A machine learning-based fault prediction method for automobile motor production equipment

Abstract

The application relates to the technical field of industrial data processing, and specifically discloses a fault prediction method for automobile motor production equipment based on machine learning, which comprises the following steps: synchronously acquiring vibration data of a main bearing and working condition data of a main shaft motor, calculating a learning access factor based on the deviation degree of the working condition data from a preset standard working condition, weighting and gating a basic drift learning rate of a dynamic baseline model by using the learning access factor, and then updating a baseline mean value and a slow-changing volatility rate; calculating an instantaneous volatility rate based on the instantaneous energy deviation of the vibration data and the updated baseline mean value, determining a working condition mutation index as the ratio of the instantaneous volatility rate to the slow-changing volatility rate, identifying a current working condition type and acquiring a corresponding adaptive threshold value, and determining that a fault occurs when the working condition mutation index exceeds the adaptive threshold value. The method can simultaneously realize immunity to transient working conditions and sensitive detection of weak mutations in a steady state, and greatly improves the accuracy and robustness of fault prediction.

Description

Technical Field

[0001] This invention relates to the field of industrial data processing technology. Specifically, it relates to a fault prediction method for automotive motor production equipment based on machine learning. Background Technology

[0002] In automated production lines for automotive motors, high-speed winding machines are core equipment for achieving precise stator winding. Their main bearings withstand complex radial and axial forces at ultra-high speeds for extended periods, making them critical and vulnerable components. Pitting, peeling, or breakage of the bearings can lead to abnormal winding tension, coil damage, and even spindle seizure, causing unplanned production line downtime and severely impacting production efficiency and product quality. Therefore, real-time monitoring and fault prediction of the winding machine's main bearings are crucial for predictive maintenance and ensuring stable production line operation.

[0003] Vibration sensors are widely used in the industry to monitor the health of bearings. The degradation process of bearings manifests in two typical modes in vibration signals: one is a slow and continuous rise in the vibration energy baseline caused by normal wear, a phenomenon commonly known as drift; the other is a sudden and violent jump in the vibration signal caused by severe defects such as pitting and spalling on the bearing surface, a phenomenon commonly known as abrupt change.

[0004] In the prior art, a common fault prediction method is to use a dual-window exponential smoothing algorithm. This method sets a slow-responding long-time window to fit the drift baseline of the vibration signal, and sets a sensitive short-time window to capture the real-time changes in vibration. When the absolute difference between the output values ​​of the two windows suddenly surges and exceeds a pre-set fixed threshold, it is determined that a sudden fault has occurred.

[0005] However, the technical problem with this method is that the smoothing coefficient and fault judgment threshold used are fixed and do not take into account the influence of changes in operating conditions. This method uses the same smoothing parameters and the same judgment threshold to process all collected vibration data indiscriminately. High-speed winding machines are devices with complex and variable operating conditions. Their operation process includes at least a stopped state, a starting and accelerating state, a full-load constant speed state, and a deceleration and stopping state. During the starting and accelerating state and the deceleration and stopping state, the violent jumps in the vibration signal are normal transient responses and not faults. However, during the full-load constant speed state, the vibration signal should be highly stable. At this time, any slight fluctuation may be a sign of a fault.

[0006] If the threshold is set too low in order to capture the slight changes in the constant speed state, then in the startup acceleration state, normal transient fluctuations will frequently exceed the threshold, generating a large number of false alarms. Conversely, if the threshold is forced to be set very high in order to avoid false alarms in the startup acceleration state, then in the full-load constant speed state, those early change signals with small amplitudes but extremely fatal effects will be ignored because they are below this high threshold, resulting in missed fault detection.

[0007] Therefore, existing technologies cannot accurately and adaptively distinguish between normal operating condition fluctuations and real fault mutations under complex and ever-changing operating conditions, resulting in low robustness and accuracy of predictions. Summary of the Invention

[0008] To address the issues that using a dual-window exponential smoothing algorithm with a fixed threshold cannot simultaneously account for both false alarm and false negative rates under multiple operating conditions, and cannot adaptively distinguish between fluctuations in normal operating conditions and sudden fault changes, this invention proposes a fault prediction method for automotive motor production equipment based on machine learning, comprising:

[0009] Take any point in time as the current time, and obtain the vibration data of the main bearing of the winding machine and the operating condition data of the main spindle motor at the current time;

[0010] Based on the deviation between the current operating condition data and the pre-acquired standard operating condition data, the learning admission factor at the current moment is calculated. A dynamic drift baseline model is constructed based on the learning admission factor and vibration data to output the vibration benchmark value and the slow-varying volatility benchmark value at the current moment.

