Production line fault inspection method based on regional attention learning mechanism

Abstract

The invention discloses a production line fault inspection method based on a regional attention learning mechanism, and relates to the technical field of production line fault inspection. Low-frequency vibration, high-frequency acoustic emission and temperature time sequence signals are synchronously acquired, and multi-dimensional fault features such as mechanical wear, shaft system misalignment and heat effect abnormality are covered; the problem of incomplete coverage of traditional single signal detection is solved. For different signal characteristics, key abnormal characteristics are focused, redundant noise is filtered, and the characteristic effectiveness is improved. A collaborative degradation factor model is constructed based on health historical data, a weight coefficient is dynamically adjusted by feeding back a drop rate after maintenance, equipment working condition changes are adapted, and the problem of poor adaptability of a static model is avoided. And the maintenance effect (collaborative degradation factor decline rate) is quantitatively evaluated, accurate maintenance is guided, invalid maintenance is reduced, and the operation and maintenance efficiency of a production line is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of production line fault detection, and specifically to a production line fault detection method based on a regional attention learning mechanism. Background Art

[0002] In modern industrial production, production line rotating equipment (such as motors, fans, pumps, etc.) serves as the core power unit, and its operating status directly affects the stability and production efficiency of the production line. Timely and accurate detection of equipment failures and early warning are key to avoiding unplanned downtime, reducing maintenance costs, and ensuring safe production. Traditional production line equipment fault inspection methods mainly rely on single signal detection technology, such as bearing wear detection based on vibration signals or thermal overload early warning based on temperature signals. However, with the increasing complexity of industrial equipment, equipment failures often manifest as multiple types of composite degradation (such as the coexistence of bearing wear and shaft misalignment, mechanical damage and thermal effects coupling), and single signal detection has the following significant defects:

[0003] Focusing only on a certain physical quantity (such as vibration energy or temperature value) makes it difficult to capture the collaborative degradation characteristics of multiple fault types, which can easily lead to missed or misjudgment; moreover, existing methods mostly use static weight models (such as fixed weight coefficients of vibration, acoustic emission, and temperature characteristics), which cannot adapt to the operating condition fluctuations (such as load changes, differences in ambient temperature and humidity) or feature distribution shifts caused by aging during long-term operation of the equipment, and the detection accuracy decreases after long-term use. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a production line fault inspection method based on regional attention learning mechanism.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows:

[0006] This application discloses a production line fault detection method based on a regional attention learning mechanism, comprising the following steps:

[0007] Synchronously collect low-frequency vibration signals, high-frequency acoustic emission signals, and temperature timing signals of rotating equipment on the production line;

[0008] Performing frequency band decomposition on the low-frequency vibration signal to obtain a vibration eigenvector; performing pulse density analysis on the high-frequency acoustic emission signal to obtain an acoustic emission eigenvector; performing data processing on the temperature time series signal to extract the absolute value of the temperature change acceleration;

[0009] Obtain historical data of the same type of equipment in a healthy state and build a collaborative degradation factor calculation model based on the historical data;

[0010] Input the vibration eigenvector, acoustic emission eigenvector, and absolute value of temperature change acceleration into the collaborative degradation factor calculation model to obtain the collaborative degradation factor;

[0011] Determine whether the collaborative degradation factor exceeds the preset warning threshold, and if so, trigger a fault alarm instruction;

[0012] Obtaining low-frequency vibration signals, high-frequency acoustic emission signals, and temperature time series data after maintenance is performed by staff according to the fault alarm instructions;

[0013] Recalculate the synergistic degradation factor based on the re-acquired low-frequency vibration signal, high-frequency acoustic emission signal and temperature time series data, and calculate the decline rate of the synergistic degradation factor before and after maintenance;

[0014] Determine whether the decline rate of the synergistic degradation factor is lower than a preset decline threshold. If so, adjust the model parameters of the synergistic degradation factor calculation model, and recalculate the decline rate of the synergistic degradation factor based on the adjusted model parameters until the decline rate exceeds the preset decline threshold, triggering a model optimization completion instruction.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1. By covering three core physical characteristics of the equipment: mechanical vibration (reflecting bearing wear), acoustic emission (reflecting shaft misalignment), and thermal effect (reflecting abnormal heat dissipation), it solves the problem that traditional single signal detection (such as vibration or temperature alone) cannot cover complex faults;

[0017] 2. By quantitatively integrating the week-to-week vibration change rate, the monthly mean acoustic emission offset, and the absolute value of temperature acceleration, a multi-dimensional assessment of equipment short-term anomalies, long-term degradation, and thermal effect coupling is achieved;

[0018] 3. In response to the fault feature offset caused by equipment aging, environmental changes or load fluctuations, a dynamic adjustment mechanism of the weight coefficient is used to ensure the accuracy of the test results and maintain the accuracy of the test results after long-term use. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0020] Figure 1 is a flow chart of the method of the present invention;

[0021] Figure 2 It is a data flow diagram of the present invention. DETAILED DESCRIPTION

[0022] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present invention and should not be regarded as the entire invention or as a limitation or restriction of the technical solution of the present invention.

[0023] Application Overview

[0024] In traditional production line rotating equipment fault detection, single signal detection technology has difficulty effectively capturing multiple types of composite degradation characteristics. Vibration signal analysis focuses on monitoring mechanical component wear, and temperature signals reflect changes in thermal effects. However, the two exhibit nonlinear correlation in scenarios where bearing wear and shaft misalignment coexist, resulting in a single signal detection model being unable to accurately identify composite fault modes. Static weight models are unable to adapt to feature distribution shifts caused by load fluctuations during long-term equipment operation. For example, under high-load conditions, the correlation between vibration energy baseline drift and the temperature change rate changes, causing the weight coefficient to mismatch the current operating conditions. Fault detection accuracy decreases with extended operation time.

