Adaptive monitoring method and system for transmission equipment based on physical constraints and acoustic array

Abstract

The application discloses a kind of based on physical constraint and acoustic array transmission equipment adaptive monitoring method and system, in the multiple domains of signal transmission and processing, anti-interference means is arranged hierarchically: in physical domain, through the multi-interface reflection of acoustic impedance mismatch layer, solid wave conduction crosstalk is significantly inhibited, and cascade priori guided blind source separation is used to realize acoustic field decoupling;In frequency domain, through the time-varying comb filter driven by rotational speed derivative, adaptive tracking and compensation of the frequency spectrum excursion caused by variable working condition are carried out;In decision domain, the TTA parameter updating direction is constrained by physical prior regularization. Three layers act on different types of interference sources respectively, and independently take effect in their respective domains, form the sub-domain cooperation architecture of "physical domain→frequency domain→decision domain", and realize the system-level robustness better than single-domain anti-interference scheme.

Description

Technical Field

[0001] This invention belongs to the field of mechanical transmission equipment condition monitoring technology, specifically relating to an adaptive monitoring method and system for transmission equipment based on physical constraints and acoustic arrays. Background Technology

[0002] Mechanical transmission equipment (referring to power transmission assemblies composed of gears, bearings, main shafts, and housings, including but not limited to wind turbine drive chains, ship main propulsion shafting, mine hoist reducers, and metallurgical rolling mill gear seats) operates in complex environments, facing the dual challenges of strong background noise and drastic fluctuations in operating conditions, posing significant difficulties for condition monitoring and early fault warning. Existing technologies mainly suffer from the following three types of deficiencies:

[0003] (1) There is severe crosstalk between the multi-channel acoustic acquisition signals.

[0004] Non-contact acoustic monitoring using microphone arrays is already available in industrial settings. However, in the chassis / shell environment of transmission equipment, structural vibrations propagate laterally through the equipment shell as solid waves, causing severe acoustic and electrical coupling (hereinafter referred to as "crosstalk") between the channels of a conventional microphone array. This means that the signals from each channel interfere with each other due to solid wave propagation and circuit parasitic effects. This physical coupling destroys the statistical independence of multi-channel signals, potentially leading to ill-conditioned or singular separation matrices in subsequent blind source separation (BSS) algorithms that rely on signal independence assumptions, thus affecting the accurate extraction of sound source features. Existing solutions mostly address this only at the software level, failing to block vibration propagation at its physical source.

[0005] (2) Conventional fixed-parameter filters under drastic operating conditions cause characteristic frequency shifts and loss.

[0006] Sudden and significant changes in operating parameters such as speed and load of transmission equipment can cause spectral smearing and real-time shifts in fault characteristic frequencies. Traditional fixed-bandwidth comb filters or order tracking methods designed based on steady-state speed cannot track time-varying fault characteristic frequencies during sudden speed changes (such as rapid acceleration or impact loads). This can lead to the filtering out of critical and weak fault information, a decrease in signal-to-noise ratio, and reduced reliability of subsequent diagnostics.

[0007] (3) The Time Adaptation (TTA) algorithm at the edge is prone to catastrophic negative transfer in an unsupervised state.

[0008] To address operational condition drift, TTA (Tracking and Tagging) technology has been gradually introduced at the edge, using unlabeled data to update model parameters (such as statistics for batch normalization layers) online during the testing phase. However, unsupervised objectives such as simple entropy minimization lack physical guidance. Under industrial data distributions characterized by strong noise and non-stationarity, the parameter update direction is prone to deviating from the physical manifold, leading to over-adaptation of the model to normal data and decreased sensitivity to real faults—a phenomenon known as "negative migration," resulting in numerous false positives or false negatives. This problem is particularly prominent when relying solely on the limited computing power of the edge for autonomous decision-making. Summary of the Invention

[0009] To address the aforementioned issues, this invention discloses an adaptive monitoring method and system for transmission equipment based on physical constraints and acoustic arrays. Through a three-layer collaborative architecture consisting of a dual decoupling design of underlying physical and algorithmic cascades, dynamic reconstruction driven by intermediate operating conditions, and top-level physical regularization TTA updates, a closed-loop monitoring system with strong bidirectional coupling between the sensing end and the edge computing end is constructed, achieving highly robust and sensitive state monitoring under varying operating conditions.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows:

[0011] An adaptive monitoring method for transmission equipment based on physical constraints and acoustic arrays includes the following steps:

[0012] Step S1: Obtain multi-channel high signal-to-noise ratio speakerprint features and perform dual decoupling of physics and algorithms.