[0011] Based on the instantaneous energy deviation between the vibration data at the current moment and the dynamic vibration benchmark value, the instantaneous volatility at the current moment is calculated, and the ratio of the instantaneous volatility to the slowly varying volatility benchmark value is used as the operating condition change index at the current moment.

[0012] The operating condition type is determined based on the current operating condition data, and an adaptive threshold for the operating condition mutation index corresponding to the operating condition type is obtained. When the operating condition mutation index is greater than the adaptive threshold, the winding machine is judged to have a fault, so as to realize the fault prediction of automotive motor production equipment.

[0013] This technical solution integrates mechanical vibration and electrical operating condition data to establish a multi-dimensional perception space. The process of calculating and learning the admission factor is essentially assessing the credibility of the current operating condition. It acts like a smart valve, physically isolating the baseline model from transient shocks under unsteady conditions, ensuring that the model updates its memory only when the equipment is in a thermodynamically and kinetically stable state. Subsequently, the operating condition mutation index, constructed by comparing the instantaneous energy with the slowly varying baseline energy, is essentially a self-normalized signal-to-noise ratio index. It eliminates the influence of the equipment's basic energy level, making fault determination no longer dependent on absolute amplitude, but on the degree of mutation in the signal structure. Finally, in conjunction with the operating condition adaptive threshold, it tightens in steady state to capture weak defects and relaxes in transient state to tolerate normal fluctuations, thereby accurately predicting faults in automotive motor production equipment.

[0014] Preferably, the current operating condition data and the standard operating condition data include: the current operating condition data includes the spindle motor speed and load current at the current moment; the standard operating condition data includes the spindle motor rated speed and rated load current.

[0015] Preferably, the learning admission factor at the current moment is determined by the following relationship:

[0016] ;in, for Learning access factors at all times It is a natural exponential function. The preset sensitivity coefficient, for Rotation speed at any given moment for Load current at any given time and These are the rated speed and rated load current of the spindle motor, respectively.

[0017] This technical solution employs attenuation logic based on Gaussian radial basis functions to map the multivariate coupling deviation of rotational speed and current into a continuous probability space. Through this nonlinear gating mechanism, the model jump caused by traditional fixed threshold cutting is avoided, thereby strictly ensuring the purity of the data input into the baseline model, so that the model only records the effective equipment aging trend rather than operational interference.

[0018] Preferably, a dynamic drift baseline model is constructed based on the learning admission factor and vibration data to output the current vibration benchmark value and the slow-varying volatility benchmark value, which is based on the following relationship:

[0019] ;

[0020] ;

[0021] in, for Vibration reference value at time, for Vibration reference value at time, for The benchmark value of slow-varying volatility at time t. for The benchmark value of slow-varying volatility at time t. The preset base drift learning rate, for Learning access factors at all times for Vibration data at any given moment.

[0022] This technical solution constructs a variable-weight recursive filter with dynamic inertia. When the device is in a steady state, the model reduces inertia and sensitively tracks the natural physical drift caused by temperature changes. When the device experiences drastic changes in operating conditions, the model freezes the baseline at the previous healthy state. This design decouples normal slow drift from normal transient shocks, solving the problem that traditional filters cannot distinguish between the two, resulting in the baseline being biased by noise.

[0023] Preferably, another way to determine the sensitivity coefficient is through a data-driven self-calibration method: during the calibration phase when the winding machine is running at full load and constant speed, a continuous speed sequence and load current sequence are collected; the first variance of the normalized deviation of the speed sequence relative to the rated speed and the second variance of the normalized deviation of the load current sequence relative to the rated load current are calculated respectively; based on the statistical characteristics of Gaussian distribution, the reciprocal of twice the sum of the first and second variances is taken as the sensitivity coefficient.

[0024] This technical solution enables data-driven adaptive configuration of model hyperparameters. It automatically calibrates the bandwidth of the Gaussian admission window based on the unique noise level of each device, ensuring that both high-precision new devices and older devices with some wear and tear can obtain the optimal learning threshold that matches their own stability. This avoids the risk of the model failing to update due to excessively high sensitivity or absorbing noise due to excessively low sensitivity.

[0025] Preferably, the time required for the main bearing of the winding machine to reach a thermally stable state from a cold start, and the sampling frequency of the vibration data are obtained. Based on the definition of the time constant of exponential smoothing, the reciprocal of the product of the time and the sampling frequency is used as the basic drift learning rate.

[0026] Preferably, the instantaneous volatility at the current moment is determined by the following relationship:

[0027] ;in, for Instantaneous volatility at time t, for Instantaneous volatility at time t, The preset instantaneous response coefficient, for Vibration data at any given time for The vibration reference value at any given time.

