Compressed garbage truck fault judgment method and system based on voiceprint recognition

Abstract

The invention relates to the technical field of equipment acoustic state monitoring and fault diagnosis, in particular to a compressed garbage truck fault judgment method and system based on voiceprint recognition. Comprising the steps of synchronously obtaining an audio signal, real-time working condition data and an action instruction of a vehicle, and performing adaptive noise reduction on the audio signal by using the working condition data; then double-path diagnosis is performed on the noise-reduced signals in parallel, systematic relevance faults in one path are diagnosed through causal time sequence analysis, and early transient faults of components in the other path are diagnosed through auto-encoder reconstruction error analysis; and finally, giving an alarm according to the diagnosis result of any path. According to the method, the working condition self-adaptive noise reduction and the double-path parallel diagnosis model are fused, so that the problems of low diagnosis accuracy and single fault identification type under strong interference and variable working conditions in the prior art are solved, and comprehensive and accurate diagnosis of systematic associated faults and early transient faults is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment acoustic state monitoring and fault diagnosis, in particular to a compression garbage truck fault judgment method and system based on voiceprint recognition. BACKGROUND

[0002] Acoustic analysis of a complex sounder such as a compression garbage truck is the key to judging the health state of the equipment, which is similar to diagnosing the pathological characteristics of the human body by analyzing the speech signal.

[0003] The prior art attempts to directly migrate the voiceprint recognition technology for humans to the industrial diagnosis field. The method is usually to extract the Mel frequency cepstral coefficients (MFCCs) or spectrogram of machine sound, and then combine support vector machines (SVM) or conventional neural networks for classification. However, the original intention of these methods is to process relatively stable human speech, and they ignore the core problem that machine "speech" will be strongly modulated by operating conditions (such as speed, load) and environmental noise. When the operating conditions change, the voiceprint features will drift, resulting in poor model generalization ability, and it is difficult to establish a stable fault pattern mapping. In real complex scenarios, the diagnosis accuracy is low, and false negatives and false positives occur frequently

[0004] Therefore, there is an urgent need for a new acoustic analysis method that can decouple operating conditions and faults, and can deeply analyze different types of abnormal information in machine speech, to achieve accurate diagnosis of machine complex problems.

[0005] To this end, a compression garbage truck fault judgment method and system based on voiceprint recognition are proposed. SUMMARY

[0006] The purpose of the present application is to provide a compression garbage truck fault judgment method and system based on voiceprint recognition, which deeply analyzes and identifies the systematic correlation faults and early transient faults contained in the machine voiceprint by fusing operating condition adaptive noise reduction and a dual-path parallel diagnosis model, aiming to solve the problems of low diagnosis accuracy and incomplete fault type recognition in complex operating conditions.

[0007] To achieve the above purpose, the present application provides the following technical scheme:

[0008] The compression garbage truck fault judgment method and system based on voiceprint recognition comprises:

[0009] Real-time acquisition of multi-channel audio signals; and synchronous acquisition of real-time operating condition data and action instructions of the vehicle; according to the real-time operating condition data, dynamically selecting a model matched with the current operating condition from a background noise model library, and performing adaptive noise reduction processing on the multi-channel audio signals to obtain clean audio signals;

[0010] Parallel execution of a dual-diagnosis path on the clean audio signals:

[0011] establishing a real-time time-series event chain composed of a plurality of sound events from the clean audio signal with the action instruction as a causal event starting point; analyzing whether there is a deviation in timing, sequence and / or feature by comparing the real-time time-series event chain with the causal acoustic template, diagnosing whether there is a correlation fault;

[0012] inputting the clean audio signal into a self-encoder model pre-trained based on health state data; detecting whether there is a pulse peak caused by early damage of a component by analyzing a reconstruction error signal sequence between the input and output of the self-encoder model, diagnosing whether there is a transient fault;

[0013] when there is a correlation fault and / or a transient fault, issuing a fault alarm.

[0014] Preferably, the real-time working condition data specifically includes engine speed, power take-off engagement state, hydraulic system main pump pressure, and action instruction of the hydraulic valve group.

[0015] Preferably, the background noise model library is a combination of health state sounds of the compression garbage truck under a plurality of micro working condition combinations, wherein each micro working condition combination corresponds to one background noise model, forming a mapping relationship; the dynamically selecting a model matched with the current working condition from the background noise model library according to the real-time working condition data is to combine the obtained plurality of real-time working condition data, and use the combination as an index to directly call the background noise model uniquely corresponding to the index in the background noise model library, so as to perform high-target noise elimination for the accurate running state of the current vehicle.

