A method and system for monitoring a processing line of a fan controller

Abstract

This invention relates to the field of data processing, specifically to a method and system for monitoring a production line for wind turbine controllers. The method includes: acquiring multimodal time-series data; constructing a production line topology diagram; processing the data using a spatiotemporal graph convolutional network to extract spatiotemporal correlation features; real-time determination of whether the production line exhibits abnormal conditions; calculating the gradient contribution of the production line health entropy to the input features of each workstation, and locating the workstation with the highest contribution as the root cause of the abnormality; inputting the spatiotemporal correlation features into a trajectory prediction module based on Kalman filtering and attention mechanisms to generate an evolution trajectory of the production line health entropy within a future preset time window. This evolution trajectory includes predicted values ​​and confidence intervals, used for early warning of potential production line deterioration. The method of this invention can effectively predict systemic risks in wind turbine controller processing lines, improving the efficiency and accuracy of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of data processing. More specifically, this invention relates to a method and system for monitoring a production line using a fan controller. Background Technology

[0002] As the core control unit of wind turbine generator sets, wind turbine controllers are manufactured with a high degree of automation and complex processes, placing extremely stringent requirements on product quality and production line stability. Currently, production line monitoring for high-end electronic equipment like wind turbine controllers mainly relies on Manufacturing Execution Systems (MES) to record and statistically analyze data such as production cycle time, material consumption, and output at each workstation, or traditional Statistical Process Control (SPC) methods to monitor key physical parameters such as temperature and torque at individual workstations by setting static thresholds. Some production lines also deploy industrial cameras for video monitoring, but the acquired image data is usually only used for post-production manual quality traceability or isolated visual inspection of specific defects. These methods generally suffer from insufficient data utilization; data sources from different modalities such as sensors, MES systems, and video form "information silos" that are not effectively integrated to provide a comprehensive and multi-dimensional understanding of the production line status. At the same time, they treat each processing station as an independent entity for analysis, ignoring the close spatial correlation and dynamic temporal dependence between stations formed by the process flow, making it difficult to capture the chain anomalies caused by the transmission and accumulation of small deviations from upstream stations to downstream.

[0003] Due to the aforementioned limitations, existing monitoring technologies fall short in handling complex changes in production line status. Furthermore, regarding anomaly detection, the lack of a comprehensive indicator that can quantify the overall health level of the production line makes it difficult for systems to identify systemic risks caused by the collaborative actions of multiple workstations that are still within the tolerance range of a single parameter. Often, a reactive response is only made after serious quality defects or equipment downtime occur, resulting in irreversible losses. Secondly, in terms of anomaly tracing, when an alarm is triggered, locating the true root cause of the problem relies heavily on the personal experience of engineers and step-by-step troubleshooting, a time-consuming and inefficient process. There is a lack of intelligent means to automatically and accurately locate the source of anomalies from massive amounts of data. Finally, in terms of risk warning, existing technologies are mostly reactive, lacking the ability to proactively predict future trends in production line status. They cannot anticipate potential performance degradation risks and provide early warnings, causing companies to miss the best opportunity to take preventative maintenance measures and avoid production interruptions.

[0004] In summary, existing monitoring methods for wind turbine controller manufacturing lines suffer from several drawbacks: they cannot provide a comprehensive and holistic understanding of the line's status, cannot predict systemic risks associated with the wind turbine controller manufacturing line, and have low efficiency and accuracy in troubleshooting. Summary of the Invention

[0005] To address the problems of existing monitoring methods for wind turbine controller processing lines, such as the inability to provide a comprehensive and three-dimensional understanding of the line status, the inability to predict systemic risks in the wind turbine controller processing line, and the low efficiency and accuracy of fault diagnosis, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a method for monitoring a production line for a fan controller, comprising:

[0007] Real-time physical parameters, manufacturing execution system data, and workstation video image data of each processing station on the production line are acquired to form multimodal time series data; and a production line topology diagram with processing stations as nodes and process flow relationships as edges is constructed. The data is processed using a spatiotemporal graph convolutional network to extract spatiotemporal correlation features that integrate the spatial dependence and temporal evolution of workstations.

[0008] Based on the spatiotemporal correlation features, the production line health entropy, which characterizes the current operational stability of the production line, is calculated, and by comparing it with a preset health baseline, it is determined in real time whether there is an abnormal state in the production line.

[0009] When an abnormal state is detected in the production line, the reverse gradient propagation algorithm is used to locate the abnormal root cause workstation by calculating the gradient contribution of the production line health entropy to the input features of each workstation.

