Feeder fault detection and processing method and system

Abstract

The invention relates to the technical field of fault detection, and provides a fault detection and processing method and system for a feeder, and the method comprises the steps: obtaining the multi-source data of the feeder in real time, and extracting the multi-source features, so as to sense the operation state of the feeder in real time; comparing the multi-source features with a determined fault threshold value to judge a fault type and a fault level, processing the multi-source features through a graph neural network to generate a fault diffusion probability of the feeder, and judging and outputting a fault conduction path and the fault diffusion probability according to the fault diffusion probability so as to accurately recognize the fault of the feeder; then, a multi-stage response mechanism is generated, and the execution effect of the feeder is monitored for closed-loop updating, so that the timeliness and accuracy of fault detection are remarkably improved, the non-planned shutdown risk of the feeder is reduced, the production loss and the maintenance cost caused by faults are reduced, high efficiency, stability and intelligent management of operation of the feeder are realized, and the operation efficiency of the feeder is improved. A powerful guarantee is provided for the continuity and the safety of industrial production.

Description

Technical Field

[0001] This application relates to the field of fault detection technology, and in particular to a fault detection and handling method and system for a feeder. Background Technology

[0002] In modern industrial production, feeders are key equipment in the material conveying and processing process. Their operational stability directly affects the continuity and production efficiency of the production line. With the development of industrial automation and intelligence, feeder fault detection technology is also constantly evolving. In the early days, manual inspection and simple instrument monitoring were the main fault detection methods, which relied on the experience and judgment of maintenance personnel. The detection efficiency was low and the accuracy was poor, making it difficult to detect early potential faults.

[0003] With the advancement of sensor and electronic information technologies, automated fault diagnosis technologies based on single-signal detection are becoming increasingly popular. For example, vibration sensors monitor the vibration frequency of equipment, or current sensors analyze changes in motor current to determine faults. While these technologies improve detection efficiency, they have limitations. Relying on a single signal, they cannot fully reflect the complex operating conditions of the feeder, and are prone to missed detections or misjudgments. Fixed fault judgment thresholds are difficult to adapt to different working conditions such as changes in material characteristics and fluctuations in equipment load, resulting in poor adaptability in practical applications.

[0004] In recent years, some studies have attempted to use multi-sensor data fusion technology for fault detection, aiming to improve diagnostic accuracy by integrating various signal data. However, existing fusion methods often employ simple data overlay or weighted averaging, failing to fully explore the deep correlations between multi-source data and making it difficult to effectively extract fault-sensitive features. Furthermore, in terms of fault handling, traditional systems often use preset fixed response strategies, lacking dynamic assessment of fault risks and cost-benefit analysis. This can easily lead to over-maintenance or delayed maintenance, increasing operating costs and compromising the reliable operation of the feeder. Most studies have not addressed how to achieve deep fusion and accurate feature extraction of multi-source data under complex operating conditions to construct an intelligent fault detection and handling system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a method and system for fault detection and handling of a feeder.

[0006] In a first aspect, this application provides a fault detection and handling system for a feeder, the system comprising: a sensor fusion module, a fault detection module, and an execution processing module;

[0007] The sensor fusion module is used to acquire multi-source data of the feeder in real time. The multi-source data includes vibration signals, acoustic signals and motor current. The multi-source data is processed by wavelet packet decomposition algorithm, and multi-source features are extracted and verified. The multi-source features include equipment surface defects, material collision acoustics and motor harmonic distortion.

[0008] The fault detection module is used to dynamically correct the fault threshold based on the density and humidity of the material. When the multi-source features are greater than or equal to the fault threshold, the fault type and fault level are output. When the multi-source features are less than the fault threshold, the multi-source features are processed by a graph neural network to generate the fault propagation probability of the feeder. The nodes of the graph neural network represent the physical components of the feeder, and the edge weights are generated by training from historical fault data. At the same time, a warning threshold is configured. When the fault propagation probability is greater than or equal to the warning threshold, the fault propagation path and fault propagation probability are output.