[0016] Preferably, the causal acoustic template is a pre-established database, and for each action instruction as a causal event, a series of sound events triggered under the health state are stored, including reference occurrence sequence, reference time delay, and reference acoustic feature;

[0017] The deviation includes timing deviation comparison, sequence deviation comparison, and feature deviation comparison;

[0018] The timing deviation comparison diagnoses an action slow and / or fast fault caused by hydraulic leakage and / or pipeline blockage by judging whether the actual occurrence time of the sound event in the real-time time-series event chain exceeds the preset tolerance range of the reference time delay;

[0019] The sequence deviation comparison diagnoses a logic error fault caused by electromagnetic valve sticking and / or mechanical component interference by judging whether the actual occurrence sequence of the sound event in the real-time time-series event chain is consistent with the reference occurrence sequence and / or whether there is a missing key sound event;

[0020] Feature deviation comparison determines whether the actual acoustic characteristics of a sound event in a real-time time-series event chain, including frequency, amplitude, and energy envelope, differ significantly from the reference acoustic characteristics, thus diagnosing faults caused by changes in operating status due to abnormal load and / or component wear.

[0021] Preferably, the autoencoder model is trained by inputting health status audio data of the compressed garbage truck under various operating conditions;

[0022] The reconstructed error signal sequence includes point-by-point subtraction of the input and output of the autoencoder model to obtain a time-synchronized error value sequence, and applying a peak detection algorithm to the error value sequence to find isolated energy pulses whose amplitude exceeds a preset statistical threshold; the pulse peak is used to characterize the occurrence of transient acoustic events that are discontinuous in the time domain caused by early bearing cracks, gear surface pitting, and loosening of structural connections.

[0023] Preferably, the fault alarm includes a first type of alarm, a second type of alarm, and a third type of alarm; the first type of alarm indicates the existence of a related fault with performance degradation and / or logical errors, and indicates the specific stage at which the related fault occurs; the second type of alarm indicates the existence of a transient fault with component-level physical damage, and indicates the frequency of abnormal pulse peaks; when both a related fault and a transient fault exist simultaneously, a third type of alarm is generated; the third type of alarm indicates the existence of a complex serious fault caused by physical damage to components that has affected the overall operation of the equipment.

[0024] A fault diagnosis system for compressed garbage trucks based on voiceprint recognition includes:

[0025] Multi-source acquisition module: Real-time acquisition of multi-channel audio signals; and synchronous acquisition of real-time vehicle operating data and action commands;

[0026] Adaptive noise reduction module: Based on the real-time operating data, dynamically selects a model matching the current operating conditions from the background noise model library, and performs adaptive noise reduction processing on the multi-channel audio signal to obtain a clean audio signal;

[0027] A dual diagnostic path is executed in parallel for the clean audio signal:

[0028] Causal analysis module: Taking the action command as the starting point of the causal event, it establishes a real-time time-series event chain composed of multiple sound events from the clean audio signal; by comparing the real-time time-series event chain with the causal acoustic template, it analyzes whether there are deviations in timing, sequence and / or characteristics, and diagnoses whether there are related faults.

[0029] The pulse detection module inputs the clean audio signal into a self-encoder model pre-trained based on health state data; by analyzing the reconstruction error signal sequence between the input and output of the self-encoder model, whether a pulse peak caused by early damage of a component exists is detected, and whether a transient fault exists is diagnosed;

[0030] The fusion decision module issues a fault alarm when there is a correlation fault and / or a transient fault.

[0031] Compared with the prior art, the present application has the following beneficial effects:

[0032] 1. The present application fundamentally solves the core pain point that the prior art cannot effectively separate faults and normal noises when facing application scenarios such as compression garbage trucks with variable working conditions and strong background noise, which leads to inaccurate diagnosis. The present application deeply binds the real-time running condition data of the vehicle and the noise reduction process, and realizes high-targeted elimination of normal working noise by dynamically calling a noise model accurately matched with the current micro working condition. This method ensures the quality of the audio signal input into the diagnosis module from the source of signal processing, so that weak fault features can be highlighted instead of being overwhelmed by strong background noise. The direct effect is to significantly reduce the false positive rate caused by working condition changes and the missed alarm rate caused by too low signal-to-noise ratio, so that the stability and reliability of the entire diagnosis system in real complex environments are qualitatively improved.

[0033] 2. The present application improves the scope of fault diagnosis from recognizing a single abnormal sound to understanding the action flow of the machine to a new dimension through correlation fault diagnosis based on causal time sequence voiceprints. The prior art cannot diagnose soft faults that do not have obvious abnormal sounds but affect the running efficiency and logic of the equipment. The present application can judge the fault by analyzing whether the rhythm and logic of the sound event chain are normal by establishing the causal relationship template of the action instruction and the subsequent series of sound events. This enables the present application to effectively diagnose correlation faults such as performance degradation and function failure at the system level, greatly expanding the depth and breadth of acoustic diagnosis.