[0010] The spatiotemporal correlation features are input into a trajectory prediction module based on Kalman filtering and attention mechanism to generate the evolution trajectory of the production line health entropy within a future preset time window. The evolution trajectory includes predicted values ​​and confidence intervals, which are used to provide early warning of potential deterioration of the production line status.

[0011] Preferably, the acquisition of real-time physical parameters, manufacturing execution system data, and workstation video image data for each processing station on the production line to form multimodal time-series data includes:

[0012] The system collects physical parameters of each workstation through an industrial controller; acquires manufacturing data associated with production batches and workpieces through a manufacturing execution system; captures image data of workstation operations through an industrial camera; and aligns the physical parameters, manufacturing data, and image data in time using a unified timestamp to form multimodal time-series data.

[0013] Preferably, the spatiotemporal graph convolutional network is composed of alternately stacked graph convolutional modules and temporal convolutional modules; wherein, the graph convolutional module is used to aggregate the information of adjacent workstation nodes in the production line topology graph at each time step to capture spatial dependencies; the temporal convolutional module is used to extract the dynamic change pattern of its own multimodal time series data at each workstation node to capture temporal dependencies.

[0014] Preferably, the calculation of the production line health entropy, which characterizes the current operational stability of the production line, includes:

[0015] Clustering algorithms are used to group the spatiotemporal correlation feature vectors of all workstations within the current time window to obtain multiple clusters; the proportion of feature vectors in each cluster to the total number is counted to form a discrete probability distribution; Shannon entropy is calculated based on the probability distribution as the health entropy of the production line.

[0016] Preferably, the calculation of the production line health entropy, which characterizes the current operational stability of the production line, includes: inputting the extracted spatiotemporal correlation features into a multilayer perceptron, whose output layer uses the Softmax function to calculate the probability distribution of the production line in various preset states, the types of which include normal, slight deviation and severe abnormality; and then calculating the information entropy of the probability distribution according to the Shannon entropy formula to obtain a single scalar value, namely the production line health entropy.

[0017] Preferably, the health baseline is determined based on the statistical distribution of production line health entropy calculated over multiple historical normal operation cycles; when the real-time calculated production line health entropy is higher than a certain proportion of the health baseline, it is determined that the production line is in an abnormal state.

[0018] Preferably, real-time determination of whether there is an abnormal state on the production line includes:

[0019] During the normal operation cycle of the production line, a large number of health entropy values ​​are collected, and their mean and standard deviation are obtained by fitting a Gaussian distribution. The mean plus three times the standard deviation is set as a dynamic threshold for anomaly judgment. If the health entropy of the production line calculated in real time exceeds the dynamic threshold, it is judged as an anomaly.

[0020] Preferably, the workstations with the highest location contribution to the root cause of the anomaly include:

[0021] Using the production line health entropy as the objective function, backpropagation is performed along the spatiotemporal graph convolutional network to calculate the gradient of the health entropy with respect to the input features of each workstation within a preset time window before the anomaly occurs. The absolute values ​​of the gradients of all input features at each workstation are accumulated to obtain the total gradient contribution of that workstation. The total gradient contributions of all workstations are sorted, and the workstation with the highest contribution is identified as the root cause workstation of the anomaly.

[0022] Preferably, generating the evolution trajectory of the production line health entropy within a future preset time window includes:

[0023] An attention mechanism is used to process the historical spatiotemporal correlation feature sequence to assign weights to the spatiotemporal correlation features at each time step;

[0024] The attention-weighted feature sequence is input into the Kalman filter, and through iterative prediction and update steps, a sequence of predicted values ​​of the production line health entropy within a preset future time window is generated; and the confidence interval for each predicted value is calculated using the state covariance matrix of the Kalman filter.

[0025] In a second aspect, the present invention provides a production line monitoring system for a wind turbine controller, comprising a memory and a processor, wherein the memory stores computer program instructions, and the computer program instructions, when executed by the processor, implement the production line monitoring method for a wind turbine controller of the present invention.

