Multi-dimensional state operation monitoring method and system based on yaw collector ring

Abstract

The application relates to the technical field of yaw slip ring state monitoring, and specifically discloses a multi-dimensional state operation monitoring method and system based on a yaw slip ring, which comprises the following steps: obtaining an original operation data set of the yaw slip ring, wherein the original operation data set contains operation parameters in the electrical, mechanical, attitude and environmental dimensions which are collected via sensors and subjected to redundancy checking; constructing a noise suppression matrix based on the physical mechanism of the yaw slip ring, filtering isolated fluctuations and cross-modal abnormal values in the original operation data set by using the noise suppression matrix, and inputting the noise suppression matrix into a convolutional neural network model to obtain a fusion feature vector. By introducing the noise suppression matrix constructed based on the physical mechanism and the channel interaction mechanism guided by the physical topology, the technical problem that a lightweight model is difficult to extract cross-modal strong coupling features at an edge end is effectively solved, and the sensitivity and accuracy for early composite faults are significantly improved without increasing the computing power burden.

Description

Technical Field

[0001] This invention relates to the field of yaw collector ring condition monitoring technology, and in particular to a multi-dimensional condition operation monitoring method and system based on yaw collector ring. Background Technology

[0002] As a key component connecting the fixed tower and the rotating nacelle in a wind turbine generator, the yaw slip ring bears the core function of transmitting electrical load and control signals. Its operational stability directly affects the generator's power generation efficiency and safety. Existing monitoring technologies typically employ a single-dimensional threshold alarm mechanism, independently collecting vibration, temperature, or current signals, triggering an alarm when a parameter exceeds a fixed limit. However, this discrete monitoring method ignores the inherent correlations between various physical quantities of the slip ring under complex operating conditions, making it difficult to adapt to operating scenarios with drastic fluctuations in load and ambient temperature. This leads to a high likelihood of missed alarms or false alarms due to environmental interference in the early stages of minor faults.

[0003] More importantly, existing technologies face irreconcilable technical conflicts when dealing with fault feature extraction and production maintenance scheduling:

[0004] On the one hand, in order to adapt to the limited computing power at the edge, the monitoring model is often forced to adopt a minimalist and lightweight structure. However, this makes it unable to capture the cross-modal nonlinear strong coupling characteristics of mechanical wear-contact resistance-temperature rise in the yaw collector ring, and easily misjudges the early weak signs with physical causal relationship as noise.

[0005] On the other hand, the existing maintenance decision-making system and the production execution system are disconnected. When the health of the equipment drops to a critical point but there is an uninterrupted rigid production task, such as when the capacity trough window is zero, the system lacks the means to actively intervene in the remaining lifespan of the equipment, and is caught in the dilemma of either forcibly shutting down the machine and causing huge default losses, or operating the machine with defects and causing the equipment to be completely destroyed. Summary of the Invention

[0006] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a multi-dimensional state operation monitoring method and system based on a yaw collector ring, to improve the sensitivity and accuracy of early-stage complex faults.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a multi-dimensional state operation monitoring method based on a yaw collector ring, comprising the following steps:

[0008] Obtain the raw operating dataset of the yaw collector ring, which includes operating parameters in electrical, mechanical, attitude and environmental dimensions collected by sensors and verified for redundancy.

[0009] A noise suppression matrix is ​​constructed based on the physical mechanism of the yaw collector ring. The noise suppression matrix is ​​used to filter isolated fluctuations and cross-modal outliers in the original running dataset and input into the convolutional neural network model to obtain a fused feature vector.

[0010] Based on the basic factors of the historical fault dimension and the real-time factors of the real-time operating condition dimension, dynamic weights for different monitoring dimensions are calculated, and the dynamic weights are smoothly updated using the exponential moving average method.

[0011] Based on the current load and temperature conditions, the corresponding operating condition sub-reference is matched from the graded calibration reference library, and the frequency domain correlation coefficient and relative deviation between the fused feature vector and the operating condition sub-reference are calculated to mark abnormal features.

[0012] Combining the dynamic weights and the abnormal features, the equipment health is calculated in different dimensions, and a comprehensive health score is generated by combining the operating condition attenuation coefficient.

[0013] In response to the overall health score falling below the health alarm threshold, a time series regression model is used to predict the decay warning time window, and maintenance instructions are generated in conjunction with the idle window of the manufacturing execution system.

[0014] To achieve the above objectives, a second aspect of the present invention proposes a multi-dimensional state operation monitoring system based on a yaw collector ring, comprising:

[0015] The data acquisition and preprocessing module is used to acquire the raw operating dataset of the yaw collector ring and perform denoising processing based on the noise suppression matrix constructed based on the physical mechanism.

[0016] The feature fusion and analysis module is used to generate fused feature vectors by using a convolutional neural network with a deep separable convolutional structure, and to calculate dynamic weights based on historical fault dimensions and real-time operating condition dimensions.

[0017] The status monitoring and assessment module is used to compare the fused feature vector based on the working condition sub-benchmark to mark abnormal features, and to calculate the multi-dimensional health score and the comprehensive health score in combination with the dynamic weight.

[0018] The decision and maintenance response module is used to generate maintenance instructions by using a prediction model to output a decay warning time window when the overall health score is lower than the health alarm threshold, and combining it with the idle window of the production execution system.

[0019] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described multi-dimensional state operation monitoring method based on a yaw collector ring.

[0020] The multi-dimensional state operation monitoring method and system based on yaw collector ring of this invention effectively solves the technical problem of lightweight models having difficulty extracting cross-modal strong coupling features at the edge by introducing a noise suppression matrix and a channel interaction mechanism guided by physical topology. It significantly improves the sensitivity and accuracy of early composite faults without increasing the computing power burden. At the same time, this invention establishes a dynamic weight allocation and calibration benchmark system based on operating condition adaptation, which overcomes the false alarm problem of single threshold under changing operating conditions.

[0021] More importantly, this invention deeply integrates equipment health management with manufacturing execution systems. When faced with extreme conflicts where there is no maintenance window and production tasks cannot be interrupted, it can use a proactive load reduction and life extension strategy based on a sensitivity model to extend the remaining life of the equipment at the expense of local production efficiency. This breaks the time-space deadlock between rigid production and equipment safety, and achieves the optimal balance between economic benefits and operational safety. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the multi-dimensional state operation monitoring method based on yaw collector ring provided by the present invention.

[0023] Figure 2 This is a schematic diagram comparing the effects of multi-source parameter physical correlation filtering in the multi-dimensional state operation monitoring method based on yaw collector ring provided by the present invention;

[0024] Figure 3 This is a schematic diagram of the adaptive dynamic weight evolution curve of the operating condition in the multi-dimensional state operation monitoring method based on yaw collector ring provided by the present invention.

[0025] Figure 4 This is a heatmap of the physical topology adjacency matrix of the feature channel in the multi-dimensional state operation monitoring method based on yaw collector ring provided by the present invention;

[0026] Figure 5 This is a simulation diagram of the enhanced response of physical coupling factor features in the multi-dimensional state operation monitoring method based on yaw collector ring provided by the present invention;

[0027] Figure 6 This is a histogram showing the sensitivity of health decay rate to process parameters in the multi-dimensional state operation monitoring method based on yaw collector ring provided by this invention.

[0028] Figure 7 This is a comparison chart of life extension trajectory prediction under the active load reduction strategy in the multi-dimensional state operation monitoring method based on yaw collector ring provided by the present invention.

[0029] Figure 8This is a schematic diagram illustrating the implementation of the multi-dimensional state operation monitoring system based on yaw collector ring provided by the present invention.

[0030] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0031] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0032] The following description, with reference to the accompanying drawings, describes a multi-dimensional state operation monitoring method, system, and electronic device based on a yaw collector ring according to embodiments of the present invention.

[0033] Example 1:

[0034] This embodiment details a multi-dimensional state monitoring method based on a yaw collector ring. This method is configured for use in a monitoring server and aims to address the problem of missed early fault detection caused by neglecting the intrinsic correlations between various physical quantities of the yaw collector ring in existing technologies, as well as the spatiotemporal conflict between production maintenance and equipment health management. This embodiment achieves precise control over the entire lifecycle of the yaw collector ring by constructing a physical mechanism noise suppression matrix, a lightweight feature fusion model, and a production-maintenance collaborative mechanism.