[0034] 3. The present application solves the industry problem that early physical damage such as bearing crack and gear pitting is difficult to be found because the sound signal energy is weak and the duration is extremely short through transient fault enhancement detection based on reconstruction error sequence. This method ingeniously uses the poor reconstruction ability of self-encoder for abnormal signals to shift the focus from the original signal to the reconstruction error sequence. A weak transient impact may not be noticeable in the original signal, but it will be amplified as a clear pulse peak in the error sequence. This method is equivalent to providing a powerful amplifier for transient fault signals, which has high detection sensitivity. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A method flowchart of the method for judging the fault of the compression garbage truck based on the voiceprint recognition proposed for the embodiments of the present application is shown in the figure.

[0036] Figure 2 A construction flowchart of the causal acoustic template proposed for the embodiments of the present application is shown in the figure.

[0037] Figure 3 A system structure diagram of the system for judging the fault of the compression garbage truck based on the voiceprint recognition proposed for the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0039] Embodiment one

[0040] Please refer to Figures 1 to 2 The present application provides a method and system for judging the fault of a compression garbage truck based on voiceprint recognition, and the technical solutions are as follows:

[0041] The method for judging the fault of the compression garbage truck based on the voiceprint recognition is shown in the figure, which comprises: Figure 1

[0042] Real-time multi-channel audio signals are collected, and real-time working condition data and action instructions of the vehicle are synchronously acquired. According to the real-time working condition data, a model matched with the current working condition is dynamically selected from a background noise model library, and the multi-channel audio signals are subjected to adaptive noise reduction processing to obtain clean audio signals.

[0043] The clean audio signals are subjected to parallel execution of double diagnosis paths:

[0044] A real-time time sequence event chain composed of multiple sound events is established from the clean audio signals with the action instructions as the starting point of the causal event. By comparing the real-time time sequence event chain with a causal acoustic template, whether there is a deviation in time sequence, order and / or characteristics is analyzed to diagnose whether there is a correlation fault.

[0045] The clean audio signals are input into a self-encoder model pre-trained based on health state data. By analyzing the reconstruction error signal sequence between the input and output of the self-encoder model, whether there is a pulse peak caused by early damage of a component is detected to diagnose whether there is a transient fault.

[0046] ​When there is a correlation failure and / or transient failure, a failure alarm is issued.

[0047] Further, the real-time working condition data specifically includes engine speed, power take-off engagement state, hydraulic system main pump pressure, and hydraulic valve group action instruction.

[0048] The acquisition of the real-time working condition data is achieved by reading the data frames broadcast by the vehicle controller local area network in real time through a vehicle-mounted bus interface, and analyzing these data frames to extract information such as engine speed, power take-off engagement state, hydraulic system main pump pressure, and hydraulic valve group action instruction. This method does not require invasive modification of the vehicle, ensuring the synchronization and integrity of data acquisition.

[0049] In order to convert the real-time working condition data into discrete working condition codes that can be indexed by the model library, the following processing is performed:

[0050] The operating speed range of the engine is divided into multiple preset intervals, for example: idle speed (<800 RPM), low speed (800-1200 RPM), medium speed (1201-1800 RPM), high speed (>1800 RPM), etc. The real-time speed value will be mapped to the interval code to which it belongs.

[0051] The main pump pressure range is divided into multiple levels according to the load characteristics, for example: idle pressure (<5 MPa), light load pressure (5-12 MPa), medium load / rated pressure (12-20 MPa), overload pressure (>20 MPa). The real-time pressure value will be mapped to the pressure level code to which it belongs.

[0052] The power take-off engagement state is binary data, directly defined as "engaged" or "disengaged" in two states.

[0053] The definition of the hydraulic valve group action instruction is to define the combination of the power-on / power-off state of the core electromagnetic valve controlling different actuators (such as the blade, slide, turnover mechanism, lifting mechanism, etc.) as a specific action instruction. For example, the 'blade compression' instruction corresponds to the power-on of valve A and valve B, and the 'filler lifting' instruction corresponds to the power-on of valve C. Each independent valve group action instruction with specific function is regarded as a discrete state.

[0054] This embodiment acquires multi-dimensional real-time data such as engine, hydraulic, etc. non-invasively, and combines and discretizes them into precise micro-working condition codes. This method provides reliable and accurate indexing basis for subsequent analysis (such as noise reduction), so that the system can perform high-targeted processing according to the current precise operating state of the vehicle, greatly improving the fine analysis ability and adaptability of the system to complex and variable working conditions.