[0013] MEMS (Micro-Electro-Mechanical Systems) acoustic arrays are deployed on the monitored components of the transmission equipment (such as gearbox housings and bearing housings). An acoustic impedance mismatch layer is placed between the array base and the equipment housing. This mismatch layer utilizes the strong reflection effect of waves at the interface of different impedance media to significantly suppress the lateral propagation of low-frequency structural vibration waves (mainly bending waves and transverse waves) along the housing surface, ensuring that the signals of each channel are mutually independent at the acquisition source (laying the foundation for the statistical independence required for subsequent blind source separation). For the multi-channel signals whose solid-conducted crosstalk has been weakened by physical isolation, blind source separation matrix operations are further performed under the guidance of a preset array spatial transfer function to separate the airborne acoustic coupling caused by the simultaneous excitation of multiple MEMS sensors by the same sound source, outputting a high signal-to-noise ratio multi-channel acoustic signature sequence.

[0014] The acoustic impedance mismatch layer consists of alternating layers of high acoustic impedance metal plates and low acoustic impedance damping silicone, with a total number of layers N≥3. The thickness of each layer is designed using the transmission matrix method. The target bandgap is defined by the main spectral peak frequency band of the transmission equipment shell structure vibration (e.g., 50Hz-2kHz in a wind turbine gearbox). The thickness of each layer is optimized to ensure that the total transmission coefficient of the vibration wave within this frequency band is below a preset threshold, thus forming a broadband transmission bandgap within the target frequency band. In the specific design, the thickness of each layer is used as the optimization variable, and minimizing the average transmission coefficient within the target frequency band is the objective function.

[0015] The pre-defined array spatial transfer function was calibrated in a semi-anechoic laboratory environment. A standard omnidirectional sound source (such as a white noise point source) was placed 2m from the center of the MEMS acoustic array. The influence of shell reflection in the actual deployment environment was compensated for by adaptive iteration in the subsequent blind source separation calculation. A precision turntable was used to drive the array to perform a full-space scan with a step size of 5°. The scan range was 0-360° horizontal angle and -60° to +60° elevation angle. The amplitude ratio and phase difference of each channel relative to the reference channel were recorded, and a multi-dimensional lookup table (hereinafter referred to as the spatial mapping table) was constructed with horizontal angle, elevation angle, frequency, and channel number as indexes. The table stores the amplitude ratio and phase difference of each channel relative to the reference channel. The horizontal angle step size was 5°, the elevation angle step size was 5°, and the frequency resolution was set according to the subsequent signal processing requirements. In blind source separation operations, let the observed signal vector x(t) = A · s(t) + n(t), where A is an L×N dimensional mixing matrix (L is the number of array channels, N is the number of sound sources), and its j-th column vector represents the spatial transfer characteristics from the j-th sound source to each MEMS sensor. This array spatial transfer function is used to constrain the initialization of the mixing matrix A. Unlike conventional algorithms that use random initialization or initialization based on observed data characteristics, this invention utilizes prior knowledge of the mechanical structure of the transmission equipment (obtaining the geometric coordinates of each sound source, such as the gear meshing area and bearing position, from design drawings or 3D models), combined with current operating parameters (such as rotational speed) provided by SCADA / PLC (data acquisition and monitoring control / programmable logic controller) to predict the main lobe direction of the sound source radiation directivity, and retrieves the corresponding array spatial transfer function vector from the spatial mapping table as the initial column vector of matrix A. Considering that structural acoustic-structure interaction mode drift caused by variable loads inside the gearbox leads to changes in the radiation field, this invention sets a tolerance neighborhood with preset angles (e.g., horizontal ±15°, elevation ±10°) based on the initial column vector reference value. Furthermore, utilizing the time-frequency sparsity of the current frame's observed signal, the direction vectors of single-source point clusters are extracted within the tolerance neighborhood using local clustering algorithms (e.g., K-means clustering or DBSCAN clustering). This constrained iterative correction of the mixing matrix yields the true mixing matrix column vector. This approach leverages physical priors to overcome the permutation ambiguity of blind source separation algorithms and prevents iterative divergence caused by relying solely on unconstrained data. In addition, when the deviation between the corrected real column vector and the initial reference value exceeds a preset threshold, the system triggers an online calibration mechanism to update the transfer function value of the corresponding node in the spatial mapping table in reverse. This enables the lookup table to have online adaptive evolution capability as the equipment ages and operating conditions change, ensuring that each separated source signal has a fixed one-to-one correspondence with the physical sound source (for example, in the gearbox monitoring scenario, each separated channel can correspond to each physical sound source such as gear meshing noise and bearing vibration noise).