[0035] like Figure 1 As shown, the method in this embodiment includes the following steps:

[0036] S1, Data Acquisition and Monitoring Strategy Construction Phase.

[0037] First, the method in this embodiment performs the step of obtaining the raw operating dataset of the yaw collector ring.

[0038] Specifically, the monitoring server establishes communication connections with various types of sensors deployed at the yaw collector ring site. The raw operational dataset includes operational parameters in electrical, mechanical, attitude, and environmental dimensions, collected by the sensors and redundancy-checked. To ensure data accuracy, redundancy is implemented at the sensor level. For example, for wear monitoring, ultrasonic and eddy current sensors are used in a heterogeneous redundant deployment; for vibration monitoring, piezoelectric and capacitive sensors are used in a redundant deployment. After receiving the data, the monitoring server performs a data consistency check, eliminating explicit erroneous data caused by sensor malfunctions, thus forming the raw operational dataset.

[0039] It should be noted that, before obtaining the original operational dataset of the yaw collector ring, this embodiment also specifies in detail the process of determining the acquisition strategy for monitoring items in order to balance monitoring accuracy and data transmission and storage costs.

[0040] Specifically, the monitoring server first performs the step of acquiring historical fault analogy data. This historical fault analogy data is not randomly selected, but rather based on historical operating records of equipment with the same or similar models as the currently monitored yaw collector ring, retrieved from a big data platform. The historical fault analogy data specifically includes the number of historical faults of similar equipment, the maximum single maintenance cost, and the maximum downtime. These three dimensions represent the frequency of fault occurrence, the degree of economic loss, and the degree of impact on production efficiency, respectively.

[0041] For example, the monitoring server calculates a monitoring priority value based on the historical fault analogy data. This calculation process aims to quantify the importance of different monitoring items, such as ambient temperature, carbon brush wear rate, and vibration amplitude. The monitoring priority value is positively correlated with the number of historical faults, meaning that the more frequent the fault, the higher the priority of its corresponding monitoring item; simultaneously, the monitoring priority value is functionally related to the sum of the maximum single repair cost and the maximum downtime.

[0042] For example, the formula for calculating the monitoring priority value can be defined as:

[0043] ;

[0044] In the formula, This represents the monitoring priority value; This represents the normalized maximum cost of a single repair. This represents the maximum downtime after normalization. This represents the number of historical failures corresponding to the monitoring item. Using this formula, the system can calculate the quantitative score for each monitoring item and sort them according to the monitoring priority value, thereby determining which parameters are core monitoring objects and which are auxiliary monitoring objects.

[0045] Furthermore, the monitoring server calculates error correction values ​​based on historical acquisition error rates. The acquisition error rate reflects the sensor's stability under specific conditions. The introduction of error correction values ​​is to prevent data distortion caused by the sensor's inherent instability. Based on the numerical range of the error correction values, the system sets different data acquisition frequencies for different monitoring items in stages.

[0046] Specifically, the error correction value can be calculated based on the ratio of invalid data points to the total number of collected data points within a historical time window. The system presets several numerical ranges, such as low error, medium error, and high error ranges. When the error correction value for a monitoring item, such as dust concentration, falls within the high error range, it indicates that the parameter fluctuates significantly and is difficult to measure. The system will automatically increase the data acquisition frequency for this monitoring item, such as from once per minute to five times per minute. Oversampling techniques and subsequent filtering algorithms are used to approximate the true value. Conversely, for stable parameters with error correction values ​​in the low error range, the acquisition frequency is reduced to save bandwidth. This dynamic strategy, driven by historical data, ensures the high quality and high availability of the original operational dataset.

[0047] S2, Noise suppression and feature extraction stage based on physical mechanisms.

[0048] After obtaining a high-quality raw running dataset, the monitoring server will enter the core stage of data processing.

[0049] Specifically, this embodiment performs the step of constructing a noise suppression matrix based on the physical mechanism of the yaw collector ring, and using the noise suppression matrix to filter isolated fluctuations and cross-modal outliers in the original running dataset.

[0050] It should be noted that traditional filtering methods are usually based on the time-frequency characteristics of a single signal, such as low-pass filtering, but they cannot identify pseudo-signals that conform to spectral characteristics but violate physical laws. This embodiment introduces physical constraints by constructing a parameter correlation matrix that includes vibration parameters, yaw angle parameters, temperature parameters, and current parameters.

[0051] Specifically, physical correlation thresholds are set in the parameter correlation matrix. These physical correlation thresholds are not single values, but rather boundary conditions describing the interaction between two physical quantities. The physical correlation thresholds include a positive correlation coefficient threshold between vibration and yaw angle, and a linear correlation coefficient threshold between temperature and current.

[0052] For example, regarding the correlation between vibration and yaw angle, the physical mechanism shows that when the yaw slip ring performs a yaw action, i.e., the yaw angle changes, the vibration amplitude will inevitably increase due to mechanical friction and gear meshing. Therefore, the system sets a positive correlation coefficient threshold, such as 0.85. Then, the monitoring server monitors the fluctuation of the single parameter in the original running data in real time. If a sudden and significant increase in vibration amplitude is detected, but the yaw angle remains stationary at the same time (i.e., the yaw angle change rate is 0), and the correlation between the two is much lower than the positive correlation coefficient threshold, the system determines that the fluctuation of the single parameter (vibration) is noise, which may be caused by external impact or sensor loosening, and ultimately filters it out.

[0053] like Figure 2 The signal processing effect of the multi-source parameter physical correlation filtering method in actual operation is shown. The horizontal axis of the three sub-figures is uniformly labeled as the sampling time, and the vertical axis represents the numerical value of each physical quantity.

[0054] Figure 2 The first part, the blue curve, depicts the raw vibration acceleration signal collected by the sensor. A sudden high-amplitude spike fluctuation appears around two seconds, exhibiting obvious outlier characteristics.

[0055] Figure 2 The second part, the red curve, shows the change in yaw angle within the same time period. Around two seconds, the curve remains a horizontal straight line, indicating that the slip ring did not perform a yaw action. Based on the noise suppression matrix logic proposed in this invention, the system detects a drastic fluctuation in a single vibration parameter without a corresponding change in its associated yaw angle parameter. It determines that this fluctuation does not conform to the laws of physical friction transmission and is a pseudo-signal generated by environmental interference.

[0056] Therefore in Figure 2 In the filtered signal shown by the green curve in the third part of the image, the spike at that point has been effectively removed and restored to a smooth background noise. However, in the four- to seven-second interval, the original vibration signal shows an amplitude increase again, and the corresponding yaw angle curve shows an upward trend, indicating that the unit is yawing. The system identifies that the two meet the physical correlation threshold requirement and determines it to be a genuine mechanical vibration characteristic. Therefore, this waveform change is completely preserved in the filtered result, thus proving that this method can accurately retain real fault signs while filtering out false noise.

[0057] Similarly, for the related parameters of temperature and current, the physical mechanism is based on Joule's law (…). This means that an increase in current inevitably leads to an increase in temperature. The system sets a threshold for the linear correlation coefficient between temperature and current, such as 0.7. If a sharp rise in temperature is detected, but the current load remains low, the temperature anomaly is judged as an isolated fluctuation not caused by an electrical fault (which may be due to direct sunlight or an ambient heat source), and is therefore removed from the electrical fault diagnosis model or marked as environmental interference.

[0058] Furthermore, this step also includes setting a correlation threshold between wear rate and vibration amplitude based on the aforementioned operating condition sub-reference, and removing isolated data points that exceed the correlation threshold. Physically, there is a positive correlation between the wear rate of a carbon brush and the vibration amplitude of the contact surface. If an extremely high wear rate reading appears in the dataset, but the corresponding vibration amplitude is very low, this violates the physical laws of wear. The system identifies such cross-modal anomalies as measurement errors and performs cleanup.

[0059] After the above physical denoising process, the purified data is input into the feature extraction module. Specifically, this embodiment performs the step of inputting the data into a convolutional neural network model to obtain a fused feature vector.

[0060] Optionally, to improve the targeting of feature extraction, the monitoring server first classifies the filtered operating parameters into electrical, mechanical, and attitude categories. Electrical parameters include three-phase current, voltage, and frequency; mechanical parameters include vibration acceleration, carbon brush wear, and contact pressure; and attitude parameters include yaw angle, yaw speed, and nacelle position. These categorized data are then input into attention networks constructed for specific modalities.