[0055] Further, the background noise model library is a combination of health state sounds of the compression garbage truck under a plurality of micro working condition combinations, each of which corresponds to a background noise model, forming a mapping relationship; the dynamic selection of a model matching the current working condition from the background noise model library according to the real-time working condition data is to combine the obtained plurality of real-time working condition data, and use the combination as an index to directly call the background noise model uniquely corresponding to the index in the background noise model library, so as to accurately eliminate noise for the precise running state of the current vehicle.

[0056] The specific method of using the combination as an index is to encode each working condition data after discretization processing, splice according to a preset order, and form a unique composite working condition string or multi-dimensional key value.

[0057] For example, a specific micro working condition index can be defined as:

[0058] ‘engine speed interval code-torque converter state code-main pump pressure level code-hydraulic valve group action instruction code’

[0059] Suppose at a certain moment, the vehicle state is: engine speed 1000 RPM (belongs to the ‘low speed’ interval, coded as 02), torque converter ‘engaged’ (coded as 1), main pump pressure 15 MPa (belongs to the ‘medium load’ level, coded as 03), and ‘squeegee compression’ action is being performed (coded as V_COMPRESS). The working condition index at this moment is constructed as ‘02-1-03-V_COMPRESS’. The system directly calls the background noise model that accurately matches it from the background noise model library using this unique index for subsequent adaptive noise reduction processing.

[0060] The construction of the background noise model library specifically includes: for each micro working condition combination, collecting a health state audio signal of more than 60 seconds, and calculating the average power spectral density; using the average power spectral density as the background noise model under this real-time working condition, and storing it together with the corresponding working condition index;

[0061] The adaptive noise elimination process specifically uses a spectral subtraction algorithm: the real-time collected audio signal is subjected to short-time Fourier transform to obtain the power spectrum of the current frame, the average power spectral density of the corresponding background noise model called according to the current working condition index is subtracted from the power spectrum, and inverse Fourier transform is performed to obtain the noise-reduced audio signal. This method can accurately eliminate background noise under a specific working condition, greatly improve the signal-to-noise ratio, lay a solid foundation for subsequent accurate identification of weak fault features, and significantly reduce the false positive and false negative rates.

[0062] The embodiment realizes dynamic self-adaptation of the noise reduction process by establishing a one-to-one mapping relationship between micro working conditions and health state noise models and using real-time working condition data to form a unique index. The method can accurately separate and eliminate normal working noise under specific working conditions from complex background noise, rather than using a "one-size-fits-all" broad filter. This greatly improves the signal-to-noise ratio, making the weak fault characteristic signals previously submerged in strong noise stand out, fundamentally improving the accuracy and reliability of subsequent fault diagnosis, and effectively avoiding false positives and false negatives caused by changes in working conditions.

[0063] Further, the causal acoustic template is a pre-established database, which stores the reference occurrence sequence, reference time delay and reference acoustic feature of a series of sound events triggered under the health state for each action instruction as a causal event;

[0064] The deviation includes time sequence deviation comparison, sequence deviation comparison and feature deviation comparison;

[0065] The time sequence deviation comparison diagnoses the action delay and / or action acceleration fault caused by hydraulic leakage and / or pipeline blockage by judging whether the actual occurrence time of the sound event in the real-time time sequence event chain exceeds the preset tolerance range of the reference time delay;

[0066] The sequence deviation comparison diagnoses the logic error fault caused by electromagnetic valve sticking and / or mechanical component interference by judging whether the actual occurrence sequence of the sound event in the real-time time sequence event chain is consistent with the reference occurrence sequence and / or whether there is a missing key sound event;

[0067] The feature deviation comparison diagnoses the running state change fault caused by abnormal load and / or component wear by judging whether the actual acoustic feature of a certain sound event in the real-time time sequence event chain, including frequency, amplitude and energy envelope, has a significant statistical difference from the reference acoustic feature.

[0068] As shown in Figure 2 The establishment process of the causal acoustic template is as follows: 50 or more audio samples under the health state covering all action instructions are provided. By combining the time stamp of the vehicle-mounted bus interface instruction and the audio energy envelope analysis, the time point of the key sound event in each sample is labeled in a supervised or semi-supervised manner; the time delay and acoustic feature vector of the same event in all samples are statistically analyzed to calculate the mean and standard deviation / covariance matrix. These statistics together constitute the reference stored in the database. This method constructs the reference by statistically modeling a large number of healthy samples, rather than relying on a single template, which makes the diagnosis tolerant to normal fluctuations and greatly improves the robustness and anti-interference ability of the model.

[0069] The reference acoustic features are stored in the database in the form of a multi-dimensional feature vector, which includes at least: peak frequency, spectral centroid, root mean square amplitude, and rise time and decay time of the energy envelope.