[0025] For example, in a combined fault scenario involving motor bearing wear and cooling system degradation, the vibration signal energy increases significantly in the initial degradation phase, while the acceleration of temperature changes exhibits a delayed response due to cooling delays. A single vibration detection model triggers an early warning before the temperature anomaly reaches the threshold, but the actual equipment remains in a safe state, resulting in an increased false alarm rate. Furthermore, seasonal variations in ambient temperature cause fluctuations in the temperature signal baseline. The static weight model over-amplifies the contribution of temperature features in high-temperature summer conditions, masking the subtle early wear characteristics in the vibration signal and causing missed detections.

[0026] If these issues are not addressed, false alarms and missed detections in complex fault modes will increase the frequency of unplanned downtime and raise maintenance costs. Static weight models are unable to dynamically adapt to changing operating conditions. After long-term operation, detection accuracy continues to decline, reducing the reliability of equipment health assessments and increasing the risk of sudden failures. In complex degradation scenarios involving multiple coupled physical quantities, traditional methods struggle to establish cross-modal feature correlations, impacting the accuracy of fault root cause location, delaying the timeliness of maintenance decisions, and ultimately leading to increased equipment performance degradation and shortened service life.

[0027] When faced with the above problems, this application first considers how to capture composite degradation features through multi-source signal fusion. Traditional single signal detection is difficult to correlate the nonlinear relationship between vibration, acoustic emission and temperature, resulting in false alarms and missed detections. To this end, this application attempts to synchronously collect low-frequency vibration, high-frequency acoustic emission and temperature time series signals, extract multi-dimensional feature vectors through frequency band decomposition, pulse density analysis and acceleration calculation, and construct cross-modal feature associations. In order to solve the problem that the static weight model cannot adapt to operating condition fluctuations, this application introduces a maintenance feedback mechanism to re-evaluate the feature contribution after each maintenance, dynamically adjust the weight coefficient, and make the model continue to match the current equipment status. In addition, in response to the differences in different signal response delays in composite faults, this application proposes to establish a collaborative degradation factor benchmark based on historical health data, quantify the overall degradation degree through weighted calculation, and combine the dual judgment mechanism of warning threshold and decline rate to achieve closed-loop optimization of fault detection and maintenance effect verification.

[0028] In this regard, Figure 1 As shown, this application proposes a production line fault detection method based on a regional attention learning mechanism, comprising the following steps:

[0029] Synchronously collect low-frequency vibration signals, high-frequency acoustic emission signals, and temperature timing signals of rotating equipment on the production line;

[0030] Low-frequency vibration signals, high-frequency acoustic emission signals, and temperature time-series signals can establish a complete monitoring chain for equipment, from microscopic damage (high-frequency acoustic emission signals) to macroscopic dynamics (low-frequency vibration signals) to system thermodynamics (temperature time-series signals). High-frequency acoustic emission signals capture microscopic fracture sounds within the material, low-frequency vibration signals record the vital signs of the mechanical structure, and temperature time-series signals monitor the system's energy metabolism status. This combination achieves full lifecycle coverage of the equipment's health status at a reasonable cost, avoiding missed early faults (the advantage of acoustic emission) while identifying complex failure modes (multi-signal correlation). Temperature signals provide a safety guarantee for the system, making it the optimal solution for predictive maintenance of industrial equipment. Low-frequency vibration signals refer to the basic physical quantities used to monitor the mechanical vibration status of rotating equipment on a production line. This can be achieved by using an acceleration sensor to collect vibration signals on the surface of the equipment housing. The frequency band is typically set to 0-10kHz to cover typical mechanical failure characteristics such as bearing wear and shaft misalignment. High-frequency acoustic emission signals are dynamic stress wave signals that reflect damage to the equipment's internal microstructure. These signals can be acquired using a resonant acoustic emission sensor in the 100kHz-1MHz frequency range, capturing transient events such as bearing crack propagation and lubrication failure. Temperature time series signals continuously record temperature changes in key equipment components. These signals can be generated using infrared thermometers or embedded thermocouples and are used to monitor the equipment's operating thermal status.

[0031] The low-frequency vibration signal is subjected to frequency band decomposition to obtain a vibration feature vector; the high-frequency acoustic emission signal is subjected to pulse density analysis to obtain an acoustic emission feature vector; the temperature time series signal is subjected to data processing to extract the absolute value of the temperature change acceleration; the vibration feature vector refers to a multidimensional feature set that characterizes the vibration state of the equipment. Specifically, the wavelet packet decomposition technology can be used to perform frequency band energy analysis on the low-frequency vibration signal, and the energy proportion of each sub-band is extracted to form a feature vector for quantifying the degree of mechanical degradation of the equipment. The acoustic emission feature vector refers to a characteristic indicator that characterizes the damage activity inside the equipment. Specifically, the pulse density statistical method can be used to calculate the number of acoustic emission events that exceed the threshold per unit time to form a feature vector that reflects the damage activity. The absolute value of the temperature change acceleration refers to the absolute value of the second-order derivative of the temperature time series signal. Specifically, the least squares method can be used to fit the temperature time series curve and extract the quadratic term coefficient to quantify the abnormal temperature rise rate of the equipment.

[0032] Obtain historical data of the same type of equipment in a healthy state, and construct a collaborative degradation factor calculation model based on the historical data; the collaborative degradation factor calculation model refers to a mathematical model that integrates multi-source signal characteristics to evaluate the overall degradation degree of the equipment. Specifically, the gradient descent method can be used to optimize the weight coefficients of vibration, acoustic emission, and temperature characteristics, and the weighted calculation can comprehensively reflect the multi-dimensional degradation coupling effects such as mechanical damage, microcracks, and thermal anomalies.

[0033] The vibration eigenvector, acoustic emission eigenvector, and absolute value of temperature change acceleration are input into the collaborative degradation factor calculation model, and the collaborative degradation factor is obtained after weighted calculation;

[0034] Determine whether the collaborative degradation factor exceeds the preset warning threshold. If so, generate a fault alarm instruction; the preset warning threshold refers to the critical value of the degradation factor that triggers the fault alarm. Specifically, the upper limit of the degradation factor distribution of healthy equipment can be calculated using historical data statistical methods to distinguish normal fluctuations from abnormal degradation states.