[0061] Specifically, the attention network simulates the diagnostic thinking of human experts, focusing on different parameters under different fault modes. The attention network calculates the weights of each feature and selects core features whose weight ranking meets a preset ratio. For example, in a suspected contact failure mode, the weights of current fluctuations and temperature changes are automatically increased. The system sorts the features by weight and selects the top-ranked (e.g., top 30%) features as the core features. These core features include at least the effective current value, wear rate, yaw angle variation trend, and vibration dominant frequency amplitude; these features most significantly characterize the operating status of the slip ring.

[0062] Subsequently, the core features are input into a convolutional neural network employing a depthwise separable convolution structure, outputting the fused feature vector. It's important to note that depthwise separable convolution is used here instead of standard convolution, aiming to achieve a lightweight model that can be deployed on computationally limited edge monitoring servers or embedded boards. Depthwise separable convolution decomposes standard convolution into depthwise convolution and pointwise convolution, significantly reducing computational cost while maintaining effective extraction of time-series features. The resulting fused feature vector is a highly compressed data representation rich in state information.

[0063] S3, Dynamic weight allocation and benchmark comparison stage.

[0064] After obtaining the fused feature vector, in order to overcome the problem of high false alarm rate of traditional fixed threshold monitoring method under varying operating conditions, this embodiment introduces dynamic weight and adaptive benchmark mechanism.

[0065] Specifically, this embodiment performs the step of calculating dynamic weights for different monitoring dimensions based on the basic factors of the historical fault dimension and the real-time factors of the real-time operating condition dimension.

[0066] First, the system normalizes the monitoring priority values ​​calculated in the preceding steps to obtain the basic factor. The basic factor reflects the importance of this monitoring dimension in historical experience and is a relatively static quantity. For example, if overheating faults in the annular system have historically been the most frequent, the basic factor for the temperature dimension will be initialized to a large value, such as 0.4.

[0067] However, static weights cannot adapt to the complex and ever-changing operating environment of wind turbines. Therefore, the monitoring server further uses a pre-built mapping table of operating conditions and parameter contribution to match and obtain the real-time factors based on the current real-time operating conditions. The mapping table of operating conditions and parameter contribution is constructed based on a large number of simulation experiments and field measurement data. It defines the sensitivity, or contribution, of each monitoring parameter to fault indication under different load rates, ambient temperatures, and yaw frequency.

[0068] For example, when the unit is under high load and high temperature conditions, the risk of electrical insulation aging increases dramatically. In this case, the mapping table will indicate an increase in the weight of the electrical dimension and a decrease in the weight of mechanical vibration. Conversely, when the unit is under high wind and frequent yaw conditions, the risk of mechanical wear increases, and the mapping table will indicate an increase in the weight of the mechanical dimension. The system obtains the current real-time factors by looking up the table.

[0069] Next, the system performs a weighted summation of the basic factor and the real-time factor according to a preset proportional coefficient to obtain the real-time weight. This proportional coefficient is used to balance the influence of historical experience and current operating conditions; for example, the basic factor is set to account for 40% and the real-time factor for 60%.

[0070] To avoid instability in the evaluation results due to drastic changes in weights, this embodiment utilizes the Exponential-Moving-Average (EMA) method to calculate the final dynamic weights based on the currently calculated real-time weights and the weights from the previous cycle. The calculation formula can be expressed as:

[0071] ;

[0072] In the formula, It is the final dynamic weight at the current moment; This is the currently calculated real-time weight; It is the dynamic weight of the previous moment; This is the smoothing coefficient, such as 0.3. Through the EMA algorithm, the changes in weights exhibit a smooth transition characteristic, which can respond to changes in operating conditions while avoiding sudden disturbances.

[0073] like Figure 3 The paper demonstrates the response characteristics of the adaptive dynamic weight allocation mechanism in this invention during actual operation. The horizontal axis represents the running time of the monitoring system, the blue vertical axis on the left represents the real-time load rate of the yaw collector ring, and the red vertical axis on the right represents the dynamic weight value of a specific monitoring dimension.

[0074] Figure 3 The blue stepped curve in the figure reflects the drastic changes in external operating conditions. For example, at the twentieth minute, the load rate jumps from 30% to 90% instantly, which means that the unit has entered a high-load operating state. At this time, the target weight calculated according to the mapping table of operating conditions and parameter contribution also changes accordingly.

[0075] and Figure 3 The red solid line in the figure depicts the dynamic weight evolution trajectory after being processed by the exponential moving average algorithm. It can be seen that when the working condition changes suddenly, the red curve does not rise vertically, but shows a gradual upward trend, and only gradually approaches the new target weight value after about ten to fifteen minutes.

[0076] Figure 3 The red solid line in the figure clearly shows that the hysteresis and smoothing effect demonstrates that the algorithm used in this invention can effectively suppress the oscillation of the evaluation system caused by instantaneous fluctuations in operating conditions, and avoid the health score from fluctuating due to drastic changes in weight parameters, thereby ensuring the robustness of the monitoring system and the continuous stability of the evaluation results under varying operating conditions.

[0077] While determining the weights, the monitoring server performs the following steps: based on the current load and temperature status, it matches the corresponding operating condition sub-reference from the graded calibration reference library, calculates the frequency domain correlation coefficient and relative deviation between the fused feature vector and the operating condition sub-reference, and marks abnormal features.

[0078] Specifically, the graded calibration benchmark library is no longer a single normal value, but rather multiple sub-benchmarks divided into load-temperature two-dimensional grids. For example, [load 80%-100%, temperature 40℃-60℃] corresponds to a specific benchmark model. When the system detects that it is currently in this range, it calls the corresponding sub-benchmark. The calculation process not only compares the numerical magnitude (relative deviation), but more importantly, it calculates the frequency domain correlation coefficient. This is because some faults, such as eccentricity, may not exceed the amplitude limit in the time domain, but will exhibit specific octave components in the frequency domain. By calculating the similarity between real-time features and benchmark features in the frequency domain, i.e., the correlation coefficient, the system can sensitively capture weak waveform distortions, thereby marking those features with low correlation coefficients or large deviations as the abnormal features.

[0079] S4, Health Measurement Assessment and Comprehensive Scoring Stage.

[0080] Based on the dynamic weights calculated above and the marked abnormal features, the system enters the health measurement stage.

[0081] Specifically, this embodiment performs the following steps: combining the dynamic weights and the abnormal features to calculate the equipment health across dimensions, and combining the operating condition attenuation coefficient to generate a comprehensive health score.

[0082] The monitoring server first calculates electrical health, mechanical health, and attitude health. Electrical health is calculated based on the deviation of the effective current value and the degree to which harmonic components exceed the standard. For example, a certain score is deducted for every 1% increase in current deviation; non-linear deductions are applied when the harmonic distortion rate exceeds the national standard limit. Mechanical health is calculated based on wear rate, vibration amplitude, and dynamic damping ratio. Wear rate is a core indicator for evaluating carbon brush life, vibration amplitude reflects contact stability, and dynamic damping ratio reflects the system's shock absorption capacity. Attitude health is calculated based on yaw angle deviation and speed fluctuation. Lag or overshoot of the yaw angle relative to the command value, as well as speed instability, directly translate to a decrease in attitude health.

[0083] After obtaining the original health scores for each sub-dimension, the system also needs to incorporate environmental impact corrections. Specifically, the operating condition attenuation coefficient is determined based on the current load rate and ambient temperature. This coefficient reflects the accelerated wear and tear on equipment lifespan caused by harsh operating conditions. For example, under high load (≥80%) and low temperature (<-5℃, causing materials to become brittle), the operating condition attenuation coefficient will be set to a value greater than 1.0, such as 1.2, which means that under the same physical conditions, the health risk is higher.

[0084] Finally, the system performs a comprehensive calculation. Specifically, the monitoring server calculates the electrical health, mechanical health, and attitude health scores obtained above, and then performs a weighted summation based on their respective dynamic weights to obtain a basic weighted score. Considering the accelerated wear and tear on equipment lifespan caused by harsh operating conditions, the system needs to correct this basic weighted score using an operating condition attenuation coefficient to accurately reflect the actual health status of the equipment in the current environment.