[0028] This technical solution constructs a dimensionless relative energy mutation rate. This ratio-type index eliminates the influence of the difference in absolute vibration amplitude of the equipment under different loads, so that the fault characteristics are normalized and amplified. Regardless of whether the equipment is in the low-speed or high-speed operating range, as long as the internal structure is intact, the ratio will always approach 1. Once structural damage occurs, the instantaneous energy ratio relative to the background will be unbalanced, and the index will immediately surge, thereby achieving high signal-to-noise ratio extraction of fault signals.

[0029] Preferably, the method for determining the operating condition type at the current moment based on the current speed and load current is as follows: the operating condition type is preset, including the stop state, the start-up acceleration state, the full-load constant speed state, and the deceleration stop state, and the speed and load current at each moment under each operating condition type are used to form an operating condition feature vector to form a feature vector library for each operating condition type;

[0030] The current speed and load current are combined to form the current operating condition feature vector. The current operating condition feature vector is then matched with the feature vector library for each operating condition type. Based on the principle of maximum similarity, the operating condition type at the current moment is determined.

[0031] Preferably, the vibration data of the main bearing of the winding machine at the current moment, as well as the speed and load current of the main spindle motor at the current moment, are obtained in real time by accessing the servo driver or programmable logic controller of the winding machine through the communication interface of the winding machine, and a unified timestamp is configured for the vibration data, speed and load current.

[0032] Preferably, the adaptive threshold of the working condition mutation index corresponding to the working condition type is determined in the following way: when the winding machine is in a healthy operating state, the working condition mutation index sequence under each working condition type is collected in advance, the mean and standard deviation of the working condition mutation index sequence are calculated, and based on the statistical principle of three standard deviations, the mean plus three standard deviations is set as the adaptive threshold of the working condition mutation index corresponding to that working condition type.

[0033] The present invention has the following effects:

[0034] This solution protects the baseline model under dynamic operating conditions by introducing a learning admission factor. At the same time, it constructs an operating condition mutation index using the ratio of instantaneous volatility to slow-varying volatility, and, in conjunction with an adaptive threshold, can simultaneously achieve immunity to transient operating conditions and sensitive detection of weak mutations under steady state, thereby improving the accuracy and robustness of fault prediction for automotive motor production equipment. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0036] Figure 2 This is a comparison diagram of the effects of the present invention and the prior art. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0038] This invention provides a fault prediction method and system for automotive motor production equipment based on machine learning, such as... Figure 1 As shown, it includes:

[0039] S1: Obtain the vibration data of the main bearing of the winding machine at the current moment and the operating condition data of the main spindle motor at the current moment.

[0040] The core objective of this step is to establish the spatiotemporal correlation between mechanical vibration and electrical operating conditions. The spindle speed of a high-speed winding machine changes extremely rapidly (e.g., accelerating from 0 to 3000 rpm takes only a few hundred milliseconds). If asynchronous acquisition is used, assuming the vibration data lags behind the current data by 50 ms, the system may incorrectly match the high vibration caused by the acceleration phase to the current signal that has already entered the constant-speed phase, thus misjudging it as abnormal vibration under constant-speed conditions. Therefore, this step employs hardware-level PTP synchronization, the physical basis of which is to eliminate signal transmission delay jitter and ensure that subsequent algorithms process the equipment state at the same physical instant.

[0041] Specifically, when the winding machine is running, the real-time rotational speed (unit: ) of the servo drive's internal register is read via the winding machine's communication interface using an industrial Ethernet (such as Profinet or EtherCAT) at a 1ms cycle. ) and torque current (converted to load current, unit: Simultaneously, vibration acceleration signals (unit: ) are collected by an IEPE accelerometer mounted on the main bearing housing. The data is converted to digital values ​​using a high-speed acquisition card. The acquisition card and PLC are synchronized via PTP (Precise Time Protocol) to ensure that each vibration data point corresponds to a unique speed and load current, thereby achieving timestamp alignment. This eliminates the bias of multi-source heterogeneous data and lays a solid data foundation for the subsequent establishment of an accurate working condition-vibration correlation model.

[0042] by The current moment is used to acquire the vibration data of the winding machine's main bearing and the rotational speed of the main spindle motor at the current moment. and load current The operating conditions can be reflected at the current moment through speed and load current. Combining the operating condition data at any time Operating condition feature vector at time step This vector will serve as the core basis for determining the device context, i.e., the current operating condition, in subsequent steps.

[0043] Thus, by synchronously collecting vibration and operating condition data, the necessary data foundation is provided for subsequent operating condition identification and adaptive modeling.

[0044] S2: Calculate the learning admission factor for the current moment based on the degree of deviation between the current working condition data and the pre-acquired standard working condition data.