[0026] The beneficial effects of this invention are as follows: Compared with existing technologies, this invention, by constructing a production line topology and utilizing a spatiotemporal graph convolutional network, deeply integrates multi-source heterogeneous information such as physical parameters, manufacturing execution system data, and video images. This overcomes the limitations of data silos and enables the capture of spatial dependencies and temporal evolution patterns between workstations from a global perspective, thereby gaining a more comprehensive and profound understanding of the production line's operational status. Based on this, the production line health entropy proposed in this invention, as a quantitative indicator, can accurately assess the overall stability of the production line and effectively identify systemic, early-stage anomalies caused by the collaborative influence of multiple workstations, which are difficult to detect using traditional methods. When an anomaly occurs, this invention can automatically and quickly locate the root cause workstation, significantly improving the efficiency and accuracy of troubleshooting. Furthermore, by predicting the future trend of health entropy, this invention achieves a shift from passive response to proactive prevention, providing forward-looking early warnings for potential production line deterioration, which helps to take preventative measures and ensure production continuity and product quality. Attached Figure Description

[0027] Figure 1 This is a schematic flowchart illustrating a production line monitoring method for a fan controller according to an embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram illustrating multimodal timing data according to an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram illustrating a spatiotemporal graph convolutional network according to an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram illustrating the production line health entropy and anomaly judgment according to an embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram illustrating the evolution trajectory of the production line health entropy generated according to an embodiment of the present invention;

[0032] Figure 6This is a schematic diagram illustrating the structure of a production line monitoring system for a fan controller according to an embodiment of the present invention. Detailed Implementation

[0033] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] Example of a production line monitoring method for fan controllers:

[0035] like Figure 1 As shown, the production line monitoring method for a fan controller according to the present invention includes:

[0036] S101. Obtain spatiotemporal correlation features using a spatiotemporal graph convolutional network, specifically as follows: Figure 2 As shown, real-time physical parameters, manufacturing execution system data, and workstation video image data of each processing station on the production line are obtained to form multimodal time series data; and a production line topology diagram with processing stations as nodes and process flow relationships as edges is constructed. The data is processed using a spatiotemporal graph convolutional network to extract spatiotemporal correlation features that integrate the spatial dependence and temporal evolution of workstations.

[0037] In this embodiment, acquiring real-time physical parameters, manufacturing execution system data, and workstation video image data of each processing station on the production line to form multimodal time-series data includes: collecting physical parameters of each workstation through an industrial controller; acquiring manufacturing data associated with production batches and workpieces through the manufacturing execution system; capturing image data of workstation operations through an industrial camera; and aligning the physical parameters, manufacturing data, and image data using a unified timestamp to form multimodal time-series data.

[0038] In this embodiment, Modbus or OPC UA protocols are used when acquiring physical parameters of each workstation through the industrial controller. Manufacturing data can be obtained from the Manufacturing Execution System via API interface. Image data of workstation operations captured by industrial cameras can be transmitted via RTSP protocol.

[0039] The collected physical parameters include chip mounter torque and reflow oven temperature. The collected manufacturing execution system data includes work order number and cycle time.

[0040] Based on the process flow, equipment such as pick-and-place machines, reflow soldering machines, and optical inspection machines are defined as graph nodes, and the material flow direction is defined as directed edges. An adjacency matrix is ​​generated to represent the production line topology.

[0041] Visual features are extracted from video data via a pre-trained ResNet network and input as node attributes along with other data into a composite spatiotemporal graph convolutional network model consisting of spatial graph convolutional layers and gated recurrent units. The spatiotemporal graph convolutional network model learns and outputs a feature vector that incorporates global spatiotemporal dependencies, which is used to characterize spatiotemporal correlation features.

[0042] like Figure 3 As shown, in this embodiment, the spatiotemporal graph convolutional network is composed of alternately stacked graph convolutional modules and temporal convolutional modules; wherein, the graph convolutional module is used to aggregate the information of adjacent workstation nodes in the production line topology graph at each time step to capture spatial dependencies; the temporal convolutional module is used to extract the dynamic change pattern of its own multimodal time series data at each workstation node to capture temporal dependencies.

[0043] Specifically, assume the production line includes three sequential workstations: loading, processing, and unloading. At time T, the graph convolution module processes the data of the processing workstation itself, while simultaneously integrating product arrival information from the upstream loading workstation and the idle status information of the downstream unloading workstation, reflecting spatial mutual influence. Subsequently, the temporal convolution module focuses on analyzing the changing trends of a series of data at the processing workstation from time T minus 10 seconds to time T, such as whether the temperature continues to rise, in order to capture dynamic features in the time dimension.

[0044] S102. Real-time determination of whether there is an abnormal state in the production line, specifically: based on the spatiotemporal correlation characteristics, calculate the production line health entropy that characterizes the current operating stability of the production line, and determine whether there is an abnormal state in the production line in real time by comparing it with a preset health baseline.