[0035] Re-acquire the low-frequency vibration signal, high-frequency acoustic emission signal, and temperature time series data after the maintenance performed by the staff according to the fault alarm instruction;

[0036] Recalculate the synergistic degradation factor based on the re-acquired low-frequency vibration signal, high-frequency acoustic emission signal and temperature time series data, and calculate the decline rate of the synergistic degradation factor before and after maintenance;

[0037] Determine whether the decline rate of the collaborative degradation factor is lower than the preset decline threshold. If so, adjust the weight coefficient in the weighted calculation of the collaborative degradation factor, and recalculate the decline rate of the collaborative degradation factor based on the adjusted weight coefficient until the decline rate is greater than the preset decline threshold. Post-maintenance data recalculation refers to a data closed-loop mechanism for verifying the maintenance effect. Specifically, the maintenance effectiveness can be evaluated by comparing the decline rate of the degradation factor before and after maintenance, providing feedback basis for dynamically adjusting the weight coefficient. Weight coefficient adjustment refers to the adaptive process of optimizing feature weights according to maintenance effects. Specifically, an incremental learning algorithm can be used to prioritize the adjustment of acoustic emission feature weights to solve the problem that traditional static weight models cannot adapt to working condition drift.

[0038] The core innovation of this application lies in constructing a multi-source signal collaborative degradation assessment system, breaking through the bottleneck of single signal detection in identifying complex faults by fusing the time-frequency features of vibration, acoustic emission, and temperature signals; establishing a dynamic weight adjustment mechanism, iteratively optimizing feature weights based on maintenance effect feedback, and overcoming the performance degradation problem of static models in long-term operation; forming a closed-loop control process of detection-maintenance-verification, quantitatively evaluating maintenance effectiveness through the degradation factor reduction rate, and realizing the self-optimization capability of the fault diagnosis system.

[0039] like Figure 2 As shown, it is a data flow diagram of the present application; as a preferred embodiment, the solution of the present application is specifically implemented as follows:

[0040] On a factory production line, a rotating piece of equipment was selected as the monitoring target. A vibration sensor, an acoustic emission sensor, and a temperature sensor were installed on the equipment to collect low-frequency vibration signals, high-frequency acoustic emission signals, and temperature time series signals, respectively. The sampling frequencies were set to 1 kHz for vibration signals, 100 kHz for acoustic emission signals, and 1 Hz for temperature signals.

[0041] The collected low-frequency vibration signal was decomposed into eight frequency bands using wavelet packets. The energy of each frequency band was calculated, and the frequency band with the highest energy value was selected as the vibration eigenvector. The high-frequency acoustic emission signal was segmented into 1ms time windows. The pulse density within each time window was calculated, and the average density of the ten time windows with the highest density values ​​was selected as the acoustic emission eigenvector. A quadratic curve was fitted to the temperature time series signal, and the absolute value of the quadratic coefficient was extracted as the absolute value of the temperature acceleration.

[0042] Data from the historical database on this type of equipment under normal operating conditions was extracted and used to train the collaborative degradation factor calculation model. The model uses a weighted summation approach, with initial weights set to 0.4 for vibration characteristics, 0.4 for acoustic emission characteristics, and 0.2 for temperature characteristics.

[0043] The model uses the currently acquired vibration eigenvectors, acoustic emission eigenvectors, and the absolute value of the temperature change acceleration to calculate the synergistic degradation factor. The warning threshold is set to 0.8. When the synergistic degradation factor exceeds 0.8, the system issues a fault alarm.

[0044] Maintenance personnel inspect and maintain the equipment according to the alarm instructions. After maintenance is complete, they re-collect signals and calculate the synergistic degradation factor. They calculate the rate of decrease in the synergistic degradation factor before and after maintenance, setting the expected rate of decrease threshold at 30%.

[0045] If the actual drop rate is less than 30%, adjust the model weights. First, reduce the weight of the acoustic emission feature by 0.05 at a time until it reaches 0.6 or the drop rate meets the requirement. If it still does not meet the requirement, reduce the weight of the vibration feature by 0.05 at a time until it reaches 0.7 or the drop rate meets the requirement.

[0046] Through the above scheme, the present application realizes the effective detection of complex faults of rotating equipment in the production line. By synchronously collecting multi-source signals and extracting features, the multi-dimensional operating status information of the equipment is captured, overcoming the limitation that single signal detection is difficult to identify complex faults. The collaborative degradation factor calculation model constructed based on historical health data realizes the quantitative assessment of the overall degradation degree of the equipment and improves the accuracy of fault warning. The introduced maintenance feedback mechanism and weight dynamic adjustment method enable the detection model to continuously adapt to changes in equipment status and working conditions, and solves the problem of decreased detection accuracy of static weight models after long-term use. Through the dual judgment of warning threshold and decline rate, closed-loop optimization of fault detection and maintenance effect verification is realized, which improves the pertinence and effectiveness of equipment maintenance. This method significantly reduces the false alarm and missed alarm rates in complex fault scenarios, effectively reduces unplanned downtime, and improves equipment reliability and production efficiency.

[0047] The present application further proposes that the calculation of the synergistic degradation factor includes the following steps:

[0048] Calculate the week-to-week change rate ΔEv of the vibration eigenvector Ev:

[0049] ΔEv=(Ev current -Ev weekago ) / Ev weekago ;

[0050] Where, Ev current is the current weighted vibration energy value, Ev weekago The vibration energy value under the same working conditions 7 days ago;

[0051] Calculate the monthly mean offset ΔDa of the acoustic emission characteristic vector Da:

[0052]

[0053] In the formula, Da current is the acoustic emission characteristic vector of the current day, is the monthly mean value of the acoustic emission eigenvector of the month;

[0054] The weighted calculation is used to obtain the synergistic degradation factor CDF:

[0055] CDF=α·ΔEv+β·ΔDa+γ·|ΔT′|;

[0056] Where |ΔT'| is the absolute value of the temperature change acceleration, and α, β, and γ are the weight values ​​configured in the collaborative degradation factor calculation model.