[0016] Thus, the physical isolation of the acoustic impedance mismatch layer achieves decoupling of solid-borne crosstalk, and the blind source separation guided by the array transfer function achieves decoupling of spatial acoustic coupling, completing the physical-algorithm dual decoupling of multi-channel acoustic features.

[0017] Step S2: Parallel acquisition of transient parameters of operating conditions

[0018] Real-time operating data is read from the SCADA / PLC system of the transmission equipment via interfaces such as industrial Ethernet and fieldbus, extracting operating parameters such as instantaneous speed r(t), load torque, and output power. The first-order time derivative of the instantaneous speed is calculated. (Used for real-time calculation of filter bandwidth in subsequent step S3).

[0019] Step S3: Construct a time-varying comb filter based on the rotational speed derivative and perform signal reconstruction.

[0020] A time-varying comb filter with an adaptive passband adjustment mechanism is constructed using instantaneous rotational speed and its derivative. The instantaneous notch bandwidth B(t) of this filter satisfies the following dynamic evolution model:

[0021] ;

[0022] in, The basic narrowband bandwidth required to extract specific fault characteristics under steady-state conditions is defined by k, which is a dynamic margin coefficient preset based on the equipment's moment of inertia and the order of the fault characteristics. The absolute value of |dr(t) / dt| is taken to ensure that passband broadening is triggered during acceleration and deceleration transients. When the rotational speed changes drastically, the filter passband automatically broadens to cover the spectral shift characteristics; when returning to a quasi-steady state, the passband narrows to restore high selectivity.

[0023] The time-varying comb filter described above is used to independently filter and reconstruct the acoustic signature feature sequences of each channel output in step S1, removing the characteristic frequency disturbances caused by the changing operating conditions and outputting physical guidance features decoupled from the current transient operating conditions. The reconstructed feature sequences of each channel retain their original channel numbers for use in subsequent steps.

[0024] Step S4: Detect operating condition distribution drift and trigger the adaptive engine

[0025] Based on the SCADA / PLC operating condition data stream (including parameters such as speed, load torque, and output power) acquired in real time in step S2, the sliding window statistical distribution of these parameters is continuously maintained at the edge computing node. The distribution drift metric between the current data window and the historical reference window is calculated online (e.g., using symmetric KL divergence, Jensen-Shannon divergence, or maximum mean difference). The length of the current data window is set according to the expected time scale of operating condition changes, while the historical reference window uses steady-state operating data segments under normal equipment conditions; their lengths can be the same or different, depending on the equipment's operating characteristics.

[0026] When the drift exceeds a preset safety threshold, a substantial change in equipment operating conditions is identified, and this is used as a hard trigger signal to activate the Test Time Adaptation (TTA) engine at the edge. This threshold is determined based on the historical statistical distribution of drift measurements under normal operating conditions, or a fixed threshold is set according to equipment operating experience to ensure that false alarms are not triggered within the fluctuation range of normal operating conditions. After the TTA engine is activated, the unsupervised parameter update and diagnostic inference described in step S5 are executed using the physical guidance features output in step S3.

[0027] Step S5: Perform unsupervised parameter evolution and diagnosis under physical constraints.