[0070] The timing deviation comparison determines whether the difference between the actual occurrence time of a sound event in the real-time timing event chain and the reference time delay exceeds a preset tolerance threshold. This tolerance threshold can be set to ±15% of the reference time delay, or an absolute time window of not less than 0.1 seconds. When this threshold is exceeded, a timing deviation is determined, thereby diagnosing faults caused by hydraulic leakage and / or pipeline blockage leading to excessively slow or / or excessively fast movements.

[0071] The feature deviation comparison extracts the actual acoustic feature vectors of real-time sound events and calculates the Mahalanobis distance between them and the corresponding benchmark acoustic feature vectors in the database. This distance is then compared to a preset deviation judgment threshold. When the distance is greater than the threshold, a statistically significant difference is determined. The deviation judgment threshold is pre-calibrated by performing statistical analysis on multiple samples under healthy conditions. Based on this, faults caused by abnormal load and / or component wear are diagnosed.

[0072] The deviation judgment threshold in the feature deviation comparison is determined based on the Mahalanobis distance distribution of the aforementioned 50 healthy samples. For example, the Mahalanobis distance of all healthy samples can be calculated, and the upper bound of its 99% confidence interval can be selected as the final judgment threshold. The threshold calibrated by this method can effectively distinguish between normal fluctuations and abnormal deviations.

[0073] This embodiment elevates diagnosis from identifying isolated abnormal noises to understanding the complete logical chain of machine actions by establishing a causal acoustic template. Through multi-dimensional quantitative comparison of timing, sequence, and acoustic characteristics, it can accurately diagnose "soft faults" that are difficult to detect using traditional methods, such as slow movement caused by hydraulic leakage or logical errors caused by solenoid valve sticking. This improves the depth of fault diagnosis and the accuracy of root cause localization, providing precise decision-making basis for predictive maintenance.

[0074] Furthermore, the autoencoder model is trained by inputting health status audio data of the compressed garbage truck under various operating conditions;

[0075] The reconstructed error signal sequence includes point-by-point subtraction of the input and output of the autoencoder model to obtain a time-synchronized error value sequence, and applying a peak detection algorithm to the error value sequence to find isolated energy pulses whose amplitude exceeds a preset statistical threshold; the pulse peak is used to characterize the occurrence of transient acoustic events that are discontinuous in the time domain caused by early bearing cracks, gear surface pitting, and loosening of structural connections.

[0076] The autoencoder model is a one-dimensional convolutional autoencoder, comprising three one-dimensional convolutional layers with 16, 32, and 64 filters respectively, all with a kernel size of 3 and a stride of 2, and all using ReLU as the activation function. The decoder uses a deconvolutional layer structure symmetrical to the encoder. The loss function used during model training is mean squared error.

[0077] The process of finding isolated energy pulses with amplitudes exceeding a preset statistical threshold specifically includes the following steps:

[0078] Step 1: Input an independent, untrained set of health status verification audio data into the trained autoencoder model to obtain the corresponding reconstruction error sequence. Calculate the mean and standard deviation of the reconstruction error sequence. The preset statistical threshold is determined based on the statistical results. The specific calculation method is as follows: multiply the calculated standard deviation by a coefficient preset according to the diagnostic sensitivity requirements, and then add the product to the calculated mean. The sum is the final threshold. The coefficient is typically between 3 and 6.

[0079] Step 2: Apply a peak detection algorithm to the real-time reconstruction error sequence. This algorithm can be implemented using a signal processing library. Set the peak height parameter to be greater than the threshold T calibrated in Step 1, and set the minimum horizontal distance parameter between peaks to ensure that the detected pulses are isolated energy pulses.

[0080] This method adaptively determines the detection threshold using statistical methods, avoiding the blindness of manually setting thresholds, and can flexibly balance the sensitivity and reliability of diagnosis by adjusting coefficients, resulting in more accurate results.

[0081] This embodiment cleverly shifts the focus of analysis from the original signal to the reconstruction error by training an autoencoder that only learns from healthy states. This method leverages the model's poor ability to reconstruct abnormal transient signals, amplifying the weak impact sounds emitted by early damage such as bearing cracks and gear pitting into a clearly discernible pulse peak in the error sequence. This is equivalent to providing a powerful amplifier for weak fault signals, achieving extremely high detection sensitivity and solving the industry's problem of difficulty in detecting early physical damage due to weak signals.

[0082] Furthermore, the fault alarm includes a first type of alarm, a second type of alarm, and a third type of alarm; the first type of alarm indicates the existence of a related fault with performance degradation and / or logical errors, and indicates the specific stage at which the related fault occurs; the second type of alarm indicates the existence of a transient fault with component-level physical damage, and indicates the frequency of abnormal pulse peaks; when both a related fault and a transient fault exist simultaneously, a third type of alarm is generated; the third type of alarm indicates the existence of a complex serious fault caused by physical damage to components that has affected the overall operation of the equipment.