[0057] The calculation of the week-on-week change rate ΔEv captures short-term abnormal fluctuations in vibration energy by comparing current and historical data under the same operating conditions. The calculation of the monthly mean offset ΔDa identifies long-term trend deviations in acoustic emission characteristics by comparing daily data with the monthly mean. The weight coefficients α, β, and γ are pre-configured in the model, corresponding to the contributions of vibration, acoustic emission, and temperature characteristics, respectively.

[0058] Specifically, when calculating the week-on-week rate of change, it is necessary to ensure that the data from 7 days ago is consistent with the current operating conditions to avoid errors introduced by load or environmental differences. The calculation of the monthly mean offset must be based on the complete data of the current month, and the monthly mean must be updated using the sliding average method to reflect real-time changes. During the weighted calculation process, the weight coefficients of the vibration eigenvector and the acoustic emission eigenvector are dynamically adjusted according to the model configuration. For example, when the equipment is under high load conditions, the vibration weight coefficient can be increased to 0.4, and the acoustic emission weight coefficient can be reduced to 0.4. The absolute value of the temperature change acceleration is used as a supplementary indicator, and its weight coefficient remains fixed or fine-tuned according to preset rules. Through the above steps, the collaborative degradation factor can comprehensively characterize the degree of degradation of multiple signal features, and the configurability of the weight coefficient enhances the model's adaptability to different working conditions.

[0059] Through the above technical solution, the present application realizes the comprehensive analysis of multiple signals of rotating equipment on the production line. By calculating the week-on-week change rate of the vibration characteristic vector, the short-term change trend of the vibration state of the equipment is captured. By calculating the monthly mean offset of the acoustic emission characteristic vector, the medium-term change of the acoustic emission characteristics of the equipment is reflected. Combined with the absolute value of the temperature change acceleration, the health status of the equipment is comprehensively evaluated. The synergistic degradation factor obtained by weighted calculation comprehensively considers the information of the three dimensions of vibration, acoustic emission and temperature, which improves the accuracy and reliability of fault detection. In addition, by configuring the weight values ​​of different signals in the synergistic degradation factor calculation model, the method has good adaptability and flexibility, and can be optimized and adjusted according to the characteristics of different types of equipment.

[0060] This application further proposes:

[0061] The weight coefficients for adjusting the weighted calculation of the synergistic degradation factor include:

[0062] Prioritize reducing the weight coefficient β of the acoustic emission eigenvector, and the update formula is:

[0063] β'=max(β-η·(δ th -δ),0.4);

[0064] When β is reduced to the lower limit and still ineffective, the weight coefficient α of the vibration eigenvector is reduced, and the update formula is:

[0065] α'=max(α-η·(δ th -δ),0.7);

[0066] Where, δ is the decline rate of the synergistic degradation factor, δ th is the preset drop threshold, and η is the configurable learning rate.

[0067] The weight adjustment process adopts a phased strategy. First, the weight coefficient β of the acoustic emission eigenvector is reduced, and its weight lower limit is set to 0.4. If the reduction rate requirement is still not met after adjustment, the vibration eigenvector is reduced, and its weight lower limit is set to 0.7. The configurable learning rate η controls the weight adjustment amplitude. The learning rate η value range is limited to 0.01 to 0.1, and this parameter is dynamically optimized based on historical adjustment data. δ is introduced into the weight coefficient update formula th -δ is used as the adjustment amount to associate the weight adjustment amplitude with the deviation degree of the actual maintenance effect.

[0068] The acoustic emission eigenvector can directly reflect the microscopic energy release events inside the material (such as crack initiation, friction micro-convex fracture, etc.). These events can generate high-frequency elastic waves at the early stage of the fault, while the vibration signal may not have shown obvious amplitude changes at this time. The high-frequency characteristics of the acoustic emission eigenvector make it easy to separate it from the mechanical vibration noise in the frequency domain. Pure fault features can be effectively extracted through high-pass filtering or wavelet packet decomposition. In contrast, the vibration signal is easily interfered by the fundamental frequency and harmonics of the equipment, and early fault features are easily submerged. The vibration signal reflects the overall dynamic response of the equipment (such as imbalance, misalignment, looseness, etc.), but is not sensitive to local minor damage. The fault must develop to a certain scale to stimulate significant vibration energy. Therefore, the weight coefficient of the acoustic emission eigenvector is adjusted first, and then the weight coefficient of the vibration eigenvector is adjusted. The essence of this adjustment method is the dynamic balance of fault sensitivity:

[0069] Utilize the microscopic damage sensitivity of acoustic emission to achieve early warning;

[0070] Suppress false positives caused by oversensitivity through weight adjustment;

[0071] Under the premise of ensuring that macro faults are not missed, the system's ability to perceive minor faults is restored.

[0072] Specifically, when the decline rate of the collaborative degradation factor after the maintenance operation does not meet the standard, the reduction value of the acoustic emission feature vector weight coefficient is first calculated. The reduction value is calculated by the learning rate η and the decline rate deviation δ th -δ. For example, when the descent rate deviation is 0.1 and the learning rate η is 0.05, the weight coefficient is reduced by 0.005. If the updated acoustic emission weight reaches the lower limit of 0.4 and still cannot meet the requirements, the vibration feature vector weight adjustment is started, using the same calculation logic but setting a different lower limit of 0.7. This staged adjustment mechanism gradually optimizes the model's response sensitivity to different fault characteristics by prioritizing the adjustment of acoustic emission features that are sensitive to pulse density and then adjusting vibration features that are sensitive to energy changes. For example, in the early stage of bearing wear, prioritizing the reduction of the acoustic emission weight β can improve the detection capability of microcracks; and when the shaft misalignment intensifies, reducing the vibration weight α can more effectively reflect abnormal energy fluctuations.

[0073] For example, the configurable learning rate η is initially set to 0.05, and the preset drop threshold δ th Set to 0.3.

[0074] The weight coefficient of the current acoustic emission feature vector β is 0.2, and the decrease rate of the collaborative degradation factor CDF is 0.1. The weight coefficient of the updated acoustic emission feature vector is:

[0075] max[0.2-0.05*(0.3-0.1),0.4]=0.4.