[0085] In this embodiment, the operating condition attenuation coefficient is defined as a value greater than or equal to 1.0, where a larger value indicates a more severe operating environment. Therefore, the system divides the basic weighted score by the operating condition attenuation coefficient to obtain the comprehensive health score. The calculation formula is as follows:

[0086] ;

[0087] In the formula, This represents the final calculated overall health score; This represents the calculated electrical health level; Represents the dynamic weights corresponding to the electrical dimension; This represents the calculated mechanical health status; Represents the dynamic weights corresponding to the mechanical dimension; This represents the calculated posture health score; Represents the dynamic weights corresponding to the pose dimension; This represents the operating condition attenuation coefficient determined based on the current load rate and ambient temperature.

[0088] Calculations based on this formula show that when the unit is under high load or extreme temperature conditions, i.e. The calculated comprehensive health score It will be lower than the nominal basic weighted score, thus realizing the quantitative representation of operational risks.

[0089] S5, Maintenance Decision Generation and Cross-Device Collaboration Phase.

[0090] When the overall health score drops to a certain level, the monitoring server will trigger a maintenance decision process.

[0091] Specifically, this embodiment executes the following steps in response to the overall health score falling below the health alarm threshold: using a time series regression model to predict the decay warning time window, and combining this with the idle window of the production and manufacturing execution system to generate maintenance instructions.

[0092] The monitoring server has a pre-set health alarm threshold, such as 60 points. Once the real-time score falls below this value, the system first inputs the comprehensive health score into a Long Short-Term Memory (LSTM) network model. LSTM has the ability to remember both short-term and long-term information, and can predict the health change curve within a preset time period (such as the next 30 days) based on the health decline trend over the past 30 days or longer.

[0093] Next, the system determines the time interval within which the health status change curve drops to the critical failure threshold (e.g., 40 points, representing that the equipment is about to be completely damaged), and defines this interval as the attenuation warning time window. This time window not only provides the latest repair time but also the earliest repair time, i.e., the current moment.

[0094] To resolve the conflict between production and maintenance, the monitoring server extracts production idle windows from the Manufacturing Execution System (MES) via an API interface. These idle windows refer to non-operational time reserved in the production plan, specifically including periods of low capacity (such as order gaps) and changeover intervals (such as time for changing product molds).

[0095] The system executes a matching algorithm to find and match the longest idle production window within the time range of the attenuation warning window that meets the maintenance requirements (e.g., replacing the collector ring requires 4 hours). Once a match is successful, it is marked as the optimal maintenance window, and the system generates the maintenance instruction containing the maintenance type (e.g., replacing carbon brushes, cleaning the toroidal surface) and resource scheduling information (required spare parts, personnel qualifications).

[0096] It should also be noted that wind farms typically consist of multiple turbines, and these devices are interconnected. Therefore, this embodiment also executes cross-device collaboration logic.

[0097] Specifically, the monitoring server's generation of maintenance instructions in conjunction with the idle window of the manufacturing execution system also includes cross-device collaborative logic. First, the server establishes a device association influence matrix containing the main and slave slip rings of the generator set. In a doubly-fed wind turbine generator set, the slip rings (or slip rings) on the stator and rotor sides are tightly coupled in electrical connection and are physically adjacent. This matrix quantifies the degree of impact of a failure in one device on another.

[0098] When any device, such as the main collector ring, triggers the maintenance command, the system not only monitors that device but also automatically queries the device association influence matrix to find all related devices, such as the collector rings. At this time, the system adjusts the data acquisition frequency of the related devices to a second frequency, which is higher than the initial data acquisition frequency. For example, it increases the frequency from once per minute to five times per minute for enhanced monitoring.

[0099] Subsequently, the system assesses the health trend of the associated devices. Even if the current health of an associated device is not yet below the alarm threshold, if its health shows a downward trend and falls below a specific maintenance coordination threshold, such as 75 points (indicating a sub-healthy state), the system will make a merging decision to avoid a second downtime in the short term. Specifically, the maintenance tasks of the associated devices will be merged into the optimal maintenance window.

[0100] For example, if the main slip ring requires a 4-hour downtime for replacement, while the secondary slip ring is in a sub-optimal condition and requires 2 hours to repair separately, combining both operations would reduce the total time to only 5 hours due to shared downtime preparation, personnel access to the tower, and safety isolation. This is significantly less than the total time required for separate repairs (4 + 2 = 6 hours + losses from two downtimes). This cross-device collaborative mechanism maximizes the use of valuable downtime windows and significantly reduces maintenance costs.

[0101] In summary, this embodiment constructs a logically closed-loop, technologically advanced multi-dimensional status monitoring system for yaw collector rings by optimizing data acquisition and physical mechanism denoising at the bottom layer, lightweight feature extraction and dynamic weight evaluation at the middle layer, and production-maintenance collaboration and cross-equipment linkage at the top layer. This effectively solves the pain points in existing technologies and achieves a win-win situation for equipment operation safety and production economic benefits.

[0102] Example 2:

[0103] This embodiment, based on Embodiment 1, features in-depth technical optimizations to the feature extraction stage. Specifically, this embodiment focuses on describing a residual coupling enhancement mechanism guided by physical topology. This mechanism aims to address the problem of limited channel interaction capabilities in deeply separable convolutional structures commonly found in lightweight convolutional neural networks when processing cross-modal correlated features, ensuring accurate capture of complex physical coupling fault features in yaw collector rings even with limited edge computing resources.

[0104] Specifically, in the step of inputting core features into a convolutional neural network with a depthwise separable convolutional structure mentioned in Example 1, although depthwise separable convolution significantly reduces the number of parameters and computational complexity by decomposing standard convolution into two stages: depthwise convolution and pointwise convolution, this structure completely cuts off the information interaction between channels in the depthwise convolution stage. In the operation monitoring scenario of yaw slip rings, faults are often not isolated manifestations of a single dimension, but rather exhibit strong cross-modal coupling characteristics. For example, uneven pressure on the contact surface between the carbon brush and the slip ring can lead to mechanical vibration, which in turn causes fluctuations in contact resistance, ultimately manifesting as local temperature rise and current waveform distortion. This vibration-electricity-heat chain reaction manifests as a highly nonlinear correlation between different feature channels in the feature space. If only the subsequent linear or simple nonlinear pointwise convolution in standard depthwise separable convolution is relied upon, it is difficult to reproduce this strong physical coupling relationship in deep networks. Therefore, this embodiment introduces a parallel physical enhancement branch at the output of the depthwise convolutional layer.

[0105] For example, this embodiment first performs the step of constructing a physical topology map of the feature channels. This process is not automatically learned by a neural network, but rather defined manually or semi-automatically based on prior physical knowledge. The system maps each feature channel in the core features to a node in the physical topology map. Here, the feature channels are not merely abstract numerical matrices, but signal carriers with explicit physical meaning.

[0106] Specifically, assuming the input core feature tensor contains multiple channels, the system will semantically label these channels based on the physical structure and operating mechanism of the yaw slip ring. For example, the first channel is labeled as the A-phase current RMS value node, the second channel is labeled as the slip ring surface temperature node, the third channel is labeled as the radial vibration amplitude node, and the fourth channel is labeled as the yaw drive torque node.

[0107] After mapping the nodes, the system establishes physical coupling edges between causally related nodes based on the heat generation mechanism and vibration transmission mechanism of the yaw collector ring. The heat generation mechanism mainly follows Joule's law and the principle of frictional heat generation; therefore, strong coupling edges are established between current nodes and temperature nodes, and between vibration nodes (representing the frictional state) and temperature nodes. The vibration transmission mechanism follows the principles of mechanical dynamics; therefore, coupling edges are established between yaw torque nodes and radial vibration nodes. In this way, a sparse graph structure containing rich physical causal information is constructed, which is the physical topology graph of the characteristic channels.

[0108] It should be noted that, in order to mathematically describe this topology, this embodiment defines a physical adjacency matrix, denoted as . This matrix is ​​a matrix with dimension 1. The square array, in which The total number of channels representing the core features. The element values ​​in the matrix indicate whether there is a physical coupling edge between the corresponding two channel nodes. If the... The first channel and the first If there is a strong physical correlation between channels, the matrix element is assigned a value of 1; otherwise, it is assigned a value of 0. This binary definition method can effectively guide the subsequent calculation process and avoid the network wasting computational resources on irrelevant channel combinations.