[0083] When the causal analysis module diagnoses a correlational fault, it generates a first-type alarm. This alarm indicates a correlational fault involving performance degradation and / or logical errors. When the diagnosed deviation (timing, sequence, or characteristic deviation) exceeds a preset "severe" threshold, the alarm will indicate the specific stage where the correlational fault occurred.

[0084] When the pulse detection module diagnoses a transient fault, it generates a second type of alarm. The alarm indicates the presence of a component-level, physical damage transient fault. It also indicates the frequency of abnormal pulse peaks, calculated as the total number of pulse peaks detected within a single action command execution cycle.

[0085] The third type of alarm is generated when both the causal analysis module and the pulse detection module report a fault within the execution time window of the same action command. In this case, the system determines that the two faults are causally related and generates the third type of alarm. The content of the third type of alarm indicates a complex and serious fault caused by physical damage to components that has affected the overall operation of the equipment.

[0086] This embodiment establishes a tiered alarm mechanism that clearly distinguishes between related faults indicating performance degradation (Type 1 alarms) and transient faults indicating physical damage to components (Type 2 alarms). A third type of alarm, triggered when both types of faults occur concurrently within the same action window, intelligently reveals that the physical damage has evolved into a complex and severe fault affecting the entire process. This progressively sophisticated alarm logic provides users with a clear classification of fault severity, enabling maintenance decisions to shift from reactive response to precise predictive maintenance, making it highly valuable in practice.

[0087] This invention fundamentally solves the core pain point of current applications in this field, which suffer from inaccurate diagnosis due to the inability to effectively separate faults from normal noise in application scenarios with variable operating conditions and strong background noise. By deeply binding real-time vehicle operating data with the noise reduction process, this invention achieves highly targeted elimination of normal operating noise, ensuring the quality of the audio signal input to the diagnostic module. Its innovative dual-diagnostic path parallel execution can diagnose correlated "soft faults" affecting equipment operating efficiency and logic through causal temporal acoustic pattern analysis; and it can also enhance and detect weak transient faults caused by early physical damage with high sensitivity by analyzing and reconstructing error sequences. This method greatly expands the depth and breadth of acoustic diagnosis, significantly improving the accuracy and reliability of the diagnostic system.

[0088] Example 2

[0089] This embodiment aims to specifically illustrate the application process of the present invention in detecting a complex fault caused by early physical damage that has begun to affect the overall operation of the equipment. A detailed structural diagram is shown below. Figure 3 As shown.

[0090] A garbage compactor truck is performing a scraper compaction cycle. The vehicle has a potential complex fault:

[0091] Transient fault source: Early pitting has occurred on the surface of a transmission gear that drives the scraper. This will produce a weak, short-lived impact sound at a specific angle of gear meshing.

[0092] Related fault manifestation: The additional vibration and load caused by the aforementioned gear pitting accelerated the loosening of a joint in the connected hydraulic line, resulting in a slight hydraulic oil leak. This made the latter half of the scraper compression (slide forward push) movement slightly slower than normal.

[0093] In traditional detection methods, the transient sound of gear pitting is easily drowned out by the strong background noise of the engine and hydraulic pump and cannot be detected; and the sluggish movement is also difficult for the operator to detect without precise timing. Therefore, this combination of faults is extremely insidious.

[0094] When the operator issues the "scraper compression" action command, the system of this invention begins to work:

[0095] Multi-source acquisition module:

[0096] A multi-channel microphone array installed in key parts of the vehicle (such as near the hydraulic pump and scraper drive mechanism) begins to collect audio signals in real time.

[0097] Preferably, for monitoring hydraulic system malfunctions, at least one microphone in the microphone array is mounted on a fixed bracket no more than 20 cm from the surface of the main hydraulic pump housing. For monitoring malfunctions of mechanical structures such as scrapers and slides, additional microphones are mounted inside the rear beam of the vehicle body, facing the linkage joint of the pusher mechanism. All microphones should be vibration-damped to reduce interference from vehicle vibrations.

[0098] Meanwhile, the vehicle's operating data is read in real time through the vehicle bus interface: the engine speed is 1100 RPM (mapped to the "low speed" range, code 02); the power take-off is in the "engaged" state (code 1); the hydraulic main pump pressure fluctuates around 16 MPa (mapped to the "medium load" pressure level, code 03); and the current hydraulic valve group's action command is "scraper compression" (code V_COMPRESS).

[0099] Adaptive noise reduction module:

[0100] The system encodes and concatenates the real-time acquired operating condition data to form a unique composite operating condition index: "02-1-03-V_COMPRESS".