[0076] When the weight coefficient of the acoustic emission eigenvector is reduced to the lower limit of 0.4 and is still ineffective, the weight coefficient α of the vibration eigenvector is reduced.

[0077] For example, if the weight coefficient α of the current vibration eigenvector is 0.5 and the decrease rate of the cooperative degradation factor CDF is 0.15, the weight coefficient of the updated vibration eigenvector is:

[0078] max[0.5+0.05*(0.3-0.15),0.7]=0.7.

[0079] Therefore, by dynamically adjusting the weight coefficients of acoustic emission and vibration eigenvectors, the calculation of the synergistic degradation factor is made more flexible and accurate.

[0080] Through the above technical solution, the present application can dynamically adjust the feature weights according to the actual maintenance effect and improve the calculation accuracy of the collaborative degradation factor. Specifically, by prioritizing the reduction of the weight coefficient of the acoustic emission feature vector, it is possible to more sensitively capture early minor equipment faults; when the acoustic emission characteristics are insufficient to reflect the fault, the weight of the vibration feature vector is reduced to comprehensively evaluate the equipment status. This adaptive weight adjustment mechanism overcomes the limitations of the fixed weight model, making the fault detection results more consistent with the actual equipment status and improving the accuracy and reliability of production line fault detection.

[0081] This application further proposes:

[0082] When maintenance is performed for 3 consecutive times, δ<0.5δ th When , the temperature weight correction is started, and the correction formula is:

[0083] γ'=γ-λ·|ΔT'-ΔT ref |;

[0084] Where λ is the configurable attenuation factor, ΔT ref It is the benchmark value of temperature change acceleration for similar equipment in a healthy state.

[0085] The trigger condition for temperature weight correction is that the drop rate δ does not reach the preset threshold δ after three consecutive maintenances. th Half of the temperature parameter is determined by the relationship between the number of maintenance and the descent rate. In the correction formula, the attenuation factor λ controls the weight adjustment range to prevent over-adjustment; the temperature change acceleration reference value ΔT ref The historical data from the health status of similar equipment ensures that the correction direction is aligned with the normal status of the equipment. The absolute value of the temperature change acceleration ΔT' and the reference value ΔT ref The difference reflects the degree to which the current temperature characteristic deviates from the healthy state. The larger the difference, the greater the weight adjustment.

[0086] Compared with the weight coefficient α of the vibration eigenvector and the acoustic emission eigenvector β, the temperature weight γ is adjusted after certain preconditions are met because the temperature signal has a certain lag, and the temperature change of the equipment needs to be transmitted through physical processes such as metal heat conduction and convection heat dissipation. In addition, the impact of ambient temperature fluctuations (such as day and night temperature differences and seasonal changes) on the temperature signal is far greater than that on the vibration / acoustic emission signal. Directly adjusting the weight is prone to misjudgment. At the same time, temperature anomalies are usually secondary effects of mechanical failures (such as bearing wear → increased friction → temperature rise). Adjusting the temperature weight may mask the true root cause of the fault. Therefore, the weight coefficient α of the vibration eigenvector and the acoustic emission eigenvector β are adjusted first to avoid adjusting the temperature weight γ to affect the determination of the true cause of the fault.

[0087] For example, set the configurable attenuation factor λ to 0.01. Assuming the temperature weight γ before the update is 0.2, the temperature change acceleration reference value ΔT of the same type of equipment in a healthy state is ref 0.5℃ / h 2 The current absolute value of the temperature change acceleration ΔT' is 0.8℃ / h 2 The updated temperature weight is calculated as follows:

[0088] 0.2-0.01*(0.8-0.5)=0.197.

[0089] Therefore, the temperature weight is appropriately reduced, which reduces the impact of temperature changes on the calculation of the cooperative degradation factor.

[0090] Through the above technical solution, the present application can adjust the calculation model of the collaborative degradation factor by modifying the temperature weight, even if the rate of decline of the collaborative degradation factor remains low after multiple consecutive maintenance. This dynamic adjustment mechanism improves the adaptability of the fault detection model to changes in operating conditions during long-term equipment operation, avoids misjudgments due to temperature changes, and thus improves the accuracy and reliability of production line fault detection.

[0091] This application further proposes a method for setting a configurable learning rate, including:

[0092] Count the number of successful adjustments in history N and the total number of adjustments M;

[0093] Update by formula:

[0094] η'=η0·(1+log 10 (N / M));

[0095] Where η0 is the initial learning rate and the value range of constraint η' is (0.01, 0.1).

[0096] Among them, the statistics of the number of historical adjustment successes N and the total number of adjustments M are used to quantify the historical experience of the learning rate adjustment effect; the update formula balances the impact of the historical adjustment success ratio on the learning rate by introducing a logarithmic function, avoiding sudden changes in the learning rate η due to too few adjustments; the constrained value range prevents the learning rate η from exceeding a reasonable range, ensuring the stability of the weight coefficient adjustment process.

[0097] For example, if the initial learning rate η0 is 0.05 and the historical adjustment success rate is 60%, the updated learning rate is:

[0098] 0.05*[1+lg(0.6)]=0.039, which is within the range of 0.01 to 0.1. This dynamic adjustment mechanism enables the learning rate to adaptively change based on historical adjustment results: when the success rate is high, the learning rate η is appropriately increased to accelerate convergence; when the success rate is low, the learning rate η is reduced to prevent weight coefficient oscillation. As a result, during long-term equipment operation, the learning rate η can be automatically optimized as operating conditions change, improving the efficiency and stability of weight coefficient adjustment and avoiding the unreliability caused by repeated manual adjustments.

[0099] By setting a value range, we prevent the learning rate η from being too large or too small, ensuring the stability and effectiveness of the weight adjustment process. This adaptive mechanism improves the adaptability of the collaborative degradation factor calculation model to different equipment and operating conditions, enhancing the accuracy and reliability of fault detection.

[0100] This application further proposes a method for setting a preset falling threshold, including:

[0101] Calculate the mean μ and standard deviation σ of the decline rate of the collaborative degradation factor CDF in the historical maintenance records;

[0102] The preset falling threshold δ is calculated according to the following formula:

[0103] δ=max(0.25,μ-2σ).