[0045] In this embodiment, the calculation of the production line health entropy, which characterizes the current operational stability of the production line, includes: using a clustering algorithm to group the spatiotemporal correlation feature vectors of all workstations within the current time window to obtain multiple clusters; counting the proportion of feature vectors in each cluster to the total number to form a discrete probability distribution; and calculating the Shannon entropy as the production line health entropy based on the probability distribution.

[0046] Specifically, in a production line with 10 workstations, the model generated a feature vector for each workstation over the past 5 minutes, resulting in 10 vectors. These 10 vectors were then divided into 3 clusters using the K-means clustering algorithm. If 8 vectors cluster in cluster A (representing stable operation), 1 vector in cluster B (representing slight fluctuations), and 1 vector in cluster C (representing anomalies), the resulting probability distribution is P(A) = 0.8, P(B) = 0.1, and P(C) = 0.1. The Shannon entropy calculated based on this distribution is low, for example, 0.92, indicating a highly consistent and stable production line operation. Conversely, if the vectors are evenly distributed across the 3 clusters, the entropy value will increase significantly, for example, approaching 1.58, indicating a chaotic or unstable production line.

[0047] In another embodiment, calculating the production line health entropy, which characterizes the current operational stability of the production line, includes: inputting extracted spatiotemporal correlation features into a multilayer perceptron, whose output layer uses a Softmax function to calculate the probability distribution of the production line in various preset states, the types of which include normal, slight deviation, and severe anomaly; then calculating the information entropy of the probability distribution according to the Shannon entropy formula to obtain a single scalar value, namely the production line health entropy. The smaller this single scalar value, the higher the certainty of the production line state and the more stable the operation.

[0048] In this embodiment, the health baseline is determined based on the statistical distribution of the production line health entropy calculated over multiple historical normal operation cycles; when the real-time calculated production line health entropy is higher than a certain percentage of the health baseline, it is determined that the production line is in an abnormal state.

[0049] In this embodiment, the method for calculating the health baseline includes: collecting a large number of health entropy values ​​during the normal operation cycle of the production line, and using the average value of the collected health entropy values ​​as the health baseline.

[0050] For example, by analyzing data from normal production line operation over the past month, the health entropy of the production line per minute is calculated, and the average value of these entropy values ​​is 0.9. Therefore, 0.9 is set as the health baseline.

[0051] In another embodiment, real-time determination of whether the production line is in an abnormal state includes: during the normal operation cycle of the production line, collecting a large number of health entropy values, obtaining their mean and standard deviation by fitting a Gaussian distribution, and setting the mean plus three times the standard deviation as a dynamic threshold for anomaly determination. If the real-time calculated health entropy of the production line exceeds the dynamic threshold, it is determined to be abnormal.

[0052] If the health baseline is 1 and the dynamic threshold is 2, then when the health entropy of the production line exceeds 2, it is determined that the production line has an anomaly.

[0053] S103. Locating the anomaly root cause workstation with the highest contribution: When it is determined that there is an abnormal state in the production line, the reverse gradient propagation algorithm is used to locate the anomaly root cause workstation with the highest contribution by calculating the gradient contribution of the production line health entropy to the input features of each workstation.

[0054] In this embodiment, as Figure 4 As shown, the workstation with the highest contribution to locating the root cause of the anomaly includes: using the production line health entropy as the objective function, backpropagating along the spatiotemporal graph convolutional network, calculating the gradient of the health entropy with respect to the input features of each workstation within a preset time window before the anomaly occurs; summing the absolute values ​​of the gradients of all input features at each workstation to obtain the total gradient contribution of that workstation; sorting the total gradient contributions of all workstations, and identifying the workstation with the highest contribution as the root cause of the anomaly. For example, if the fan controller processing line has three workstations: a pick-and-place machine, a reflow soldering machine, and an optical inspection machine, and the pick-and-place machine has the highest total gradient contribution, then the pick-and-place machine is identified as the root cause of the anomaly.

[0055] For example, when a sudden change in the production line health entropy from the normal 0.9 to an abnormal 1.8 is detected, this entropy value of 1.8 is taken as the target. The algorithm traces back to calculate the gradient of the impact of this entropy change on all input data within one minute prior to the anomaly. The calculation results might show that the sum of the absolute values ​​of the gradients of the pressure sensor data and motor current data at stamping station 3 with respect to the entropy change is 5.6, while the sum of the gradient contributions of other stations, such as loading station 1 and cutting station 2, is only 0.3 and 0.5, respectively. Since the total gradient contribution of station 3 (5.6) is much higher than that of other stations, it is identified as the root cause station causing the production line anomaly.