[0028] This invention deploys a lightweight one-dimensional convolutional autoencoder as a diagnostic model at the edge. Its structure is as follows: the encoder progressively compresses the input features into a low-dimensional latent space representation, and the decoder reconstructs the input based on this representation; the encoder's end is connected to a fault classification head, which outputs the probability distribution of each fault category. The TTA engine takes the physical guidance features reconstructed in step S3 as input and performs parameter updates under unlabeled conditions.

[0029] The core lies in constructing a total loss function that includes a physical prior regularization term:

[0030] ;

[0031] in, To minimize the entropy of the predicted probability distribution of the model output, thereby increasing the confidence in the current distribution; λ is the physical prior regularization term; λ is the balance coefficient; the regularization term is constructed as follows: based on real-time SCADA / PLC operating condition data and mechanical transmission dynamics rules, enumerate the set of fault categories that are physically impossible under the current operating condition (for example, high-order harmonic faults of gear meshing frequency are impossible at extremely low speeds), and generate a binary constraint mask vector M with a length equal to the total number of fault categories, setting the positions of impossible categories to 1. The design is to penalize the model for outputting high-confidence probabilities on these impossible categories:

[0032] ;

[0033] Among them, 1 {·} For indicator functions, For the set of impossible fault categories, This represents the model's output probability for class c. This regularization term effectively suppresses diagnostic results that violate physical laws. During parameter updates, the backbone convolution kernels are frozen to preserve the general fault features learned in the source domain. Only the statistics of the batch normalization layer and the bias term of the last linear layer of the classifier are updated to avoid catastrophic forgetting while adapting to operating condition drift. Finally, an adaptive state diagnostic result is output based on the updated model.

[0034] The balance coefficient λ of the joint loss function is adaptively adjusted based on the transient severity of the current operating condition:

[0035] ;

[0036] Where λ0 is the basic regularization coefficient and α is a preset proportional coefficient, so that the physical constraint strength is higher when the working conditions are more severe, preventing the parameter update from deviating from the region that conforms to the physical rules.

[0037] The beneficial effects of this invention are as follows:

[0038] 1. Construct a multi-domain hierarchical anti-interference architecture from the physical domain to the decision domain to improve system-level robustness.

[0039] This invention employs layered anti-interference measures across multiple domains of signal transmission and processing: In the physical domain, multi-interface reflection through an acoustic impedance mismatch layer significantly suppresses solid-wave propagation crosstalk, and spatial acoustic coupling is separated by blind source separation guided by the array transfer function in the same step (corresponding to step S1), achieving dual decoupling at the signal acquisition end through a cascaded physical and algorithmic approach; in the frequency domain, a time-varying comb filter driven by the rotational speed derivative adaptively tracks and compensates for spectral shifts caused by varying operating conditions (corresponding to step S3); in the decision domain, the TTA parameter update direction is constrained through physical prior regularization (corresponding to step S5). These three layers act independently on different types of interference sources, forming a domain-specific cooperative architecture of [physical domain → frequency domain → decision domain], achieving superior system-level robustness compared to single-domain anti-interference schemes.

[0040] 2. Significantly suppresses crosstalk at the physical source, greatly improving the signal-to-noise ratio of voiceprint features.

[0041] By introducing an alternating high and low impedance acoustic impedance mismatch layer into the MEMS acoustic array base, the lateral propagation of low-frequency mechanical structural waves along the shell surface can be effectively suppressed, enhancing the physical independence of multi-channel signals. Combined with blind source separation guided by the array spatial transfer function, residual spatial crosstalk is further suppressed, which helps improve the signal-to-noise ratio of multi-channel acoustic signatures, providing a higher quality data foundation for subsequent operational condition-fault decoupling.

[0042] 3. Achieve high-fidelity capture of fault characteristics under drastic operating conditions.

[0043] The time-varying comb filter driven by the rotational speed derivative adaptively matches the intensity of transient changes in operating conditions: the passband widens during rapid acceleration, effectively preventing characteristic frequencies from shifting out of the passband; the passband narrows during steady-state operation, restoring high selectivity. This physically guided reconstruction helps reduce the interference of varying operating conditions on acoustic signature features and improves the stability of fault feature extraction under varying operating conditions.

[0044] 4. Suppress catastrophic negative migration in TTA unsupervised updates to ensure physical consistency of diagnostic output.