[0104] The decline rate data in historical maintenance records is accumulated through long-term maintenance operations and covers the actual maintenance effects under different equipment states and operating conditions. The mean decline rate μ reflects the average improvement in the coordinated degradation factor (CDF) caused by maintenance operations, while the standard deviation σ represents the degree of dispersion in the decline rate data.

[0105] δ = max(0.25, μ - 2σ) calculates the lower limit of the threshold by combining the mean μ and standard deviation σ of the descent rate to ensure that the threshold is not lower than the reference value of 0.25. This prevents the threshold from being too low due to data fluctuations and prevents the threshold from being excessively lowered when the mean descent rate is too low due to equipment aging or abnormal operating conditions.

[0106] For example, when the mean of the drop rate is 0.4 and the standard deviation is 0.1, 0.4-2*0.1=0.2, and 0.25 is taken as the final threshold.

[0107] This threshold setting method ensures that the triggering conditions for weight adjustment after maintenance are neither too loose nor too strict, balancing the timeliness and accuracy of maintenance operations.

[0108] Through the above technical solution, the present application can dynamically adjust the preset drop threshold based on historical maintenance data, improving the rationality and adaptability of threshold settings. This avoids the misjudgment that may result from artificially fixed thresholds, while also taking into account the long-term changes in the equipment's operating status, making fault detection and maintenance effect evaluation more accurate and reliable. Furthermore, by introducing statistical methods, the impact of outliers on threshold settings is reduced, improving the robustness of the system.

[0109] This application further proposes that the process of obtaining the vibration characteristic vector is as follows:

[0110] Perform wavelet packet decomposition on the low-frequency vibration signal to obtain multiple sub-band signals;

[0111] Calculate the daily energy change rate of each sub-band signal and select the sub-band with the largest change rate as the target sub-band;

[0112] The energy value of the target sub-frequency band is multiplied by a preset weight coefficient to generate the vibration feature vector.

[0113] Among them, wavelet packet decomposition uses a multi-scale analysis method to decompose low-frequency vibration signals into different frequency bands; the daily energy change rate is obtained by calculating the ratio of the current day's energy to the energy of the previous working day; the target sub-band selection is based on a dynamic sorting mechanism of the energy change rate; and the preset weight coefficient is pre-set according to the sensitivity of different frequency bands to faults in historical data.

[0114] Specifically, the low-frequency vibration signal is divided into multiple sub-band signals through wavelet packet decomposition, and each sub-band covers a specific frequency range. For each sub-band, its daily energy change rate is calculated, which reflects the degree of energy fluctuation of the equipment's operating status between adjacent working days. By comparing the daily energy change rates of all sub-bands, the sub-band with the largest change rate is selected as the target sub-band, which corresponds to the frequency feature with the most significant current degradation of the equipment. The energy value of the target sub-band is multiplied by the preset weight coefficient to generate a vibration feature vector. The weight coefficient is determined by analyzing the correlation between energy changes in different frequency bands and the occurrence of faults in historical fault data. This method enhances the ability of vibration features to characterize early degradation by dynamically selecting the frequency band with the most significant energy changes, thereby avoiding the problem of feature omission that may be caused by fixed frequency band decomposition.

[0115] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0116] The low-frequency vibration signal is decomposed by wavelet packets to obtain multiple sub-band signals. Specifically, the db4 wavelet basis is used to perform a 5-layer wavelet packet decomposition on the low-frequency vibration signal of 0-1000 Hz, obtaining 32 sub-band signals.

[0117] Calculate the daily energy change rate of each sub-band signal and select the sub-band with the largest change rate as the target sub-band. Furthermore, calculate the 24-hour average energy value for each sub-band signal and compare it with the average energy value for the same period of the previous day to obtain the daily energy change rate. Based on this, select the sub-band with the largest daily energy change rate as the target sub-band.

[0118] The vibration eigenvector is generated by multiplying the energy value of the target sub-frequency band by a preset weight coefficient. The preset weight coefficient can be set to 0.8, and the vibration eigenvector is obtained by multiplying the energy value of the target sub-frequency band by the coefficient.

[0119] Through the above technical solution, the present application can effectively extract the most sensitive frequency band features in low-frequency vibration signals and capture the vibration energy changes caused by equipment degradation. Multi-scale analysis of vibration signals is achieved through wavelet packet decomposition. Selecting the sub-band with the largest rate of change as the target sub-band can focus on the most significant degradation features. Multiplying the target sub-band energy value by the preset weight coefficient can highlight the contribution of key features. The vibration feature vector thus generated can accurately reflect the changes in the vibration state of the equipment and provide reliable input for the subsequent calculation of the collaborative degradation factor.

[0120] This application further proposes that the process of obtaining the acoustic emission feature vector is as follows:

[0121] The high-frequency acoustic emission signal is divided into multiple pulse sequences according to the time window;

[0122] Calculate the standard deviation of the pulse density in each time window and select the pulse cluster period whose standard deviation exceeds the preset threshold;

[0123] The density mean of the pulse cluster period is used as the acoustic emission feature vector.

[0124] The time window length is a configurable parameter, for example, ranging from 5 to 30 seconds, to accommodate the pulse signal periodicity of different devices. The standard deviation of pulse density is calculated by statistically analyzing the fluctuation in the number of pulses per unit time within each time window. Pulse cluster periods with standard deviations exceeding a preset threshold indicate significant density variations within that period, potentially corresponding to abnormal equipment impact events. The mean density of pulse cluster periods is calculated by removing low-density noise periods and retaining the statistical mean of high-density pulse clusters, allowing the acoustic emission feature vector to focus more on valid fault signals.

[0125] Specifically, the high-frequency acoustic emission signal is first divided into continuous time windows of fixed length, for example, every 10 seconds is a window. The number of pulses per second in each window is counted to form a pulse density sequence. The standard deviation of the sequence is calculated. If the standard deviation exceeds the preset threshold (for example, 0.8), the window is determined to be a pulse cluster period. The density mean of all selected pulse cluster periods is calculated by weighted averaging, where the weight can be dynamically adjusted according to the period length or density peak. This method effectively suppresses background noise interference by screening high-volatility periods and extracting their density mean, so that the acoustic emission feature vector can reflect the sudden high-frequency energy release caused by faults such as wear and cracks inside the equipment, thereby improving the detection sensitivity and anti-interference ability of the collaborative degradation factor calculation model.