[0109] like Figure 4 A heatmap of the physical topology adjacency matrix of the feature channels, constructed based on the physical mechanism of the yaw collector ring, is presented. The horizontal and vertical axes of the graph list the physical semantic labels of the core feature channels input to the neural network, specifically covering electrical parameter nodes such as A-phase current, B-phase current, and C-phase current, thermal parameter nodes such as ring temperature, and mechanical parameter nodes such as vibration amplitude and yaw torque.

[0110] Figure 4 The matrix cross-color blocks in the image intuitively represent whether there is a physical coupling edge between any two feature channels through the difference in color intensity. Darker blocks represent that there is a clear physical mechanism relationship between the two corresponding nodes, such as the strong coupling connection established between the current node and the temperature node based on Joule's law, and the interactive relationship established between the vibration node and the temperature node based on the principle of frictional heat generation. Lighter blocks represent that there is no direct physical causal relationship between the channels.

[0111] Figure 4 This sparse matrix distribution clearly reveals the physical topology guidance mechanism described in this invention, which forces the lightweight convolutional neural network to perform second-order interaction operations only between these physically meaningful dark channel pairs. This saves computing resources at the edge while accurately capturing cross-modal strong coupling fault features and avoids wasting computing power on invalid feature combinations.

[0112] Specifically, during the forward propagation of the convolutional neural network, after the data stream passes through the depthwise convolutional layer, a set of feature maps is output. At this point, the processing flow is divided into two paths. The main path continues into the standard pointwise convolutional layer to perform conventional linear combination of features and dimensionality increase or decrease operations. The auxiliary path, which is the enhancement branch that is the focus of this embodiment, performs physical topology-guided channel interaction enhancement operations in parallel at the output of the depthwise convolutional layer of the convolutional neural network.

[0113] Optionally, the specific execution logic of the channel interaction enhancement operation is as follows: the system extracts paired associated feature channels based on the physical coupling edges. This step is equivalent to utilizing the aforementioned physical adjacency matrix. A masking filter was performed on all possible channel combinations, retaining only those channel pairs that were physically relevant. For example, the system extracted the current channel feature map and the temperature channel feature map as a single processing pair.

[0114] Subsequently, the system performs a second-order interactive operation based on element-wise product on each pair of associated feature channels. This second-order interactive operation aims to explicitly capture the nonlinear covariance relationship between two physical quantities. In signal processing, the product of two signals can reflect their modulation relationship or covariance characteristics, which better characterizes complex physical coupling effects than simple linear superposition.

[0115] To accurately describe this calculation process, this embodiment defines a physical coupling factor. The calculation formula is as follows: Let For the first Feature map matrix of each channel For the first The feature map matrix of each channel, and the nodes With nodes If there is a physical coupling edge between the two channels, then the physical coupling factor component generated by that pair of channels is represented as:

[0116] ;

[0117] In the formula, the symbol This represents the Hadamard product, which is the element-wise multiplication of matrices. The generated... It is a matrix of the same size as the original feature map, highlighting spatial regions or time segments where current and temperature fluctuate drastically, thus characterizing cross-modal nonlinear relationships. The system performs the above operation on all channel pairs with physically coupled edges and combines all generated data. By splicing or aggregating, a complete physical coupling factor is ultimately generated.

[0118] It should also be noted that although the generated physical coupling factor contains rich second-order physical information, its channel dimension may not be consistent with the feature dimension of the main network, and its numerical distribution may differ from the features of the main network. Therefore, it cannot be directly superimposed back onto the main network. This embodiment performs the step of inputting the physical coupling factor into a fully connected layer for dimensionality transformation to obtain the physical enhancement residual vector.

[0119] Specifically, the fully connected layer mentioned here is a lightweight auxiliary network layer, whose weight parameter matrix is ​​denoted as... This layer serves two main purposes: first, it performs a linear mapping of the physical coupling factor into the feature space, aligning its semantic space with the main network; second, it adjusts the number of channels in the output tensor to match the number of output channels in the pointwise convolutional layers of depthwise separable convolutions, facilitating subsequent addition operations. The mathematical expression for this transformation process can be defined as:

[0120] ;

[0121] In the formula, Represents the physical enhancement residual vector of the output; is the bias term; Activation is the non-linear activation function, usually ReLU or Swish functions are chosen to increase the non-linear expressive power of the model.

[0122] Finally, this embodiment performs the step of adding the physically enhanced residual vector element-wise to the output vector of the pointwise convolutional layer in the depthwise separable convolution to obtain the fused feature vector containing physical coupling information. This is a typical residual-connection structure, similar to the skip connection design in ResNet.

[0123] Specifically, let the output of the pointwise convolutional layer in a depthwise separable convolution be... The final fused feature vector The calculation formula is:

[0124] ;

[0125] In the formula, the symbol This represents the element-wise addition of tensors. Through this addition operation, cross-modal coupling features forcibly extracted by physical mechanisms are injected into the main feature stream of the network.

[0126] The brilliance of this design lies in the fact that it does not disrupt the original lightweight network framework. The main network still employs efficient depthwise separable convolutions, ensuring inference speed. Meanwhile, the physical enhancement branches undergo sparsification filtering of the topology graph, performing second-order operations only on a few key physical channel pairs. The increased computational cost is negligible compared to the attention mechanism across all channels.

[0127] For example, in a practical yaw slip ring monitoring scenario, suppose the slip ring experiences an early microscopic arc erosion fault. This fault might appear as extremely weak clutter in a single current signal, easily filtered out by conventional noise reduction algorithms; in a vibration signal, it would only manifest as slight high-frequency jitter. However, physics tells us that the generation of an arc is inevitably accompanied by instantaneous high temperature and vibration caused by plasma impact. If a traditional convolutional network is used, it is difficult to correlate these two features at a shallow layer because they are both very weak and distributed in different channels.

[0128] However, in the architecture of this embodiment, since a physical coupling edge is pre-defined between current and vibration, the reinforcement branch will be forced to calculate. Since both signals exhibit non-zero abrupt changes at the moment the arc occurs, their product produces a significant spike signal. This spike signal, after being amplified by the fully connected layer, is superimposed on the main feature as a residual, effectively highlighting the fault feature deep within the network and forcing the network to pay attention to this anomaly.

[0129] like Figure 5 The simulation response results using physical coupling factors for feature enhancement are shown, where the horizontal axis of all subplots represents the sampling time, and the vertical axis represents the normalized signal amplitude or coupling strength value.

[0130] Figure 5 The blue curve at the top center represents the weak current noise signal collected by the sensor. When the fault occurs around 0.5 seconds later, its waveform only shows a very low amplitude bulge, which is easily drowned out by the background noise and difficult to detect.

[0131] Figure 5 The red curve in the middle represents the mechanical vibration signal acquired synchronously. At the corresponding moment, it only shows slight high-frequency jitter characteristics, which are also difficult to be independently identified as fault symptoms by conventional algorithms.

[0132] However, Figure 5The green curve below illustrates the waveform of the physical coupling factor generated after the second-order interactive operation guided by the physical topology. It can be clearly seen that during non-fault periods, because there is no physical correlation between the current and vibration signal, their product is suppressed and approaches zero. However, at the moment of a fault, due to the strong physical causality between the high temperature generated by the arc and the impact vibration, the second-order operation between the two produces a significant spike signal, directionally amplifying the originally vague and weak cross-modal fault features in the feature space. This result strongly demonstrates that the method described in this invention can effectively illuminate weak fault features in deep networks using physical mechanisms, thereby significantly improving the sensitivity and recognition accuracy of the monitoring system for early complex faults.

[0133] In summary, this embodiment successfully achieved accurate extraction of cross-modal strongly coupled features in a lightweight convolutional neural network by constructing a channel interaction enhancement mechanism guided by physical topology. It utilizes explicit physical causal relationships to compensate for the blind spots of purely data-driven models, providing solid algorithmic support for the early fault warning capability of the entire monitoring system.

[0134] Example 3:

[0135] Building upon Examples 1 and 2, this embodiment further refines a proactive load reduction and life extension strategy based on process parameter sensitivity, specifically addressing the most challenging production and maintenance deadlock scenarios in industrial settings. This strategy is specifically designed to resolve the extreme conflict situation described in Example 1, where the manufacturing execution system reports no production idle window of sufficient duration within the decay warning time window, and the current production task is of a high-priority, uninterruptible type. This embodiment achieves a technological leap from passively waiting for maintenance windows to proactively creating life extension opportunities by establishing a differential sensitivity model between health decay rate and process parameters.