[0045] From a dynamic perspective, when a motor accelerates, decelerates, or experiences drastic load changes, the impact of mechanical clearances and the fluctuation of current are normal physical responses, not fault characteristics. Incorporating these data into the baseline model would contaminate the baseline and thus artificially inflate it.

[0046] This step aims to construct a dynamic probabilistic baseline model, also known as a slow window, that can adaptively track the normal drift of the equipment. The goal of this model is to output the normal vibration mean and normal slowly varying volatility of the equipment. The key point is that the model is only allowed to be updated under specific operating conditions. In order to solve the problem that transient operating conditions such as startup acceleration will contaminate the normal drift, a learning admission factor is designed. The role of this factor is to allow the model to learn only when the equipment is under full load and constant speed operating conditions that best reflect its true wear state.

[0047] Specifically, it is defined by a Gaussian radial basis function to assess the degree to which the current operating condition deviates from the ideal steady-state condition. Physically, it simulates a confidence gate: the current data is considered representative only when the equipment operating condition is thermodynamically and kinetically very close to the rated steady state, allowing it to pass through the gate to update the equipment's health baseline.

[0048] In one embodiment, the rated speed and rated load current of the spindle motor are obtained in advance based on its industrial parameters. Furthermore, the learning admission factor at the current moment is determined by the following relationship:

[0049]

[0050] in, for Learning access factors at all times It is a natural exponential function. The preset sensitivity coefficient is a preset constant used to control the sharpness of factor decay. The larger the value, the more stringent the operating condition matching requirements; it is usually set to [value missing]. , for Rotation speed at any given moment for Load current at any given time and These are the rated speed and rated load current of the spindle motor, respectively.

[0051] in, Indicates the speed deviation term. Indicates the current deviation term, when and At that time, that is, under full load and constant speed conditions, both are close to , The exponent term within () is close to , making At this point, the model is allowed to learn and drift at maximum speed; when In a stopped state, or When the operating conditions are changing drastically, such as acceleration / deceleration, keep away This leads to a large deviation in rotational speed. The exponent term within the parentheses is a very large negative number, making At this point, the model is prohibited from learning and is frozen.

[0052] This provides information on learning access factors. A simple example of calculation:

[0053] Assuming the preset parameters are: rated operating speed Rated operating load current Sensitivity coefficient .

[0054] Scenario 1: During full-load, constant-speed operation, the current rotational speed... , Speed ​​deviation: Current deviation term: , ,at this time A value close to 1 indicates that learning is permitted.

[0055] Scenario 2, Acceleration State: Current Rotational Speed Far below the rated speed Speed ​​deviation: Current deviation term: , ,at this time A value close to 0 indicates that learning is prohibited and the model is frozen.

[0056] The reason for adopting this implementation method is that learning the admission factor is the key to ensuring the purity of the model. A continuous, non-linear variable is needed to quantify how stable the current working condition is. Traditional threshold switching methods will cause the model to jump at the critical point, while the decay formula based on the Gaussian radial basis function can smoothly describe the degree of deviation of the working condition.

[0057] By learning the access factor, the interference of non-steady-state data on the baseline model can be automatically shielded. When the equipment operating conditions are unstable, the learning access factor is automatically turned off and approaches 0 to prevent the model from learning incorrect normal noise. When the operating conditions return to the rated state, the learning access factor is automatically turned on, allowing the model to capture the natural aging drift of the bearing.

[0058] In one embodiment, another way to determine the sensitivity coefficient is through a data-driven self-calibration method: during the calibration phase when the winding machine is running at full load and constant speed, a continuous speed sequence and load current sequence are collected; the first variance of the normalized deviation of the speed sequence relative to the rated speed and the second variance of the normalized deviation of the load current sequence relative to the rated load current are calculated respectively; based on the statistical characteristics of Gaussian distribution, the reciprocal of twice the sum of the first and second variances is taken as the sensitivity coefficient.

[0059] Specifically: During the equipment commissioning period, the machine was run idle at 3000 rpm for 1 minute. The normalized variance of the speed fluctuation was calculated to be 0.001, and the normalized variance of the current fluctuation was 0.002. Therefore, the total variance was 0.003, and the sensitivity coefficient was... .

[0060] The reason for adopting this implementation method is that devices of different brands and with different degrees of aging exhibit varying natural fluctuation amplitudes under steady-state conditions. A fixed coefficient might prevent some high-precision devices from ever achieving the desired level of accuracy. Or low-precision equipment in an unsteady state It's still very big.

[0061] S3: Construct a dynamic drift baseline model based on the learning admission factor and vibration data to output the current vibration benchmark value and slow volatility benchmark value.

[0062] The health baseline of mechanical equipment is not a horizontal line, but a curve that changes slowly with temperature and lubrication conditions. If the baseline is not updated, it will falsely report thermal drift; if the baseline is updated too quickly, it will also learn fault signals, resulting in missed reports.