[0126] Through the above technical solution, this application can effectively identify transient abnormal pulse clusters in high-frequency acoustic emission signals, eliminate conventional steady-state noise interference, and accurately capture the sudden stress wave signals released by mechanical damage to the equipment. Through dynamic segmentation and statistical screening mechanisms, this method enhances the sensitivity of acoustic emission feature vectors to local degradation characteristics, thereby improving the ability of collaborative degradation factors to characterize complex faults and avoiding feature extraction bias caused by uneven distribution of pulse signals.

[0127] This application further proposes that the process of extracting the absolute value of the temperature change acceleration is:

[0128] Fitting a quadratic curve of the temperature time series signal in hours;

[0129] The quadratic term coefficient of the quadratic curve is extracted as the temperature change acceleration value, and the absolute value of the temperature change acceleration is obtained.

[0130] As a preferred embodiment, the solution of the present application is specifically implemented as follows: in the temperature time series signal processing process, the continuously collected production line equipment temperature data is divided into processing units by hour. In each processing unit, the least squares method is used to fit the temperature data points to a quadratic polynomial, and the form is obtained: T(t) = at 2 The curve equation of +bt+c. The absolute value of the quadratic coefficient a is extracted as the temperature change acceleration value. When the temperature time series shows an accelerated upward or downward trend, the absolute value of the coefficient will increase significantly. For example, in the case of abnormal bearing friction, the absolute value of the quadratic coefficient of the temperature curve may increase from 0.03℃ / h under normal operating conditions. 2 Sudden increase to 0.15℃ / h 2 The absolute value of the temperature change acceleration generated thereby is input into the collaborative degradation factor calculation model to participate in the weighted operation.

[0131] Through the above technical solution, this application effectively solves the technical defects of traditional temperature monitoring methods that only focus on the absolute value of temperature and ignore the dynamic characteristics of the change trend. Through the mathematical representation of the quadratic coefficient, it is possible to accurately capture the acceleration characteristics of temperature changes and avoid misjudgments caused by slow fluctuations in ambient temperature. This technical means makes the contribution of the temperature signal in the collaborative degradation factor form a strong correlation with the actual degree of degradation of the equipment, especially in the detection of thermal effect anomalies in the early fault stage, significantly improving the characterization ability of complex fault characteristics.

[0132] This application further proposes a method for constructing a collaborative degradation factor calculation model based on historical data, comprising the following steps:

[0133] Obtain historical data on the health status of multiple devices of the same type;

[0134] Extract vibration eigenvectors, acoustic emission eigenvectors, and absolute values ​​of temperature change acceleration;

[0135] The weight coefficients of vibration eigenvector, acoustic emission eigenvector and absolute value of temperature change acceleration in the model are optimized by gradient descent method until the weighted synergistic degradation factor approaches 0.

[0136] Among them, when obtaining historical data, it is necessary to cover the full life cycle operation data of multiple devices in a healthy state to ensure data diversity; the extraction of feature vectors must adopt an algorithm consistent with real-time detection to ensure the consistency of model input features; the gradient descent method uses the minimization of the collaborative degradation factor as the objective function, and through iterative adjustment of the weight coefficient, the model output approaches the baseline value of the healthy state.

[0137] Specifically, vibration feature vectors, acoustic emission feature vectors, and absolute values ​​of temperature change acceleration are first extracted from the historical operating data of multiple devices of the same type as model training samples. Subsequently, the weight coefficients of each feature are initialized, and the error between the output value of the collaborative degradation factor under the current weight and the healthy state baseline value is calculated through the gradient descent algorithm, and the weight coefficient is updated based on the error backpropagation. For example, if the initial value of the weight coefficient of the vibration feature vector is 0.3, it may be adjusted to 0.28 during the gradient descent process to reduce the error. This process continues to iterate until the weighted calculation result of the collaborative degradation factor stably approaches 0, indicating that the model has learned the optimal weight combination of each feature under the healthy state. The model established in this way can dynamically adapt to the differences in feature distribution of different devices and improve long-term detection accuracy.

[0138] As a preferred embodiment, the solution of this application is implemented as follows: In an industrial motor health monitoring scenario, historical operating data from multiple motors of the same model is collected, including vibration signals, acoustic emission signals, and temperature time series signals. First, the vibration signal is decomposed using wavelet packets, and the target sub-band with the highest daily energy change rate is extracted as the vibration feature vector. The acoustic emission signal is segmented into time windows, and periods where the pulse density standard deviation exceeds a preset threshold are selected. The density mean is calculated as the acoustic emission feature vector. A quadratic curve is fitted to the temperature signal, and the absolute value of the quadratic term coefficient is extracted as the temperature change acceleration feature. A synergistic degradation factor calculation model is then constructed with an initial weight coefficient of 0.33. Using the health data as a benchmark, the weight coefficients are iteratively adjusted using gradient descent. Each iteration, the mean squared error (MSE) between the synergistic degradation factor at the current weight and zero is calculated, and the weight coefficients are updated using backpropagation. The optimization is terminated when the MSE for 100 consecutive iterations is less than 0.001 and the change in the weight coefficients is less than 0.005. Finally, a stable model with a vibration weight of 0.28, an acoustic emission weight of 0.51, and a temperature weight of 0.21 is obtained.

[0139] Through the above technical solution, this application effectively solves the defect that the traditional static weight model cannot adapt to the long-term operating characteristics changes of the equipment. Through the gradient optimization process of historical health data, the collaborative degradation factor is stably close to the zero-value benchmark under normal operating conditions of the equipment, which significantly improves the robustness of multi-source signal fusion detection, avoids false alarms caused by operating condition fluctuations or sensor drift, and provides an adaptive weight configuration mechanism for individual differences of different devices.