[0045] Embedding a constraint mask regularization term generated based on SCADA / PLC physical information into the TTA loss function constrains the gradient updates of network parameters within a manifold that conforms to the laws of mechanical dynamics. This reduces the risk of the model giving high-confidence diagnoses of physically impossible fault states under strong noise and variable operating conditions. It helps to avoid the counterintuitive bias in diagnosis caused by the lack of physical guidance from unsupervised targets in traditional TTA and improves the physical consistency of diagnostic results under variable operating conditions. Attached Figure Description

[0046] Figure 1 A schematic diagram of the hardware architecture of an edge computing terminal and MEMS acoustic array with integrated acoustic impedance mismatch layer components.

[0047] Figure 2 Comparison of the dynamic evolution of bandwidth of time-varying comb filters.

[0048] List of identifiers in attached diagrams:

[0049] 1. Cast aluminum housing; 2. Protective mesh; 3. ePTFE acoustic diaphragm; 4. MEMS microphone array; 5. Driver board; 6. Synchronous clock circuit; 7. Calculation motherboard; 8. Power supply; 9. Heat sink; 10. Acoustic impedance mismatch layer. Detailed Implementation

[0050] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0051] Hardware Example: Edge Computing Terminal with Integrated Acoustic Impedance Mismatch Layer

[0052] See Figure 1 The edge computing terminal in this embodiment is integrated into a single package consisting of a cast aluminum shell 1, a protective mesh 2, an ePTFE acoustic diaphragm 3, a MEMS microphone array 4, a driver board 5, a synchronous clock circuit 6, a computing motherboard 7, a power supply 8, heat sink fins 9, and an acoustic impedance mismatch layer 10. The MEMS microphone array 4 provides uniform spatial resolution within a 360° azimuth range. The computing motherboard 7 is responsible for executing all edge inference tasks from steps S1 to S5. The synchronous clock circuit 6 consists of a high-precision temperature-compensated crystal oscillator and a clock distribution chip, providing a unified sampling clock and acquisition trigger signal for the multi-channel analog-to-digital converter, and providing a hardware-level reference for the phase consistency of the array's spatial transfer function. This synchronous clock circuit uses a high-precision temperature-compensated crystal oscillator and a clock distribution chip as the master clock source for multi-channel synchronous acquisition, independently driving each channel's analog-to-digital converter without relying on external timing signals. In addition, the circuit can reserve an external second pulse input interface (such as GPS / BeiDou or IEEE 1588 PTP network synchronization signal) as an optional long-term drift calibration method to further improve synchronization accuracy; in the crystal oscillator timekeeping mode where the master clock works independently, the accumulated error in the short term can also meet the acquisition reference accuracy requirements, and the system can work normally without relying on external time signals.

[0053] The core improvement of this invention lies in replacing the conventional single damping pad with an acoustic impedance mismatch stacked assembly between the substrate of the MEMS microphone array 4 and the terminal mounting surface. This assembly, from bottom to top, consists of: a 2mm thick stainless steel plate (high acoustic impedance), a 1.5mm thick high-damping silicone sheet (low acoustic impedance), a 2mm thick aluminum alloy plate (high acoustic impedance), and a 1.5mm thick silicone sheet, forming a three-layer impedance abrupt change interface. A specified torque is applied to the connecting bolts of the device housing.

[0054] In this embodiment, the thickness of each layer of the acoustic impedance mismatch layer is determined by optimization using the transfer matrix method. Specifically, a characteristic transfer matrix for each layer is constructed based on the density and sound velocity of each material, and these matrices are cascaded and multiplied to obtain the global transfer matrix for calculating the theoretical transmission coefficient. During optimization, the upper limit of the total thickness and the lower limit of the single-layer thickness are used as constraints, the peak frequency band of the main structural vibration of the transmission equipment is used as the target frequency band, the thickness of each layer is used as the design variable, and minimizing the average transmission coefficient within the target frequency band is used as the objective function to obtain the thickness parameters of each layer. The aforementioned 2mm thick stainless steel plate, 1.5mm thick damping silicone sheet, 2mm thick aluminum alloy plate, and 1.5mm thick silicone sheet are a set of preferred implementation parameters for the target frequency band output by this optimization method, which can make the transmission coefficient of interference waves within the target frequency band lower than a preset threshold.

