variables, the mapping table is dynamically corrected and synchronously called according to the time flow, making the first judgment threshold a "smart vernier caliper" capable of sensing frequency band changes, adapting to batch differences, and resisting environmental fluctuations. Without increasing the model's computational burden, this significantly reduces the number of invalid triggers caused by external non-noise factors affecting the first detection model, optimizing the overall operating efficiency of the cascaded detection system.

[0109] Furthermore, this invention also proposes a device for detecting abnormal equipment noise, the device comprising: The audio acquisition module 10 is used to acquire the audio signal to be detected from the device under test; The processing module 20 is used to divide the audio signal to be detected into multiple time intervals and extract the acoustic features corresponding to each time interval; The first detection module 30 is used to receive the acoustic features corresponding to each of the time intervals, output the first detection result, and has the ability to update parameters online. The second detection module 40 is used to receive the acoustic features of the corresponding time interval and output the second detection result when the first detection result meets the preset conditions, and the parameters of the second detection module remain fixed. The output module 50 is used to output a third detection result based on the first detection result and the second detection result.

[0110] It should be noted that the audio acquisition module 10 is connected to the sensor module (microphone array). Its core function is to control the acquisition timing, receive and buffer the original audio signal of the device under test, and ensure the integrity of the input data source and the consistency of the timing.

[0111] The processing module 20 is integrated into the processor 1001 and the audio processing module. Its core function is to realize the temporal decomposition and digital representation of the signal, and to convert the long continuous audio slices into a standardized acoustic feature stream that can be processed by neural networks.

[0112] The first detection module 30 acts as the system's sensitive antenna, using its online update capability to maintain the tracking of production line data drift and efficiently screen out minute clues of suspected anomalies in a massive time series.

[0113] The second detection module 40, as the qualitative center of the system, avoids the consumption of full computing power through a passive triggering mechanism and uses its frozen parameter boundaries to conduct a rigorous physical law-level review of suspected clues.

[0114] Output module 50 is responsible for coordinating the decision-making logic, integrating the diverse output results of the preceding modules, generating the final good / bad judgment instruction, and driving the actuator to take action.

[0115] This noise detection device, through modular collaborative design, constructs a complete pipeline system of "time-sequence slicing - adaptive initial screening - trigger interception - deterministic final review - physical sorting". It breaks through the physical bottleneck of "mutual exclusion of sensitivity and false alarm rate" in traditional single-model detection. Without increasing the system's full-load computing power burden, it achieves high-sensitivity capture of minute noises from industrial loudspeakers and other equipment and high-reliability filtering of legitimate transient background noise, significantly improving the accuracy and cycle time efficiency of automated quality inspection on the production line.

[0116] Since this device adopts all the technical solutions of all the above embodiments, it has at least all the beneficial effects brought about by the technical solutions of the above embodiments, including solving the problem that the existing solutions cannot simultaneously achieve high sensitivity to minute abnormal sounds and low false alarm rate of legal background noise, which will not be elaborated here.

[0117] Furthermore, embodiments of the present invention also propose a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the device noise detection method described above.

[0118] It should be noted that the readable storage medium can be any one or more of the following computer-readable storage media: flash memory, embedded multimedia cards, secure digital cards, etc., all of which have non-volatility and high read / write speed characteristics, ensuring the secure and persistent storage of computer programs, weight parameters of the first and second detection models, timing parameter scheduling mapping tables, and operation log data, supporting stable read and write in complex industrial electromagnetic environments, and meeting the service life requirements of long-term high-frequency operation of industrial testing equipment.

[0119] The computer program is stored in a readable storage medium in binary machine code format. It contains all the logical steps from audio timing division, dual-model cascaded triggering inference to physical index fusion judgment in the above method embodiments. When the readable storage medium is connected to the device's processor, the processor reads the computer program in the storage medium through the storage controller, loads the program into the running memory and executes it sequentially, thereby realizing the fully automated closed-loop function of device abnormal sound detection.

[0120] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0121] In addition, for technical details not described in detail in this embodiment, please refer to the device noise detection method provided in any embodiment of the present invention, which will not be repeated here.

[0122] Furthermore, it should be noted that, in this document, 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. Unless otherwise specified, 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 that element.

[0123] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0124] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary hardware platforms, and of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium (such as flash memory, eMMC, SD card) and includes several instructions to cause a device (which may be an industrial test instrument, acoustic analysis host, etc.) to execute the methods described in the various embodiments of the present invention.