[0136] In the actual operation of the monitoring system, a dilemma often arises. On the one hand, based on the multi-dimensional health assessment model in Example 1, the system has clearly determined that the overall health score of the yaw collector ring is below the health alarm threshold, and the decay warning time window predicted by the time series regression model indicates that the equipment will completely fail in the near future (e.g., within the next 48 hours). On the other hand, when the monitoring server interacts with the manufacturing execution system, the feedback indicates that it is currently in a peak production period or a critical order delivery period, and the production idle window time for the next week is zero.

[0137] In this situation, if the maintenance order is forcibly executed, the entire production line will be shut down or the wind turbine generator will be disconnected from the grid, resulting in incalculable economic losses and liability for breach of contract. If the current status is maintained and operation continues, the yaw collector ring will inevitably suffer a catastrophic failure before the production task is completed, such as ring surface ablation, short circuit and fire, leading to equipment scrapping and longer unplanned downtime.

[0138] To break this temporal deadlock, this embodiment introduces a key judgment and processing logic before generating maintenance instructions: executing a proactive load reduction and life extension strategy. The core idea of ​​this strategy is not to try to change the physical end of the equipment's lifespan, but to reduce the rate at which the equipment's lifespan is depleted by changing the physical environment in which the equipment operates, thereby forcibly postponing the equipment's failure until after the production task is completed.

[0139] In order to scientifically implement load reduction, it is necessary to first determine what to reduce and by how much. Therefore, this embodiment performs the step of constructing a sensitivity model of the healthy decay rate of the yaw collector ring to production process parameters.

[0140] Specifically, the manufacturing process parameters refer to the control variables that directly control the behavior of the yaw slip ring. These parameters include at least yaw speed, actuation acceleration, and braking torque. Yaw speed determines the relative sliding speed between the carbon brush and the slip ring, directly affecting the rate of frictional heat generation; actuation acceleration determines the impact force during yaw start-up and stop, directly affecting the micro-vibration amplitude of the carbon brush within the brush holder; and braking torque determines the braking force at the end of yaw, with excessive torque leading to a surge in instantaneous mechanical stress.

[0141] Based on massive amounts of historically accumulated operational data, the monitoring server uses multiple regression analysis or gradient boosting tree algorithm to establish a functional mapping relationship between the aforementioned process parameters and the equipment health degradation rate. To quantify this relationship, this embodiment defines a sensitivity index.

[0142] For example, suppose the rate of health decay at a certain moment is... It is defined as the decrease in the overall health score per unit time. Let the production process parameter vector be... This includes yaw speed. Motion acceleration and braking torque Equal components. The sensitivity model is designed to calculate the rate of health decay. The partial derivatives with respect to each component of the process parameter. Their mathematical expression is as follows:

[0143] ;

[0144] In the formula, Representing the Sensitivity coefficient of each process parameter; Representing the A specific process parameter value. The sensitivity coefficient is calculated. The system can identify which parameter is the culprit causing the deterioration of the slip ring's health under the current operating conditions. For example, if the calculation results show that the sensitivity coefficient of the acceleration is much greater than the yaw speed, it indicates that the vibration during the current acceleration and deceleration process is the main cause of accelerated wear. Therefore, acceleration should be limited first in the subsequent load reduction strategy.

[0145] like Figure 6 The graph displays the output results of a sensitivity model built based on historical operational big data. The horizontal axis of the graph represents the numerical value of the sensitivity coefficient, and the vertical axis represents the types of key production process parameters that affect the health status of the yaw collector ring.

[0146] Figure 6 The horizontal bars of varying lengths clearly quantify the marginal contribution of each process parameter to the rate of equipment health degradation. Longer bars indicate that even small changes in that parameter can lead to dramatic fluctuations in the rate of health decline. Based on... Figure 6 The specific values ​​shown are most significant for the bar length representing motion acceleration, with a sensitivity coefficient of 0.88, which is significantly higher than the braking torque of 0.55 and the yaw speed of 0.25.

[0147] This distribution pattern clearly indicates that under the current specific operating conditions, the impact vibration generated during yaw start-up and stop is the primary factor leading to increased wear of the slip ring. This provides a precise navigation basis for the system to implement an active load reduction and life extension strategy. In other words, the system should prioritize limiting the magnitude of the motion acceleration in multi-objective optimization calculations in order to achieve the maximum equipment life extension effect with the least sacrifice of production efficiency.

[0148] After determining the sensitivity of each parameter, the system needs to find a set of optimal operating parameters.

[0149] Specifically, this embodiment performs the following steps: while ensuring that the product quality indicators meet the preset standards, it performs multi-objective optimization calculations based on the sensitivity model to generate a set of low-damage process parameters that can minimize the health decay rate.

[0150] It is important to note that load reduction here cannot be carried out without limits; it must be subject to rigid constraints. For wind turbine generators, the aforementioned product quality indicators are mainly reflected in wind alignment accuracy and power generation efficiency. If the yaw response speed is excessively reduced to protect the collector ring, causing the unit to fail to align with the wind direction in time, it will result in a significant decrease in power generation, which contradicts the original intention of production. Therefore, the optimization calculation is performed within a constrained solution space.

[0151] This embodiment employs Sequential Quadratic Programming (SQP) or Particle Swarm Optimization (PSO) to solve this constrained optimization problem. The objective function is set to minimize the corrected health decay rate. :

[0152] ;

[0153] The constraints are:

[0154] ;

[0155] In the formula, The set of low-damage process parameters to be solved; For the product quality indicators expected under this parameter set, such as average power generation; This is the preset minimum acceptable quality standard.

[0156] After iterative calculations, the system outputs a set of specific parameter values, namely the low-damage process parameter set. For example, the yaw speed is reduced from 0.5 degrees / second to 0.3 degrees / second, and the starting acceleration is reduced from 5 degrees / square second to 2 degrees / square second. Although this set of parameters makes the yaw action slower, it can theoretically significantly reduce the wear rate of carbon brushes and the temperature rise rate of the contact surface while meeting basic power generation requirements.

[0157] After obtaining the low-damage parameters, the system cannot be executed immediately. Simulation verification must be performed first to confirm whether these parameters can truly help the equipment survive until the task is completed.

[0158] Specifically, this embodiment performs the step of correcting the time series regression model based on the low-damage process parameter set and re-predicting the expected failure time of the equipment after the load reduction condition is implemented.

[0159] In Example 1, the time series regression model (such as LSTM) is trained and predicted based on historical operating conditions, i.e., normal high-load operating conditions. If the original model is used directly, it cannot reflect the lifespan extension effect after load reduction. Therefore, a correction factor is introduced into the system. This factor represents the expected reduction in the decay rate, calculated based on the aforementioned sensitivity model. The revised prediction logic can be expressed as:

[0160] ;

[0161] In the formula, This represents the re-predicted expected time of equipment failure. Represents the current moment; This represents the current overall health score; The critical threshold representing the complete failure of the equipment; This represents the current rate of health decay. The predicted failure time is obtained by reducing the decay rate in the denominator. The prediction will be delayed relative to the original value. The system uses this logic, combined with the long short-term memory capability of LSTM, to generate a new health decline curve with a gentler slope, and uses this to pinpoint the new failure time.

[0162] Finally, the system makes a final decision based on the results of the re-prediction.

[0163] Specifically, this embodiment performs the step of comparing the expected failure time of the equipment with the expected end time of the current uninterrupted production task. Let the expected end time of the current uninterrupted production task be... This timeframe is typically calculated by the Manufacturing Execution System (MES) based on order volume and production cycle time.

[0164] The system makes the following logical judgment: If ,in To allow for a safety buffer period, such as 2 hours, the proactive load reduction and life extension strategy is deemed feasible. At this point, the monitoring server sends the low-damage process parameter set to the underlying programmable logic controller (PLC) or the fan main control system to trigger the load reduction operation mode. After receiving the instruction, the main control system dynamically modifies the PID control parameters of the yaw drive to limit the maximum output torque and maximum speed of the motor, allowing the equipment to enter a gentle operating state.