[0055] Details of formula and parameter derivation

[0056] 1. Dynamic bandwidth parameters of time-varying comb filters

[0057] In this embodiment, the dynamic bandwidth formula of the time-varying comb filter is:

[0058] ;

[0059] Figure 2 The dynamic evolution curve of this bandwidth as a function of the rotational speed derivative is presented. Taking a single-stage planetary gear train of a certain transmission equipment as an example (such as the gearbox of a doubly-fed wind turbine), the following is set... ( (This is the current instantaneous meshing frequency) to ensure that the meshing sidebands can be separated under steady state. The dynamic margin coefficient k is determined based on the rotor's moment of inertia and the maximum rate of change of rotational speed. In this embodiment, k = 0.15s is taken. When the derivative of rotational speed increases, the filter bandwidth increases accordingly on the basis value, which can cover the sideband offset range caused by impact loads or rotational speed fluctuations.

[0060] 2. Example of Joint Loss Function Calculation

[0061] In this embodiment, the encoder end of the one-dimensional convolutional autoencoder is connected to a fault classification head, and the output of the classification head is the probability distribution of each fault category. For the t-th batch of unlabeled reconstructed feature samples, the entropy minimization loss is calculated as follows:

[0062] ;

[0063] The physical constraint mask generation rules are as follows: Read the instantaneous SCADA / PLC speed r corresponding to the sample, and set the mask bit corresponding to the fault category index when r < 500 rpm (this threshold corresponds to the rated speed of the low-speed shaft of the gearbox in this embodiment, which is about 600–800 rpm. The working condition below 500 rpm has deviated from the normal design point. Under this working condition, categories such as high-speed gear meshing faults are physically unlikely to become the dominant fault source). Set to 1. The physical regularization term is calculated as follows:

[0064] ;

[0065] At the edge, only the batch normalized layer statistics of the feature extractor and the bias of the last linear layer of the classifier are updated, while the backbone convolution kernels are frozen to avoid forgetting historically learned failure modes. The batch normalized layer statistics are updated using a small batch intra-statistic update method to adapt to the current operating condition distribution.

[0066] It should be noted that in this embodiment, the direction of the main lobe of the sound source radiation is estimated using prior knowledge of the mechanical structure and operating parameters. The main purpose is to provide a physically meaningful initialization benchmark for the blind source separation algorithm, thereby effectively reducing the search space for solutions. During subsequent adaptive iterations, the blind source separation algorithm inherently possesses a certain tolerance for initial errors in the mixing matrix. Typically, when the estimated main lobe direction deviates from the true direction within a certain range (e.g., a reasonable deviation of approximately ±20°), this initialization benchmark helps guide the algorithm towards an effective separation matrix. For extreme variable operating conditions where the estimation deviation expands due to sound field structure drift, this embodiment further compensates dynamically through the aforementioned local clustering of the limited tolerance neighborhood and the reverse update evolution mechanism of the lookup table. This strategy, while allowing for reasonable errors in the initial estimation, effectively improves the convergence reliability and physical consistency of the separation algorithm under complex operating conditions.

[0067] Application scenario action examples

[0068] Taking the monitoring of a minor spalling fault in the inner ring of a wind turbine drivetrain bearing during acceleration from 600 rpm to 900 rpm on a wind turbine drivetrain bearing failure simulation test bench as an example:

[0069] SCADA speed starts from 600 rpm and increases by 80 rpm / s 2 The acceleration begins to increase, and S2 calculates in real time. When an increase in the rotational speed derivative is detected, the time-varying comb filter in step S3 continuously shifts its center frequency along the meshing frequency order trajectory, while the passband dynamically broadens according to B(t). During acceleration, the filter's passband center frequency continuously tracks the meshing frequency order trajectory. When the rotational speed reaches approximately 700 rpm, although the bearing's characteristic frequency spectrum peaks broaden due to the spectral smearing effect, the filter's passband width has broadened synchronously, and the characteristic energy has not overflowed the passband range.