[0140] The technical scope of the present invention is not limited to the contents of the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A production line fault detection method based on regional attention learning mechanism, characterized by: The steps include: Synchronously collect low-frequency vibration signals, high-frequency acoustic emission signals, and temperature timing signals of rotating equipment on the production line; Performing frequency band decomposition on the low-frequency vibration signal to obtain a vibration eigenvector; performing pulse density analysis on the high-frequency acoustic emission signal to obtain an acoustic emission eigenvector; Performing data processing on the temperature time series signal to extract the absolute value of the temperature change acceleration; Obtain historical data of similar equipment in a healthy state, and construct a collaborative degradation factor calculation model based on the historical data; Inputting the vibration eigenvector, the acoustic emission eigenvector, and the absolute value of the temperature change acceleration into the collaborative degradation factor calculation model to obtain a collaborative degradation factor; Determining whether the collaborative degradation factor exceeds a preset warning threshold, and if so, triggering a fault alarm instruction; Obtaining low-frequency vibration signals, high-frequency acoustic emission signals, and temperature time series data after maintenance is performed by staff according to the fault alarm instructions; recalculating the synergistic degradation factor based on the reacquired low-frequency vibration signal, the high-frequency acoustic emission signal, and the temperature time series data, and calculating the decrease rate of the synergistic degradation factor before and after maintenance; Determine whether the decrease rate of the synergistic degradation factor is lower than a preset decrease threshold. If so, adjust the model parameters of the synergistic degradation factor calculation model, and recalculate the decrease rate of the synergistic degradation factor based on the adjusted model parameters until the decrease rate exceeds the preset decrease threshold, thereby triggering a model optimization completion instruction.

2. The production line fault detection method based on regional attention learning mechanism according to claim 1 is characterized in that: The synergistic degradation factor calculation model obtains the synergistic degradation factor including: Calculate the week-to-week change rate ΔEv of the vibration eigenvector Ev: Get the current weighted vibration energy value Ev current , Vibration energy value Ev under the same working condition 7 days ago weekago ; Weekly change rate ΔEv=(Ev current -Ev weekago ) / Ev weekago ; Calculate the monthly mean offset ΔDa of the acoustic emission characteristic vector Da: Get the acoustic emission characteristic vector Da of the current day current , the monthly mean of the acoustic emission characteristic vector Monthly mean deviation The weighted calculation is used to obtain the synergistic degradation factor CDF: CDF=α·ΔEv+β·ΔDa+γ·|ΔT'|; Where |ΔT'| is the absolute value of the temperature change acceleration, and α, β, and γ are the weight values ​​configured in the collaborative degradation factor calculation model.

3. The production line fault detection method based on regional attention learning mechanism according to claim 2 is characterized in that: The weight coefficients for adjusting the weighted calculation of the synergistic degradation factor include: Prioritize reducing the weight coefficient β of the acoustic emission eigenvector, and the update formula is: β'=max(β-η·(δ th -d),0.4); When β is reduced to the lower limit and still ineffective, the weight coefficient α of the vibration eigenvector is reduced, and the update formula is: α'=max(α-η·(δ th -d),0.7); Where, δ is the decline rate of the synergistic degradation factor, δ th is the preset drop threshold, and η is the configurable learning rate.

4. The production line fault detection method based on regional attention learning mechanism according to claim 3 is characterized by: The weight coefficient for adjusting the weighted calculation of the synergistic degradation factor also includes: When maintenance is performed for 3 consecutive times, δ<0.5δ th When , the temperature weight correction is started, and the correction formula is: γ'=γ-λ·|ΔT'-ΔT ref |; Where λ is the configurable attenuation factor, ΔT ref It is the benchmark value of temperature change acceleration for similar equipment in a healthy state.

5. The production line fault detection method based on regional attention learning mechanism according to claim 3 is characterized by: The setting of the learning rate η includes: Count the number of successful adjustments in history N and the total number of adjustments M; Update by formula: η'=η0·(1+log 10 (N / M); Where η0 is the initial learning rate and the value range of constraint η' is (0.01, 0.1).

6. The production line fault detection method based on regional attention learning mechanism according to claim 1 is characterized in that: The setting of the preset falling threshold includes: Calculate the mean μ and standard deviation σ of the decline rate of the collaborative degradation factor CDF in the historical maintenance records; The preset falling threshold δ is calculated according to the following formula: δ=max(0.25,μ-2σ).

7. The production line fault detection method based on regional attention learning mechanism according to claim 1 is characterized in that: The process of obtaining the vibration characteristic vector is as follows: Perform wavelet packet decomposition on the low-frequency vibration signal to obtain multiple sub-band signals; Calculate the daily energy change rate of each sub-band signal and select the sub-band with the largest change rate as the target sub-band; The energy value of the target sub-frequency band is multiplied by a preset weight coefficient to generate the vibration feature vector.

8. The production line fault detection method based on regional attention learning mechanism according to claim 1 is characterized in that: The acquisition process of the acoustic emission feature vector is as follows: The high-frequency acoustic emission signal is divided into multiple pulse sequences according to the time window; Calculate the standard deviation of the pulse density in each time window and select the pulse cluster period whose standard deviation exceeds the preset threshold; The density mean of the pulse cluster period is used as the acoustic emission feature vector.

9. The production line fault detection method based on regional attention learning mechanism according to claim 1, characterized in that: The process of extracting the absolute value of the temperature change acceleration is as follows: Fitting a quadratic curve of the temperature time series signal in hours; The quadratic term coefficient of the quadratic curve is extracted as the temperature change acceleration value, and the absolute value of the temperature change acceleration is obtained.

10. The production line fault detection method based on regional attention learning mechanism according to claim 1, characterized in that: The construction of the collaborative degradation factor calculation model based on historical data includes: Obtain historical data on the health status of multiple devices of the same type; Extract vibration eigenvectors, acoustic emission eigenvectors, and absolute values ​​of temperature change acceleration; The weight coefficients of vibration eigenvector, acoustic emission eigenvector and absolute value of temperature change acceleration in the model are optimized by gradient descent method until the weighted synergistic degradation factor approaches 0.

Images