[0063] The vibration amplitude of a bearing will naturally drift as the running time increases due to thermal expansion and changes in lubricating oil viscosity. This is normal. However, fault shocks or start-stop shocks are transient changes. By learning the admission factor to adjust the inertia of the filter, in steady state, the inertia is reduced to track thermal drift; in non-steady state, the inertia is increased to freeze the memory and prevent start-stop shocks from being mistaken for thermal drift.

[0064] In one embodiment, a dynamic drift baseline model is constructed based on the learning admission factor and vibration data to output the current vibration benchmark value and the slow-varying volatility benchmark value, based on the following relationship:

[0065] ;

[0066] ;

[0067] in, for Vibration reference value at time, for Vibration reference value at time, for The benchmark value of slow-varying volatility at time t. for The benchmark value of slow-varying volatility at time t. The preset base drift learning rate, for Learning access factors at all times for Vibration data at any given moment.

[0068] This is a variable-weighted exponentially weighted moving average process, only when... Under full-load constant-speed operating conditions, the composite learning rate Only if the value is not zero will the model adapt to the new vibration data. Slow updates and To track normal drift; in all other operating conditions Composite learning rate The above formula degenerates into and This means that the model parameters are frozen, the model retains the state of the previous moment, and does not learn any transient noise, thus achieving automatic immunity to noise from non-steady-state operating conditions.

[0069] Here is a simple example of building a dynamic drift baseline model based on a learning admission factor and vibration data to output the current vibration benchmark value and the slow-varying volatility benchmark value:

[0070] Assume the preset parameters are: base drift learning rate , The vibration reference value at time t is: V, V, using As a gating mechanism for the learning rate, for and baseline slow volatility Perform exponential smoothing updates.

[0071] Case 1, under full load and uniform speed, Current vibration Normal fluctuations, compound learning rate ;

[0072] V;

[0073] ;

[0074] As can be seen, the model parameters are based on the new vibration data. Updates have been slow.

[0075] Scenario 2, when starting the accelerated state, Vibration at this time Transient and drastic fluctuations, the compound learning rate equals ;

[0076] V;

[0077] ;

[0078] It is evident that, despite High volatility occurred, but due to The learning rate is close to 0, and the compound learning rate is extremely small, resulting in and Almost equal to The value at time, ; The model is frozen, thus achieving immunity to transient noise.

[0079] Thus, utilizing As a gating switch, it achieves the intelligent characteristics of steady-state tracking drift and transient memory retention, reducing the false alarm rate. Through the gating effect of learning the admission factor, the dynamic baseline model is only updated under full-load constant speed conditions, realizing automatic immunity to transient noise such as start-up and stop, and ensuring the accuracy of baseline mean and baseline slow fluctuation rate.

[0080] In one embodiment, another way to determine the baseline drift learning rate is to obtain the time required for the main bearing of the winding machine to reach a thermally stable state from a cold start, and the sampling frequency of the vibration data. Based on the definition of an exponentially smoothed time constant, the reciprocal of the product of this time and the sampling frequency is used as the baseline drift learning rate. This is because the learning rate determines the memory length of the model. If the learning rate is too large, the model will be too sensitive to following noise; if it is too small, the model will lag behind the thermal drift. This method scientifically determines the hyperparameters from a physical thermodynamic perspective, ensuring that the evolution speed of the baseline model is consistent with the physical thermal inertia of the equipment.

[0081] Specifically, it was measured that the main bearing required 30 minutes (1800 seconds) to reach thermal equilibrium from 20°C to 60°C. If the sampling frequency is 1000Hz, the total number of sampling points is... ,set up .

[0082] S4: Calculate the instantaneous volatility based on the instantaneous energy deviation between the vibration data at the current moment and the dynamic vibration benchmark value, and determine the operating condition mutation index based on the instantaneous volatility and the slowly varying volatility benchmark value.

[0083] This step aims to construct a dimensionless fault index. Physically, motors with different power or loads have huge differences in absolute vibration amplitude, which cannot be measured by a unified standard. The operating condition mutation index is essentially a real-time signal-to-noise ratio. It compares the energy at the current moment with the historical background energy, thereby being able to keenly capture those weak impacts that are submerged in strong background noise, such as the vibration caused by early pitting corrosion.

[0084] In one embodiment, the instantaneous volatility at the current moment is determined by the following relationship:

[0085]

[0086] in, for Instantaneous volatility at time t, for Instantaneous volatility at time t, The preset instantaneous response coefficient, for Vibration data at any given time for Vibration reference value at time, instantaneous response coefficient The ratio is usually set. Much bigger This means that the model has almost no inertia and can respond to current changes instantaneously, thereby accurately capturing impact signals with extremely short durations and avoiding missed detections due to excessive smoothing.