[0070] During acceleration, the SCADA / PLC operating condition vector distribution drifts. In step S4, when the KL divergence exceeds a preset threshold (0.3 in this embodiment), the TTA engine is activated. In step S5, the TTA engine updates the BN statistics online based on the reconstructed features. In this embodiment, since the rotational speed is consistently higher than the threshold set by the masking rules (500 rpm in this embodiment), the constraint mask M does not mask any fault categories. The physical regularization term primarily penalizes the ultra-high confidence output to prevent the model from overconfidently judging normal operating condition fluctuations. After acceleration, at the stable 900 rpm stage, the model output indicates a significant increase in the probability of inner-circle faults from the normal baseline level (background value). Warning information is displayed locally at the edge, and the compressed feature package and diagnostic logs are stored in the onboard non-volatile memory for local offline viewing or access by the host computer via the field LAN.

[0071] In contrast, if the traditional TTA method, which relies solely on entropy minimization, is used, the model may adapt the distribution characteristics of the spectrum stretching during the acceleration process to a "normal" mode in order to reduce reconstruction error, thereby reducing sensitivity to real fault characteristics and leading to missed detections. This invention can effectively suppress this problem through physical constraints.

[0072] To verify the diagnostic performance of different strategies, preliminary comparative tests were conducted on the aforementioned testbed. The dataset was set as follows: the offline training set was collected under steady-state conditions at 600 rpm, containing 800 normal and 800 faulty samples, used for supervised pre-training of the model; the online test set was collected during the acceleration process from 600 rpm to 900 rpm, containing 200 normal and 200 faulty samples (streamed input). During the testing phase, only unlabeled features were used to trigger TTA updates, and the training and test sets did not overlap in time.

[0073] The experiment compared three diagnostic strategies: Method A (traditional fixed-parameter monitoring model without adaptive mechanism), Method B (introducing traditional TTA, using only entropy minimization loss), and the proposed method. All three methods used the same network backbone and were trained and evaluated using the same training and test sets. The diagnostic accuracy and false negative rate of the experiment are shown in the table below.

[0074] Comparison Methods accuracy Fault false alarm rate Method A 64.3% 18.3% Method B 76.5% 24.7% This method 92.4% 5.2%

[0075] Table 1 shows that Method A, lacking an adaptive mechanism, struggles to capture feature drift caused by changes in operating conditions, resulting in relatively low accuracy and a high false negative rate. Method B, by introducing entropy minimization TTA, improves accuracy, but due to the lack of physical guidance during adaptive updates, it easily misclassifies some fault features as operating condition fluctuations, leading to a higher false negative rate compared to Method A. In contrast, our proposed method, under physical prior regularization constraints, shows significant improvements in both accuracy and false negative rate.

[0076] Through the three-layer collaborative architecture consisting of physical domain decoupling, frequency domain reconstruction, and decision domain constraints, the system proposed in this invention has better diagnostic reliability than traditional monitoring schemes under strong noise and variable operating conditions, and is especially suitable for unattended industrial scenarios with drastic changes in operating conditions, such as wind power, ships, and mines.

[0077] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. An adaptive monitoring method for transmission equipment based on physical constraints and acoustic arrays, characterized in that, Includes the following steps: S1: Deploy a MEMS acoustic array on the component to be monitored in the transmission equipment. Set an acoustic impedance mismatch layer between the array base and the equipment shell to achieve structural vibration isolation. For the isolated multi-channel signal, perform a hybrid matrix operation for blind source separation under the guidance of the preset array spatial transfer function. The initialization of the hybrid matrix is ​​based on the prior lookup table value. Iterative correction is performed in the tolerance neighborhood by sparse clustering in the time and frequency domain. The lookup table is updated in reverse according to the correction deviation to separate the acoustic coupling in the air and output a high signal-to-noise ratio multi-channel acoustic feature sequence. S2: Real-time acquisition of the operating condition data stream of the transmission equipment, wherein the operating condition data includes at least instantaneous speed r(t), load torque, and output power, and calculation of the first-order time derivative of the instantaneous speed. ; S3: Construct a time-varying comb filter with an adaptive passband adjustment mechanism using the instantaneous rotational speed r(t) and its derivative, and independently perform filtering and reconstruction on the multi-channel acoustic signature sequence channel by channel to remove the characteristic frequency disturbances caused by the variable operating conditions and output the physical guidance features; S4: Based on the real-time acquired working condition data stream, maintain the sliding window statistical distribution of the edge computing node, and calculate the distribution drift metric between the current data window and the historical reference window online. When the drift exceeds the preset safety threshold, activate the test adaptation engine. S5: Using the physical guidance features as input, perform unsupervised parameter updates under physical constraints on the diagnostic model deployed at the edge; wherein, construct a total loss function containing a physical prior regularization term, the physical prior regularization term generates a binary constraint mask vector with a length equal to the total number of fault categories based on the current operating condition information, to penalize the model for outputting high confidence probabilities for fault categories that are physically impossible to occur under the operating condition, and output adaptive state diagnostic results based on the updated model.