[0056] S104. Generate the evolution trajectory of the production line health entropy within a future preset time window. Specifically, the spatiotemporal correlation features are input into a trajectory prediction module based on Kalman filtering and attention mechanism to generate the evolution trajectory of the production line health entropy within a future preset time window. The evolution trajectory includes predicted values ​​and confidence intervals, which are used to provide early warning of potential deterioration of the production line status.

[0057] In this embodiment, the trajectory prediction module employs an encoder-decoder architecture based on a long short-term memory network and introduces an attention mechanism. The encoder receives a sequence of spatiotemporal correlation features from the past period, while the decoder generates predicted values ​​of future health entropy step-by-step. During decoding, the attention mechanism enables the decoder to focus on the most historically relevant spatiotemporal features when generating each prediction point. The decoder's output sequence is used as the observation value of a Kalman filter. Through the filter's prediction and update steps, the predicted trajectory is smoothed, and the covariance of the prediction error is estimated. This covariance is used to calculate the confidence interval for each prediction point, thereby forming an evolutionary trajectory that includes a central prediction line and upper and lower boundaries. In this embodiment, the generated evolutionary trajectory is as follows: Figure 5 As shown.

[0058] In this embodiment, generating the evolution trajectory of the production line health entropy within a future preset time window includes:

[0059] S201. An attention mechanism is used to process the historical spatiotemporal correlation feature sequence to assign weights to the spatiotemporal correlation features at each time step.

[0060] The attention mechanism can automatically learn the importance of spatiotemporal correlation features. It can dynamically calculate weights to assign higher importance to spatiotemporal correlation features that are more relevant to the prediction of the current production line status. Spatiotemporal correlation features that are more relevant to the prediction of the current production line status refer to spatiotemporal correlation features that have a greater impact on the prediction of the current production line status.

[0061] S202. Input the feature sequence weighted by the attention mechanism into the Kalman filter, and generate a sequence of predicted values ​​of the production line health entropy within a preset time window through iterative prediction and update steps; and use the state covariance matrix of the Kalman filter to calculate the confidence interval for each predicted value.

[0062] For a given predicted value y, the confidence interval is constructed as follows:

[0063] (1) Set confidence level ;

[0064] In this embodiment, It can be set to 95%.

[0065] (2) Obtain the corresponding standard normal distribution quantiles by looking up a table. .

[0066] (3) Calculate the confidence interval. The expression for the confidence interval is: ;

[0067] in, Let y be the variance corresponding to the predicted value y.

[0068] For example, to predict the production line health entropy for the next 10 minutes, feature data from the past hour is first analyzed. An attention mechanism might identify that a brief material supply delay that occurred 15 minutes ago has the greatest impact on the current state, thus assigning it a weight of 0.6, while the feature weight for a stable operating state 30 minutes ago is only 0.1. This weighted feature sequence is then fed into a Kalman filter, which predicts a health entropy of 0.95 for the first minute, 0.98 for the second minute, and so on, generating a 10-minute prediction sequence. Simultaneously, the filter provides a confidence interval for each prediction value; for example, a prediction of 0.95 for the first minute has a 95% confidence interval of 0.92 to 0.98, providing decision-makers with a quantitative assessment of the prediction's reliability.

[0069] Example of a production line monitoring system for fan controllers:

[0070] This invention also provides a production line monitoring system for fan controllers. For example... Figure 6 As shown, the processing line monitoring system for the wind turbine controller includes a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, a processing line monitoring method for the wind turbine controller described in the above embodiments is implemented.

[0071] The production line monitoring system for the wind turbine controller also includes other components well known to those skilled in the art, such as communication buses and communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0072] In the description of this specification, "multiple" and "several" mean at least two, such as two, three or more, etc., unless otherwise clearly defined.

[0073] While several embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Numerous modifications, variations, and alternatives will occur to those skilled in the art without departing from the concept and spirit of the present invention. It should be understood that various alternatives to the embodiments of the present invention described herein may be employed in practicing the present invention.