[0125] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for detecting abnormal noise in equipment, characterized in that, include: Acquire the audio signal to be tested from the device under test; The acoustic features of the audio signal to be detected are input into the first detection model to obtain the first detection result; wherein, the first detection model is an adaptive model built based on machine learning and has the ability to update parameters online; Determine whether the first detection result meets the preset conditions; If the first detection result meets the preset conditions, the acoustic features are input into the second detection model to obtain the second detection result; wherein, the second detection model is a pre-configured qualitative model with unchanged model parameters; Based on the first detection result and the second detection result, a third detection result is output.

2. The method for detecting abnormal noise in equipment according to claim 1, characterized in that, Before the step of inputting the acoustic features of the audio signal to be detected into the first detection model, the method further includes: The audio signal to be detected is divided into multiple consecutive time intervals along the time axis; Extract the acoustic features corresponding to each time interval; The step of inputting the acoustic features of the audio signal to be detected into the first detection model to obtain the first detection result includes: The acoustic features corresponding to the current time interval are input into the first detection model to obtain the first detection result of the current time interval.

3. The equipment noise detection method as described in claim 2, characterized in that, The step of determining whether the first detection result meets the preset conditions includes: Extract a first probability value from the first detection result indicating that there is an anomaly in the current time interval; The first probability value is compared numerically with the first judgment threshold; If the first probability value is greater than or equal to the first determination threshold, then the first detection result is determined to meet the preset condition.

4. The equipment noise detection method as described in claim 3, characterized in that, The third detection result includes normal results and abnormal results; the step of outputting the third detection result based on the first detection result and the second detection result includes: Extract a second probability value indicating an anomaly in the current time interval from the second detection result; Determine whether the second probability value is greater than or equal to a pre-configured second determination threshold; If the result is greater than or equal to the second determination threshold, the abnormal result is generated; if the result is less than the second determination threshold, the normal result is generated.

5. The equipment noise detection method as described in claim 3, characterized in that, When performing the step of inputting the acoustic features into the second detection model, the signal physical indicators of the current time interval are calculated simultaneously; wherein, the signal physical indicators include at least one of short-time energy fluctuation rate, high-frequency subband energy ratio, zero-crossing rate standard deviation, and spectral centroid shift; the second detection result includes a second probability value of an anomaly existing in the current time interval; the step of outputting a third detection result based on the first detection result and the second detection result includes: Determine whether the second probability value is greater than or equal to a pre-configured second determination threshold, and determine whether the signal physical index has reached a preset physical transition boundary condition; If the second probability value is greater than or equal to the second judgment threshold, and the signal physical index reaches the physical transition boundary condition, then the abnormal result is generated; if the second probability value is less than the second judgment threshold, or the signal physical index does not reach the physical transition boundary condition, then the normal result is generated.

6. The equipment noise detection method as described in claim 3, characterized in that, The first determination threshold is obtained through the following steps: Construct a timing parameter scheduling mapping table, and independently configure the initial judgment threshold corresponding to each timing interval in the timing parameter scheduling mapping table; Obtain the product batch identifier of the device under test, and / or collect the environmental reference noise data of the current test environment; Based on the product batch identifier and / or the environmental reference noise data, the initial judgment threshold in the time-series parameter scheduling mapping table is dynamically corrected, and the corrected threshold is synchronously called as the first judgment threshold according to the time flow during the detection process.

7. The method for detecting abnormal noise in equipment as described in any one of claims 1 to 5, characterized in that, The steps for constructing the second detection model include: Obtain the offline acoustic features of pre-collected historical abnormal sound sample sets and historical normal sound sample sets; An initial ensemble model is constructed using multiple decision trees, and a preset upper limit for the maximum depth of the trees and a lower limit for the minimum number of samples required for node splitting are configured for the ensemble model. The ensemble model is trained under supervision using the offline acoustic features. During the growth of each decision tree, splitting stops when the preset upper limit is reached or the number of samples in the current node is lower than the minimum number of samples. Freeze the network weights of the ensemble model after training is completed, and use the frozen ensemble model as a second detection model with the model parameters remaining unchanged.

8. A device for detecting abnormal noise in equipment, characterized in that, The abnormal noise detection device includes: The audio acquisition module is used to acquire the audio signal to be tested from the device under test. The processing module is used to divide the audio signal to be detected into multiple time intervals and extract the acoustic features corresponding to each time interval; The first detection module is used to receive the acoustic features corresponding to each of the time intervals, output the first detection result, and has the ability to update parameters online. The second detection module is used to receive the acoustic features of the corresponding time interval and output the second detection result when the first detection result meets the preset conditions, and the parameters of the second detection module remain fixed. The output module is used to output a third detection result based on the first detection result and the second detection result.

9. A device for detecting abnormal noise in equipment, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the device noise detection method as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the device noise detection method as described in any one of claims 1 to 7.

Images