2. The method according to claim 1, characterized in that, In step S1, the acoustic impedance mismatch layer is composed of alternating layers of high acoustic impedance metal plates and low acoustic impedance damping silicone, with a total number of layers N≥3; the thickness of each layer is designed by the transmission matrix method so that the main frequency peak band of the vibration of the transmission equipment shell structure is the target band gap, and the total transmission coefficient is lower than the preset threshold.

3. The method according to claim 1, characterized in that, In step S1, the array spatial transfer function is pre-generated and stored as a multidimensional lookup table; the initialization process of the hybrid matrix for blind source separation is to estimate the main lobe direction of the radiation directivity of each sound source based on the prior knowledge of the mechanical structure of the transmission equipment and the current working parameters, and retrieve the corresponding spatial transfer function vector from the lookup table as the initial column vector of the hybrid matrix. During the blind source separation process, a tolerance neighborhood with a preset angle is set based on the initial column vector. The time-frequency domain sparsity of the current observed signal is used to perform local clustering of single-source point clusters within the tolerance neighborhood. The true hybrid matrix column vector is obtained through iterative correction to break the permutation ambiguity of the blind source separation algorithm. When the deviation between the true hybrid matrix column vector and the reference value of the initial column vector exceeds a preset threshold, the transfer function value of the corresponding spatial grid point in the lookup table is updated in reverse using the corrected vector as an index, thereby realizing the online evolution of the lookup table.

4. The method according to claim 1, characterized in that, In step S3, the instantaneous notch bandwidth B(t) of the time-varying comb filter satisfies: ； in, The basic narrowband bandwidth required to extract specific fault characteristics under steady-state conditions, where k is a dynamic margin coefficient preset based on the equipment's moment of inertia and the order of the fault characteristics.

5. The method according to claim 1, characterized in that, In step S5, the total loss function is: ； in, To minimize the entropy of the predicted probability distribution of the model output, thereby increasing the confidence in the current distribution; For physical prior regularization terms, This is the balance coefficient; The physical prior regularization term is calculated as follows: ； Among them, 1 {·} For indicator functions, For the set of impossible fault categories, This represents the output probability of the model in class c.

6. The method according to claim 5, characterized in that, In step S5, the balance coefficient of the total loss function Adaptive adjustment based on the transient intensity of the current operating conditions: ； Where λ0 is the basic regularization coefficient and α is the preset proportional coefficient.

7. The method according to claim 1, characterized in that, In step S5, the model is a lightweight one-dimensional convolutional autoencoder. During the parameter update process, the backbone convolutional kernel is frozen, and only the statistics of the batch normalization layer and the bias term of the last linear layer of the classifier are updated.

8. An adaptive monitoring system for transmission equipment based on physical constraints and acoustic arrays, used to perform the method as described in any one of claims 1 to 7, characterized in that, include: MEMS acoustic arrays are mounted on the components to be monitored in transmission equipment through an acoustic impedance mismatch layer to acquire multi-channel acoustic signals. The edge computing terminal is connected to the MEMS acoustic array for receiving operating data and performing signal processing and diagnostic inference. The edge computing terminal includes a computing motherboard, which is used to execute all edge inference tasks in steps S1 to S5.

9. The system according to claim 8, characterized in that, The edge computing terminal also includes a synchronous clock circuit, which consists of a high-precision temperature-compensated crystal oscillator and a clock distribution chip. This circuit provides a unified sampling clock and acquisition trigger signal for the multi-channel analog-to-digital converter to ensure the phase consistency of the multi-channel signals.

Images