[0087] Here, through smoothing Relative to its dynamic baseline The instantaneous energy deviation measures not the absolute value of the signal, but rather the magnitude of the signal's energy deviation from its dynamic normal baseline. This is a detrended instantaneous energy extraction process, unlike... Focusing on long-term trends, this formula first utilizes a dynamic benchmark value. Will The thermal drift component is removed, retaining only the pure high-frequency vibrational energy. Subsequently, by introducing an instantaneous response coefficient, it is possible to... Instead of deeply smoothing historical data, it jumps instantaneously to follow the energy fluctuations of the current moment, thus enabling it to sensitively capture weak and brief impact signals caused by mechanical damage that are submerged in background noise on a millisecond timescale.

[0088] This provides information on instantaneous volatility. Example of calculation: The old value at time ,exist A sudden fault occurred at a certain moment, and the vibration data... Instantaneous jump to At this time, the equipment is in a constant speed state. Still within the normal range V;

[0089] The instantaneous energy deviation is: ;

[0090] ;

[0091] visible, The output value jumped drastically from 0.6 to 13.06 instantaneously, sensitively capturing this mutation.

[0092] The formula for calculating the operating condition mutation index is as follows:

[0093]

[0094] in, for The index of sudden changes in operating conditions at any given time. for Instantaneous volatility at time t, for The benchmark value of slow-varying volatility at time t. The preset volatility base is typically set to a very small positive number, 0.1, to prevent the denominator from fluctuating during cold starts or when the equipment is stationary. Approaching 0 leads to computational overflow.

[0095] The reason for adopting this implementation method is that, This represents the long-term background noise level, while This represents the energy level at the current instant, under normal operating conditions (including steady-state drift). , and They are all slowly tracking the signal. exist Nearby fluctuations, their fluctuation energy Approximately equal to ,because Also tracking , Following Change, therefore The ratio of the two It will naturally stabilize around 1, eventually leading to At the moment of sudden failure, , Instantaneous change, resulting in Instantaneous maximum It will immediately follow this instantaneous jump and increase significantly, thus realizing a function to track a rapidly changing window. Because the base drift learning rate is extremely small and almost constant, therefore This achieves normalized amplification of the fault signal.

[0096] This ratio calculation method essentially constructs a self-normalized signal-to-noise ratio index. Regardless of whether the equipment is under light or heavy load, as long as the current fluctuation does not change abruptly relative to the historical background under that state, the index remains around 1, thereby eliminating the influence of operating condition differences on fault determination and achieving true operating condition immunity.

[0097] Here is a simple example of calculating the operating condition mutation index:

[0098] Scenario 1, under normal operating conditions , It also fluctuates around 0.5, therefore V, A value closer to 1 indicates a more normal value.

[0099] Scenario 2: When a sudden fault occurs (instantaneous jump), , , This value is much larger than the normal value. This achieves normalized amplification of fault signals.

[0100] Thus, by constructing the ratio of instantaneous volatility to baseline slow volatility, this method compares the instantaneous fluctuations of the fault signal with the slow fluctuations of the normal baseline, thereby achieving normalized amplification of the fault signal and improving the signal-to-noise ratio of the detection.

[0101] S4: Determine the current operating condition type based on the current operating condition data, and determine whether the winding machine has malfunctioned based on the comparison result of the operating condition mutation index and the adaptive threshold corresponding to the operating condition type.

[0102] In one embodiment, the method for determining the current operating condition type based on the current rotational speed and load current is as follows:

[0103] Pre-defined operating condition types include stop state, start-up acceleration state, full-load constant speed state, and deceleration stop state. For each operating condition type, the rotational speed and load current at each moment are used to construct an operating condition feature vector, forming a feature vector library for each operating condition type. The rotational speed and load current at the current moment are combined to form the operating condition feature vector for the current moment. The operating condition feature vector at the current moment is then compared with the feature vector library for each operating condition type. The Euclidean distance between the current operating condition feature vector and each operating condition feature vector in the feature vector library for each operating condition type is calculated. The reciprocal of the Euclidean distance measures the similarity, and the operating condition type at the current moment is determined based on the principle of maximum similarity.

[0104] Vibration characteristics vary greatly under different operating conditions. High vibration during acceleration is normal, while the same vibration during steady state may indicate bearing damage. It is necessary to identify the operating condition first in order to select the appropriate judgment criteria. This step enables refined management by operating condition and avoids the failure problem of a single threshold under complex operating conditions.