[0165] like Figure 7 The results show the predicted changes in equipment health after the system implements an active load reduction and life extension strategy in the extreme scenario of production and maintenance time-space deadlock. The horizontal axis represents the future prediction time in hours, and the vertical axis represents the comprehensive health score of the yaw collector ring.

[0166] Figure 7 The red dashed line represents the health degradation trajectory under the original high-load conditions. It can be seen that the downward slope is relatively steep. The prediction results show that the equipment will fall below the critical failure threshold of 40 points in the 24th hour. At this time, there is still a 6-hour time difference before the end of the uninterrupted production task in the 30th hour. This indicates that if no intervention measures are taken, the equipment will be completely destroyed in the middle of the task execution.

[0167] In comparison, Figure 7 The solid green line represents the health evolution trajectory after the system issued low-damage process parameters such as reducing yaw speed and limiting motion acceleration. Due to the significant reduction in health decay rate, the curve shows a more gradual downward trend, and its intersection with the critical failure threshold was successfully postponed to the thirty-eighth hour.

[0168] This significant time delay clearly demonstrates that, after proactive load reduction adjustment, the equipment's expected remaining lifespan not only fully covers the entire production task cycle but also reserves sufficient safety buffer time, thus powerfully verifying in a visual way the effectiveness of this invention in breaking rigid production constraints through a space-for-time strategy.

[0169] Simultaneously, the monitoring server modified its original maintenance plan, postponing the execution time of the maintenance instructions until after the expected end time. That is, in At the moment a production task is completed, the system immediately triggers an emergency shutdown for maintenance. Since there is no longer any production pressure, the maintenance work can be carried out at a leisurely pace.

[0170] It should also be noted that if the comparison results show that even if a load reduction strategy is adopted, Still earlier than ,or Although later than However, this is insufficient to cover the safety buffer time. If this condition is not met, it indicates that the equipment is critically damaged and no load reduction measures can sustain it until the mission ends. In this extreme situation, the system will terminate the load reduction attempt and immediately send a highest-level shutdown risk confirmation letter to the factory management, forcing manual intervention to make a decision: either shut down immediately to minimize losses or accept the risk of complete equipment failure.

[0171] Through the steps described above, this embodiment constructs a closed-loop process-health feedback adjustment mechanism. It no longer treats equipment health as a variable that can only be passively monitored, but rather as a control objective that can be actively managed by adjusting the production process. This proactive load reduction and lifespan extension method based on a sensitivity model, without increasing hardware costs, utilizes algorithms to extract the remaining value of the equipment, cleverly resolving the sharp contradiction between rigid tasks and equipment lifespan in industrial production, and greatly improving the system's intelligence level and emergency response capabilities.

[0172] Example 4:

[0173] like Figure 8 As shown, this embodiment provides a multi-dimensional status monitoring system based on a yaw collector ring. This system is configured to execute any one of the methods described in Embodiments 1 to 3, aiming to solve the technical problems of existing monitoring equipment when facing complex operating conditions of yaw collector rings, such as difficulty in extracting fault features, high false alarm rates, and rigid production and maintenance scheduling, through deep hardware and software collaboration. The system described in this embodiment not only possesses efficient computing capabilities at the edge but also enables data interaction and decision-making linkage with factory-level manufacturing execution systems.

[0174] Specifically, the system described in this embodiment mainly consists of four core functional units in its logical architecture: a data acquisition and preprocessing module, a feature fusion and analysis module, a status monitoring and evaluation module, and a decision-making and maintenance response module. These modules are connected via a high-speed data bus and run on a monitoring server or edge computing gateway equipped with a high-performance processor.

[0175] First, the data acquisition and preprocessing module serves as the system's sensing front-end, primarily addressing the issue of a single, easily interfered data source mentioned in the background section. This module physically connects to multiple sensor groups deployed at the yaw collector ring site, including but not limited to current sensors for monitoring electrical contact, ultrasonic sensors for monitoring mechanical wear, and rotary encoders for monitoring attitude. To ensure the reliability of the raw data, this module is configured to execute redundant verification logic, simultaneously reading data from heterogeneous sensors of the same type and performing consistency comparisons. More importantly, this module incorporates a noise suppression matrix built upon physical mechanisms. Addressing the false alarms caused by environmental interference, such as sudden temperature changes due to direct sunlight, as described in the background section, this module utilizes the noise suppression matrix to denoise the acquired raw operating dataset of the yaw collector ring. Based on the intrinsic correlation thresholds between physical quantities such as vibration and yaw angle, temperature and current, it filters out isolated fluctuations and cross-modal anomalies that violate physical laws in real time, thus providing a clean, high-quality data stream for subsequent feature analysis.

[0176] Secondly, the feature fusion and analysis module is the computational core of the system, primarily responsible for extracting key state information from massive amounts of data. Addressing the technical challenge of existing lightweight models failing to capture cross-modal strongly coupled features, this module integrates a convolutional neural network employing a deep separable convolutional structure. This network goes beyond conventional lightweight design, further incorporating the physical topology guidance mechanism described in Example 2. Specifically, this module is configured to generate fused feature vectors. During computation, it performs second-order interactive operations on strongly correlated channels such as current and vibration based on preset physical coupling edges, thereby accurately capturing early, subtle signs of composite faults. Furthermore, to solve the evaluation challenges under varying operating conditions, this module is also responsible for calculating dynamic weights. Based on stored historical fault dimension base factors, combined with real-time received operating condition parameters such as load rate and ambient temperature, it dynamically updates the weight coefficients of different monitoring dimensions using an exponential moving average algorithm, ensuring that the evaluation criteria can adaptively adjust with changes in the operating environment.

[0177] Furthermore, the status monitoring and assessment module serves as the system's diagnostic center, quantifying the health level of the equipment. This module stores a graded calibration benchmark library, which includes standard operating condition sub-benchmarks under different load and temperature combinations. During operation, the module compares the frequency domain correlation coefficient and relative deviation of the real-time generated fused feature vector with the sub-benchmark corresponding to the current operating condition, thereby accurately identifying abnormal features. Subsequently, the module combines the aforementioned calculated dynamic weights and the current operating condition attenuation coefficient to calculate the electrical, mechanical, and attitude health across dimensions, and finally synthesizes a comprehensive health score. This mechanism overcomes the limitations of traditional single-threshold alarms, achieving refined, multi-dimensional quantitative grading of equipment health status.

[0178] Finally, the decision-making and maintenance response module acts as the system's execution brain, responsible for resolving the spatiotemporal conflicts between production tasks and equipment maintenance. When the comprehensive health score output by the status monitoring and evaluation module falls below the preset health alarm threshold, this module immediately activates its built-in time series regression model, such as a long short-term memory network, to predict the future health change curve of the equipment and output a decay warning time window accordingly. At this time, the module actively queries the production scheduling plan of the manufacturing execution system through the communication interface, extracting the low-capacity periods and changeover intervals as production idle windows. If an idle window of sufficient duration is matched within the warning time window, the module will automatically generate a maintenance instruction containing the optimal maintenance time and resource scheduling scheme.

[0179] It is particularly important to emphasize that, for extreme deadlock scenarios where there are no available windows for configuration and production tasks cannot be interrupted, this decision-making and maintenance response module also integrates the proactive load reduction and life extension logic described in Example 3. When the system determines that it is impossible to schedule downtime maintenance before equipment failure, this module will call a preset sensitivity model and generate a set of low-damage process parameters based on a multi-objective optimization algorithm, such as reducing yaw speed or limiting torque. Subsequently, this module sends a command to the underlying control system to trigger a load reduction operation mode, which delays the rate of health degradation of the equipment by sacrificing some non-critical process efficiency, thereby forcibly postponing the expected failure time of the equipment until after the end of the production task, achieving equipment safety management under rigid production constraints.

[0180] In summary, the multi-dimensional status monitoring system based on yaw collector ring provided in this embodiment achieves multi-source sensing and edge computing at the hardware level, the fusion of physical mechanisms and deep learning at the software level, and closed-loop linkage of status monitoring and production scheduling at the business level through the close collaboration of the above four modules, thus comprehensively achieving a highly reliable and intelligent industrial equipment operation and maintenance solution.