Claims

1. A production line monitoring method for a fan controller, characterized in that, include: Real-time physical parameters, manufacturing execution system data, and workstation video image data of each processing station on the production line are acquired to form multimodal time series data; and a production line topology diagram with processing stations as nodes and process flow relationships as edges is constructed. The data is processed using a spatiotemporal graph convolutional network to extract spatiotemporal correlation features that integrate the spatial dependence and temporal evolution of workstations. Based on the spatiotemporal correlation features, the production line health entropy, which characterizes the current operational stability of the production line, is calculated, and by comparing it with a preset health baseline, it is determined in real time whether there is an abnormal state in the production line. When an abnormal state is detected in the production line, the reverse gradient propagation algorithm is used to locate the abnormal root cause workstation by calculating the gradient contribution of the production line health entropy to the input features of each workstation. The spatiotemporal correlation features are input into a trajectory prediction module based on Kalman filtering and attention mechanism to generate the evolution trajectory of the production line health entropy within a future preset time window. The evolution trajectory includes predicted values ​​and confidence intervals, which are used to provide early warning of potential deterioration of the production line status. The calculation of the production line health entropy, which characterizes the current operational stability of the production line, includes: using a clustering algorithm to group the spatiotemporal correlation feature vectors of all workstations within the current time window to obtain multiple clusters; counting the proportion of feature vectors in each cluster to the total number to form a discrete probability distribution; and calculating the Shannon entropy as the production line health entropy based on the probability distribution. Alternatively, the calculation of the production line health entropy, which characterizes the current operational stability of the production line, includes: inputting the extracted spatiotemporal correlation features into a multilayer perceptron, whose output layer uses the Softmax function to calculate the probability distribution of the production line in various preset states, the types of which include normal, slight deviation and severe abnormality; and then calculating the information entropy of the probability distribution according to the Shannon entropy formula to obtain a single scalar value, namely the production line health entropy.

2. The production line monitoring method for a fan controller as described in claim 1, characterized in that, The acquisition of real-time physical parameters, manufacturing execution system data, and workstation video image data from each processing station on the production line constitutes multimodal time-series data, including: The system collects physical parameters of each workstation through an industrial controller; acquires manufacturing data associated with production batches and workpieces through a manufacturing execution system; captures image data of workstation operations through an industrial camera; and aligns the physical parameters, manufacturing data, and image data in time using a unified timestamp to form multimodal time-series data.

3. The production line monitoring method for a fan controller as described in claim 1, characterized in that, The spatiotemporal graph convolutional network consists of alternating stacked graph convolutional modules and temporal convolutional modules. The graph convolutional modules are used to aggregate information of adjacent workstation nodes in the production line topology graph at each time step to capture spatial dependencies. The temporal convolutional modules are used to extract the dynamic change patterns of the multimodal time series data of each workstation node to capture temporal dependencies.

4. The production line monitoring method for a fan controller as described in claim 1, characterized in that, The health baseline is determined based on the statistical distribution of the production line health entropy calculated over multiple historical normal operating cycles. When the real-time calculated production line health entropy is higher than a certain percentage of the health baseline, it is determined that the production line is in an abnormal state.

5. The production line monitoring method for a fan controller as described in claim 1, characterized in that, Real-time assessment of whether there are abnormal conditions on the production line includes: During the normal operation cycle of the production line, a large number of health entropy values ​​are collected, and their mean and standard deviation are obtained by fitting a Gaussian distribution. The mean plus three times the standard deviation is set as a dynamic threshold for anomaly judgment. If the real-time calculated health entropy of the production line exceeds the dynamic threshold, it is judged as an anomaly.

6. The production line monitoring method for a fan controller as described in claim 1, characterized in that, The anomaly root cause workstations with the highest location contribution include: Using the production line health entropy as the objective function, backpropagation is performed along the spatiotemporal graph convolutional network to calculate the gradient of the health entropy with respect to the input features of each workstation within a preset time window before the anomaly occurs. The absolute values ​​of the gradients of all input features at each workstation are accumulated to obtain the total gradient contribution of that workstation. The total gradient contributions of all workstations are sorted, and the workstation with the highest contribution is identified as the root cause workstation of the anomaly.

7. The production line monitoring method for a fan controller as described in any one of claims 1 to 6, characterized in that, The evolution trajectory of the production line health entropy within a future preset time window includes: An attention mechanism is used to process the historical spatiotemporal correlation feature sequence to assign weights to the spatiotemporal correlation features at each time step; The attention-weighted feature sequence is input into the Kalman filter, and through iterative prediction and update steps, a sequence of predicted values ​​of the production line health entropy within a preset future time window is generated; and the confidence interval for each predicted value is calculated using the state covariance matrix of the Kalman filter.

8. A production line monitoring system for a fan controller, comprising a memory and a processor, wherein the memory stores computer program instructions, characterized in that, When the computer program instructions are executed by the processor, they implement the processing production line monitoring method for the fan controller as described in any one of claims 1 to 7.

Images