[0101] The system uses this index to accurately call a healthy noise model from a pre-established background noise model library that perfectly matches the micro-operating condition of "engine low speed - PTO engagement - medium load pressure - scraper compression action".

[0102] By subtracting this noise model from the real-time acquired multi-channel audio signal, the system effectively eliminates strong but normal background noise (such as engine noise and steady-state operation noise of hydraulic pump), and finally outputs a "clean audio signal" containing weak fault characteristics.

[0103] The processed "clean audio signal" is simultaneously fed into two parallel diagnostic modules:

[0104] Path 1: Causal Analysis Module

[0105] Starting with the "scraper compression" command, the module begins to construct a real-time event chain. It sequentially identifies a series of sound events, such as "solenoid valve engagement sound," "scraper start sound," and "slide impact sound when the slide plate retracts into position."

[0106] The module compares this real-time event chain with the "scraper compression" causal acoustic template in the database.

[0107] Timing deviation comparison revealed that the actual time interval from the "scraper start sound" to the "slideboard retracting and impact sound" was 2.1 seconds, while the reference time delay stored in the template was 1.8 seconds, with a set tolerance range of ±15% (i.e., 1.53 seconds to 2.07 seconds). The actual time of 2.1 seconds exceeded the upper limit of the tolerance.

[0108] Based on this, the causal analysis module diagnosed the timing deviation as a related fault caused by "slow action due to hydraulic leakage and / or pipeline blockage", and generated an internal flag for a first-class alarm.

[0109] Path 2: Pulse Detection Module (Transient Fault Diagnosis)

[0110] The "clean audio signal" is input into a pre-trained one-dimensional convolutional autoencoder model.

[0111] During scraper operation, a tiny "click" sound is produced when the pitted gear teeth mesh. Because the autoencoder model has only learned steady audio patterns under healthy conditions, it cannot accurately reconstruct this sudden, abnormal transient impact signal.

[0112] By subtracting the input and output of the model point by point, an isolated energy pulse with an amplitude far exceeding the preset statistical threshold (e.g., the mean of the healthy error sequence + 4 times the standard deviation) appears in the reconstructed error signal sequence at the moment the "click" sound occurs.

[0113] The pulse detection module successfully detected the pulse peak, determined it to be a transient fault "caused by pitting on the gear surface," and generated an internal flag for a Class II alarm. Simultaneously, the module calculated the frequency of such pulse peaks within a complete scraper compression cycle.

[0114] The fusion decision module receives fault markers from both diagnostic paths.

[0115] Module determination: Within the same "scraper compression" action command execution window, the system simultaneously detected a correlation fault (overly slow action) and a transient fault (abnormal pulse).

[0116] Based on this, the system makes a logical decision: the physical damage to the components (gear pitting) has had an observable impact on the overall operation of the equipment (hydraulic drive efficiency). This is a typical, worsening, complex, and severe fault.

[0117] Finally, the system generated and issued a third-class alarm. The alarm clearly stated: "Complex Critical Fault: High-frequency transient impact was detected in the scraper compression mechanism (suspected damage to transmission components), which has caused the operation process to time out (overly slow action)."

[0118] A thorough examination is recommended immediately.

[0119] This system addresses the core pain point of inaccurate diagnosis due to variable operating conditions and strong background noise. By deeply integrating real-time operating data with the noise reduction process, it achieves highly targeted elimination of normal noise, ensuring the quality of the input signal. Its innovative dual-diagnostic path works in parallel: first, through causal analysis, it elevates the diagnostic scope from individual abnormal noises to understanding the equipment's operational process, effectively identifying related faults such as performance degradation; second, by analyzing the autoencoder reconstruction error, it amplifies weak transient impacts into clear pulse peaks, thereby detecting early physical damage such as bearing cracks with high sensitivity. This invention significantly improves the accuracy, depth, and sensitivity of diagnosis.

[0120] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing faults in compressed garbage trucks based on voiceprint recognition. Includes, characterized in that: Real-time acquisition of multi-channel audio signals; and simultaneous acquisition of real-time vehicle operating condition data and action commands; based on the real-time operating condition data, dynamically selecting a model matching the current operating condition from the background noise model library, and performing adaptive noise reduction processing on the multi-channel audio signals to obtain clean audio signals; A dual diagnostic path is executed in parallel for the clean audio signal: Using the action command as the starting point of the causal event, a real-time time-series event chain consisting of multiple sound events is established from the clean audio signal; by comparing the real-time time-series event chain with the causal acoustic template, it is analyzed whether there are deviations in timing, sequence and / or characteristics, and whether there are correlation faults. The clean audio signal is input into an autoencoder model pre-trained based on health status data; by analyzing the reconstruction error signal sequence between the input and output of the autoencoder model, the presence of pulse peaks caused by early damage to components is detected, and the presence of transient faults is diagnosed. A fault alarm is issued when associated faults and / or transient faults are present.