[0181] Example 5:

[0182] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0183] like Figure 9 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0184] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0185] Bus 102 may include a path for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0186] The memory 103 stores a computer program corresponding to the multi-dimensional state operation monitoring method based on a yaw collector ring according to the above embodiments of the present invention. This computer program is executed under the control of the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0187] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 9 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0188] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-dimensional state operation monitoring method based on a yaw collector ring, characterized in that, Applications include monitoring servers, including: Obtain the raw operating dataset of the yaw collector ring, which includes operating parameters in electrical, mechanical, attitude and environmental dimensions collected by sensors and verified for redundancy. A noise suppression matrix is ​​constructed based on the physical mechanism of the yaw collector ring. The noise suppression matrix is ​​used to filter isolated fluctuations and cross-modal outliers in the original running dataset and input into the convolutional neural network model to obtain a fused feature vector. Based on the basic factors of the historical fault dimension and the real-time factors of the real-time operating condition dimension, dynamic weights for different monitoring dimensions are calculated, and the dynamic weights are smoothly updated using the exponential moving average method. Based on the current load and temperature conditions, the corresponding operating condition sub-reference is matched from the graded calibration reference library, and the frequency domain correlation coefficient and relative deviation between the fused feature vector and the operating condition sub-reference are calculated to mark abnormal features. Combining the dynamic weights and the abnormal features, the equipment health is calculated in different dimensions, and a comprehensive health score is generated by combining the operating condition attenuation coefficient. In response to the overall health score falling below the health alarm threshold, a time series regression model is used to predict the decay warning time window, and maintenance instructions are generated in conjunction with the idle window of the manufacturing execution system.

2. The method according to claim 1, characterized in that, Before obtaining the raw operational dataset of the yaw collector loop, the process also includes determining the monitoring item acquisition strategy: Obtain historical fault comparison data, which includes the number of historical faults of similar equipment, the maximum single maintenance cost, and the maximum downtime. Calculate the monitoring priority value based on the historical fault analogy data, and sort the monitoring items according to the monitoring priority value; The error correction value is calculated based on the historical acquisition error rate, and the data acquisition frequency of different monitoring items is set according to the numerical range of the error correction value. The monitoring priority value is positively correlated with the number of historical faults and has a functional relationship with the sum of the maximum single maintenance cost and the maximum downtime.

3. The method according to claim 1, characterized in that, The noise suppression matrix is ​​constructed based on the physical mechanism of the yaw collector ring. This noise suppression matrix is ​​then used to filter isolated fluctuations and cross-modal outliers in the original running dataset, including: Construct a parameter correlation matrix that includes vibration parameters, yaw angle parameters, temperature parameters, and current parameters; In the parameter correlation matrix, a physical correlation threshold is set, which includes a positive correlation coefficient threshold between vibration and yaw angle, and a linear correlation coefficient threshold between temperature and current. The fluctuation of a single parameter in the original running dataset is detected. If a single parameter fluctuates and the change of its associated parameter in the parameter correlation matrix does not reach the physical correlation threshold, the fluctuation of the single parameter is determined to be noise and filtered out. Based on the aforementioned working condition sub-reference, a correlation threshold between wear rate and vibration amplitude is set, and isolated data points exceeding the correlation threshold are removed.

4. The method according to claim 1, characterized in that, The input to the convolutional neural network model yields a fused feature vector, including: The filtered operating parameters are categorized into electrical, mechanical, and attitude types, and then input into attention networks built for specific modalities. The attention network is used to calculate the weight of each feature and to select the core features whose weight ranking meets the preset ratio. The core features include at least the effective value of current, wear rate, yaw angle change trend and vibration main frequency amplitude. The core features are input into a convolutional neural network with a depthwise separable convolutional structure, and the fused feature vector is output.

5. The method according to claim 4, characterized in that, The process of inputting the core features into a convolutional neural network employing a depthwise separable convolutional structure and outputting the fused feature vector specifically includes: Construct a physical topology graph of the feature channels, map each feature channel in the core features to a node in the physical topology graph, and establish physical coupling edges between nodes with causal relationships based on the heat generation mechanism and vibration transmission mechanism of the yaw collector ring. At the output of the depthwise convolutional layer of the convolutional neural network, a physical topology-guided channel interaction enhancement operation is performed in parallel. The channel interaction enhancement operation includes: extracting paired associated feature channels based on the physical coupling edges; performing a second-order interaction operation based on element-wise multiplication on each pair of associated feature channels to generate a physical coupling factor characterizing the cross-modal nonlinear relationship; inputting the physical coupling factor into a fully connected layer for dimensionality transformation to obtain a physical enhancement residual vector; and adding the physical enhancement residual vector element-wise to the output vector of the pointwise convolutional layer in the depthwise separable convolution to obtain the fused feature vector containing physical coupling information.

6. The method according to claim 2, characterized in that, The calculation of dynamic weights for different monitoring dimensions based on the basic factors of historical fault dimensions and the real-time factors of real-time operating conditions includes: The basic factor is obtained by normalizing the monitoring priority value; Based on a pre-built mapping table of working conditions and parameter contributions, the real-time factors are obtained by matching the current real-time working conditions. The base factor and the real-time factor are weighted and summed according to a preset proportional coefficient to obtain the real-time weight; Using the exponential moving average method, the final dynamic weight is calculated based on the currently calculated real-time weight and the weight of the previous cycle.

7. The method according to claim 1, characterized in that, The process of combining the dynamic weights and the abnormal features to calculate the equipment health across dimensions, and generating a comprehensive health score by combining the operating condition attenuation coefficient, includes: Calculate electrical health, mechanical health, and posture health separately; The electrical health is calculated based on the deviation of the effective value of the current and the degree of exceedance of harmonic components; the mechanical health is calculated based on the wear rate, vibration amplitude and dynamic damping ratio; and the attitude health is calculated based on the yaw angle deviation and speed fluctuation. The attenuation coefficient for the operating condition is determined based on the current load rate and ambient temperature. The electrical health, mechanical health, and attitude health are weighted and summed according to their respective dynamic weights, and then multiplied by the operating condition attenuation coefficient to obtain the comprehensive health score.

8. The method according to claim 1, characterized in that, In response to the overall health score falling below the health alarm threshold, a maintenance instruction is generated by using a time series regression model to predict the decay warning time window and combining it with the idle window of the manufacturing execution system, including: The comprehensive health score is input into a long short-term memory network model to predict the health change curve within a preset time period in the future. The time interval within which the health status change curve drops to the critical failure threshold is determined as the decay warning time window; Extract production idle windows from the manufacturing execution system, the production idle windows including low capacity periods and changeover intervals; Within the attenuation warning time window, the longest idle production window whose duration meets the maintenance requirements is matched, marked as the optimal maintenance window, and the maintenance instruction containing maintenance type and resource scheduling information is generated.

9. The method according to claim 8, characterized in that, The process of generating maintenance instructions includes: A sensitivity model of the health decay rate of the yaw collector ring to production process parameters is constructed, wherein the production process parameters include at least yaw speed, actuation acceleration, and braking torque. While ensuring that the product quality indicators meet the preset standards, a multi-objective optimization calculation is performed based on the sensitivity model to generate a set of low-damage process parameters that can minimize the health decay rate. The time series regression model is modified based on the low-damage process parameter set to re-predict the expected failure time of the equipment after the load reduction condition is implemented. If the expected failure time of the equipment is compared with the expected end time of the current uninterrupted production task, and the expected failure time of the equipment is later than the expected end time, the low-damage process parameter set is sent to the control system to trigger the load reduction operation mode, and the execution time of the maintenance command is postponed to after the expected end time.

10. A multi-dimensional state operation monitoring system based on a yaw collector ring, characterized in that, The system applied to the multi-dimensional state operation monitoring method based on yaw collector ring according to any one of claims 1-9, the system comprising: The data acquisition and preprocessing module is used to acquire the raw operating dataset of the yaw collector ring and perform denoising processing based on the noise suppression matrix constructed based on the physical mechanism. The feature fusion and analysis module is used to generate fused feature vectors by using a convolutional neural network with a deep separable convolutional structure, and to calculate dynamic weights based on historical fault dimensions and real-time operating condition dimensions. The status monitoring and assessment module is used to compare the fused feature vector based on the working condition sub-benchmark to mark abnormal features, and to calculate the multi-dimensional health score and the comprehensive health score in combination with the dynamic weight. The decision and maintenance response module is used to generate maintenance instructions by using a prediction model to output a decay warning time window when the overall health score is lower than the health alarm threshold, and combining it with the idle window of the production execution system.

Images