[0105] In one embodiment, the adaptive threshold of the operating condition mutation index corresponding to the operating condition type is determined as follows: when the winding machine is in a healthy operating state, the operating condition mutation index sequence for each operating condition type is collected in advance, the mean and standard deviation of the operating condition mutation index sequence are calculated, and based on the statistical principle of three standard deviations, the mean plus three standard deviations is set as the adaptive threshold of the operating condition mutation index corresponding to that operating condition type.

[0106] In one embodiment, the method for determining whether the winding machine has malfunctioned is as follows, in response to a comparison between the operating condition mutation index and the adaptive threshold corresponding to the operating condition type:

[0107] If the current condition change index is greater than the adaptive threshold corresponding to the current condition type, it is determined that the winding machine has malfunctioned at the current moment; if the current condition change index is not greater than the adaptive threshold corresponding to the current condition type, it is determined that the winding machine has not malfunctioned at the current moment.

[0108] The dynamically adjusted threshold achieves dual self-adaptation: maintaining a low threshold in steady state to detect minor faults (such as micro-cracks in the outer ring of the bearing), and switching to a high threshold during dynamic processes to tolerate normal shocks, thus completely resolving the contradiction between false alarms and missed alarms.

[0109] Specifically, data collection under different working conditions , Indicates the stop condition The sequence, W2, represents the startup condition. sequence, Indicates uniform speed condition sequence, Indicates the deceleration condition The sequence is analyzed, and its statistical distribution characteristics, including the mean, are calculated. and standard deviation .

[0110] Calculate the adaptive threshold : .

[0111] This provides information on adaptive thresholds. A simple example:

[0112] Operating Condition 1 Under working conditions It is very stable, and its mean value can be obtained statistically. Standard deviation , This is a very low sensitivity threshold, used to detect subtle real-world faults.

[0113] Operating Condition Two Under working conditions Normal fluctuations are extremely large, and their mean is statistically derived. Standard deviation , This is a very high robustness threshold used to ignore normal transient fluctuations.

[0114] After obtaining all working conditions After pre-storing, the system enters the online fault prediction phase. At any given moment, the first step is to determine the current operating condition. Then, the corresponding threshold is retrieved from the pre-stored threshold table. ,if : Determine that the equipment is operating normally under the current conditions; if It determines that a sudden equipment failure has occurred and issues a warning signal.

[0115] Here is a simple example of fault prediction:

[0116] During normal acceleration: time The system identifies the operating condition as follows: (Initiation accelerated state). The system detected... ,at this time The index fluctuates normally in an instant. ,because The system determined it to be normal, successfully avoiding a false alarm.

[0117] During steady-state faults: time The system identifies the operating condition as follows: (Full load, constant speed state), the system query found ,at this time Calculated due to fault mutation ,because The system detected a malfunction and successfully issued a warning.

[0118] This invention solves the drift problem of the baseline model under non-steady-state conditions by introducing a learning admission factor, and constructs a dimensionless mutation index by combining the ratio of instantaneous to slowly varying volatility. Combined with adaptive thresholds for different working conditions, it realizes a high-precision fault prediction scheme that can adapt to complex industrial field environments.

[0119] The effects of the present invention and the prior art will be described in conjunction with the accompanying drawings, such as... Figure 2 As shown: Figure 2 The upper part shows that the vibration signal exhibits severe but normal fluctuations during the start-up acceleration phase. During the full-load constant speed phase, the vibration signal shows a slow drift, and at the fault abrupt change point, a sharp pulse with an amplitude significantly weaker than the start-up acceleration fluctuation appears. Figure 2 The middle section shows the instantaneous difference calculated by a conventional algorithm and a fixed threshold. The startup fluctuation zone produces a very high instantaneous difference peak. To avoid false alarms, the fixed threshold is forced to be set above the startup fluctuation peak. In the fault mutation zone, the instantaneous difference peak caused by the fault signal is much lower than this high threshold, resulting in missed alarms (fault peak < threshold), clearly demonstrating the limitations of the existing technology. Figure 2The lower part of this figure shows the operating condition mutation index and dynamically changing adaptive threshold calculated by this invention. The Y-axis is a logarithmic scale (labeled as decimal numbers "1", "10", "100", "1000"). The stepwise change of the adaptive threshold is clearly visible: in the startup fluctuation zone, the threshold synchronously rises to a high level (1000.0); in the full-load constant speed zone, the threshold decreases to a very low level (5.0). In the startup fluctuation zone, the peak value of the operating condition mutation index is lower than the synchronously rising adaptive threshold, successfully immunizing against false alarms. In the fault mutation zone, the operating condition mutation index is significantly amplified, and its peak value is much higher than the adaptive threshold at this time (5.0), as shown by the green arrow, achieving accurate alarm (index > threshold).