2. The method for diagnosing faults in a compressed garbage truck based on voiceprint recognition according to claim 1, characterized in that: The real-time operating data specifically includes engine speed, power take-off engagement status, hydraulic system main pump pressure, and hydraulic valve group action commands.

3. The method for diagnosing faults in a compressed garbage truck based on voiceprint recognition according to claim 1, characterized in that: The background noise model library is a combination of the healthy state sounds of the compressed garbage truck under various micro-operating conditions. Each combination of micro-operating conditions corresponds one-to-one with a background noise model, forming a mapping relationship. The step of dynamically selecting a model matching the current operating condition from the background noise model library based on the real-time operating condition data involves combining the acquired real-time operating condition data and using this combination as an index to directly call the background noise model uniquely corresponding to the index in the background noise model library. This allows for highly targeted noise cancellation based on the precise operating state of the current vehicle.

4. The method for diagnosing faults in a compressed garbage truck based on voiceprint recognition according to claim 1, characterized in that: The causal acoustic template is a pre-established database that stores the baseline occurrence sequence, baseline time delay, and baseline acoustic characteristics of a series of sound events triggered in a healthy state for each action instruction that is a causal event. The deviations include time-series deviation comparison, sequence deviation comparison, and feature deviation comparison; Timing deviation comparison determines whether the actual occurrence time of sound events in the real-time timing event chain exceeds the preset tolerance range of the reference time delay, thus diagnosing faults caused by hydraulic leakage and / or pipeline blockage that result in excessively slow or / or excessively fast movements. Sequence deviation comparison determines whether the actual occurrence order of sound events in the real-time timing event chain is consistent with the baseline occurrence order and / or whether there are missing key sound events, thus diagnosing logic errors caused by solenoid valve jamming and / or mechanical component interference. Feature deviation comparison determines whether the actual acoustic characteristics of a sound event in a real-time time-series event chain, including frequency, amplitude, and energy envelope, differ significantly from the reference acoustic characteristics, thus diagnosing faults caused by changes in operating status due to abnormal load and / or component wear.

5. The method for diagnosing faults in a compressed garbage truck based on voiceprint recognition according to claim 1, characterized in that: The autoencoder model is trained by inputting health status audio data of the compressed garbage truck under various operating conditions. The reconstructed error signal sequence includes point-by-point subtraction of the input and output of the autoencoder model to obtain a time-synchronized error value sequence, and applying a peak detection algorithm to the error value sequence to find isolated energy pulses whose amplitude exceeds a preset statistical threshold; the pulse peak is used to characterize the occurrence of transient acoustic events that are discontinuous in the time domain caused by early bearing cracks, gear surface pitting, and loosening of structural connections.

6. The method for diagnosing faults in a compressed garbage truck based on voiceprint recognition according to claim 1, characterized in that: The fault alarms include a first type of alarm, a second type of alarm, and a third type of alarm. The first type of alarm indicates the presence of a related fault involving performance degradation and / or logical errors, and indicates the specific stage at which the related fault occurs. The second type of alarm indicates the presence of a transient fault involving component-level physical damage, and indicates the frequency of abnormal pulse peaks. When both a related fault and a transient fault exist simultaneously, a third type of alarm is generated. The third type of alarm indicates the presence of a complex serious fault caused by physical damage to components that has affected the overall operation of the equipment.

7. A fault diagnosis system for compressed garbage trucks based on voiceprint recognition, characterized in that: Multi-source acquisition module: Real-time acquisition of multi-channel audio signals; and synchronous acquisition of real-time vehicle operating data and action commands; Adaptive noise reduction module: Based on the real-time operating data, dynamically selects a model matching the current operating conditions from the background noise model library, and performs adaptive noise reduction processing on the multi-channel audio signal to obtain a clean audio signal; A dual diagnostic path is executed in parallel for the clean audio signal: Causal analysis module: Taking the action command as the starting point of the causal event, it establishes a real-time time-series event chain composed of multiple sound events from the clean audio signal; By comparing the real-time event chain with the causal acoustic template, the system analyzes whether there are deviations in timing, sequence, and / or characteristics, and diagnoses whether there are associated faults. Pulse detection module: Inputs the clean audio signal into an autoencoder model pre-trained based on health status data; by analyzing the reconstruction error signal sequence between the input and output of the autoencoder model, detects whether there are pulse peaks caused by early damage to components, and diagnoses whether there is a transient fault; Fusion Decision Module: Issues a fault alarm when associated faults and / or transient faults are present.

Images