[0120] Thus, by setting different adaptive thresholds for different operating conditions, this method can use a high threshold to avoid false alarms during startup and acceleration, while using a low threshold to sensitively capture weak faults during full-load conditions, perfectly resolving the contradiction between robustness and sensitivity.

[0121] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only.

Claims

1. A fault prediction method for automotive motor production equipment based on machine learning, characterized in that, include: Take any point in time as the current time, and obtain the vibration data of the main bearing of the winding machine and the operating condition data of the main spindle motor at the current time; Based on the deviation between the current operating condition data and the pre-acquired standard operating condition data, the learning admission factor at the current moment is calculated. A dynamic drift baseline model is constructed based on the learning admission factor and vibration data to output the vibration benchmark value and the slow-varying volatility benchmark value at the current moment. Based on the instantaneous energy deviation between the vibration data at the current moment and the dynamic vibration benchmark value, the instantaneous volatility at the current moment is calculated, and the ratio of the instantaneous volatility to the slowly varying volatility benchmark value is used as the operating condition change index at the current moment. The operating condition type at the current moment is determined based on the operating condition data at the current moment, and the adaptive threshold of the operating condition mutation index corresponding to the operating condition type is obtained. When the operating condition mutation index is greater than the adaptive threshold, it is determined that the winding machine has a fault, so as to realize the fault prediction of automotive motor production equipment. The learning access factor at the current moment is determined by the following relationship: ;in, for Learning access factors at all times It is a natural exponential function. The preset sensitivity coefficient, for Rotation speed at any given moment for Load current at any given time and These are the rated speed and rated load current of the spindle motor, respectively. A dynamic drift baseline model is constructed based on the learning admission factor and vibration data to output the current vibration benchmark value and the slow-varying volatility benchmark value. This is based on the following relationship: ; ;in, for Vibration reference value at time, for Vibration reference value at time, for The benchmark value of slow-varying volatility at time t. for The benchmark value of slow-varying volatility at time t. The preset base drift learning rate, for Learning access factors at all times for Vibration data at any given moment.

2. The fault prediction method according to claim 1, characterized in that, The current operating condition data and standard operating condition data include: The current operating data includes: the spindle motor speed and load current at the current moment; Standard operating condition data includes: rated speed and rated load current of the spindle motor.

3. The fault prediction method according to claim 1, characterized in that, Another way to determine the sensitivity coefficient is through a data-driven self-calibration method: During the calibration phase when the winding machine is running at full load and constant speed, a continuous speed sequence and load current sequence are collected. Calculate the first variance of the normalized deviation of the speed sequence relative to the rated speed, and the second variance of the normalized deviation of the load current sequence relative to the rated load current. Based on the statistical characteristics of Gaussian distribution, take the reciprocal of twice the sum of the first and second variances as the sensitivity coefficient.

4. The fault prediction method according to claim 1, characterized in that, Another way to determine the base drift learning rate is to obtain the time required for the main bearing of the winding machine to reach a thermally stable state from a cold start, as well as the sampling frequency of the vibration data. Based on the definition of the time constant of exponential smoothing, the reciprocal of the product of this time and the sampling frequency is used as the base drift learning rate.

5. The fault prediction method according to claim 1, characterized in that, The instantaneous volatility at the current moment is determined by the following relationship: ； in, for Instantaneous volatility at time t, for Instantaneous volatility at time t, The preset instantaneous response coefficient, for Vibration data at any given time for The vibration reference value at any given time.

6. The fault prediction method according to claim 1, characterized in that, The method for determining the current operating condition type based on the current rotational speed and load current is as follows: Pre-defined operating condition types, including stop state, start-up acceleration state, full-load constant speed state, and deceleration stop state, and construct an operating condition feature vector by taking the speed and load current at each moment under each operating condition type, so as to form a feature vector library for each operating condition type; The current speed and load current are combined to form the current operating condition feature vector. The current operating condition feature vector is then matched with the feature vector library for each operating condition type. Based on the principle of maximum similarity, the operating condition type at the current moment is determined.

7. The fault prediction method according to claim 1, characterized in that, The vibration data of the main bearing of the winding machine at the current moment, as well as the speed and load current of the main spindle motor at the current moment, are obtained in real time by accessing the servo driver or programmable logic controller of the winding machine through the communication interface of the winding machine, and a unified timestamp is configured for the vibration data, speed and load current.

8. The fault prediction method according to claim 1, characterized in that, The adaptive threshold for the operating condition abrupt change index corresponding to the operating condition type is determined based on the following method: Beforehand, when the winding machine is in a healthy operating state, the working condition mutation index sequence for each working condition type is collected, and the mean and standard deviation of the working condition mutation index sequence are calculated. Based on the statistical principle of three standard deviations, the mean plus three standard deviations is set as the adaptive threshold of the working condition mutation index corresponding to that working condition type.

Images