[0009] The execution processing module is used to generate a multi-level response mechanism based on the output fault type and fault level, as well as the fault propagation path and fault propagation probability. Based on the multi-level response mechanism, the feeder is controlled, and the execution effect of the feeder is monitored to update the fault threshold and the weights of the edges of the graph neural network.

[0010] As an optional implementation, the multi-source feature extraction logic includes:

[0011] The acquired vibration signal, acoustic signal, and motor current are normalized, and the number of decomposition layers is dynamically adjusted according to the information entropy change rate of each multi-source data using an adaptive wavelet packet decomposition algorithm to obtain the decomposed frequency band signal.

[0012] Calculate the energy entropy of each frequency band signal after decomposition to generate energy entropy matrices for vibration signal, acoustic signature signal and motor current respectively;

[0013] Weights are assigned to vibration signals, acoustic signals, and motor current using an attention mechanism. The energy entropy matrices of vibration signals, acoustic signals, and motor current are then fused using tensors to form a feature tensor. This feature tensor is then processed by a convolutional neural network to extract multi-source features.

[0014] As an optional implementation, the verification logic for the multi-source features includes:

[0015] The correlation between any two of the following is calculated: surface defects of equipment, sound patterns of material collisions, and harmonic distortion of motor. A correlation matrix is ​​generated, a correlation threshold is configured, and the correlation matrix is ​​compared with the correlation threshold to obtain the verification result of modal consistency.

[0016] Set a sliding time window to monitor in real time the difference between the mean of multi-source features in the current time window and the mean of multi-source features under historical normal conditions, and obtain time series verification results based on the difference.

[0017] The physical constraints are determined based on the mechanical principle of the feeder, and the monitored multi-source features are compared with the physical constraints to obtain the verification results of the physical constraints.

[0018] As an optional implementation, the output logic of the fault propagation path includes:

[0019] The fault propagation probability is transformed into the conditional probability of a Bayesian network, where the physical components of the feeder represent the nodes of the Bayesian network, and the fault propagation relationship represents the directed edges of the Bayesian network.

[0020] The reliability of each fault propagation path is evaluated based on conditional probability using a probabilistic reasoning algorithm.

[0021] The reliability score of each fault propagation path is calculated by combining the reliability of each fault propagation path. The fault propagation paths are then sorted in descending order according to their reliability scores to output the fault propagation paths.

[0022] As an optional implementation, the fault threshold correction sub-logic includes:

[0023] Extract the correlation patterns between material density and moisture content and feeder malfunctions from historical fault data, divide the material density and moisture content into multiple intervals, and form grid cells by intersection. Each grid cell corresponds to a combination of density and moisture content.

[0024] The multi-source characteristics of historical faults in each grid cell are statistically analyzed, and fault thresholds are configured based on the correlation patterns between material density and moisture content and feeder faults.

[0025] The system monitors the feeder's operating status in real time, adjusts the fault threshold based on the feeder's operating status, monitors the measurement error of multi-source data to determine the confidence interval for the fault threshold adjustment, and expands the fluctuation range of the fault threshold based on the confidence interval for the fault threshold adjustment in order to correct the fault threshold.

[0026] As an optional implementation, the fault propagation probability generation sub-logic includes:

[0027] The physical components of the feeder are mapped to nodes of a graph neural network. The multi-source features of the feeder characterize the node features. The probability of fault propagation between physical components is calculated based on historical fault data and used as the weight of the edges of the graph neural network.

[0028] The node features are propagated according to the edge weights through spatiotemporal convolution operations to obtain the fault propagation relationship;

[0029] In a graph neural network, fault injection of nodes is simulated, and the changes in node characteristics of other nodes are observed to quantify the causal effect of the fault. By comparing the changes in node characteristics in the normal state and the fault injection state, the fault propagation path can be identified.

[0030] By combining fault propagation relationships, fault causal effects, and fault propagation paths, the probability distribution of individual physical components is generated, and the probability distribution of the physical components of the feeder is integrated to generate the fault propagation probability.

[0031] As an optional implementation, the multi-level response mechanism includes:

[0032] Based on the output fault type and fault level, as well as the fault propagation path and fault propagation probability, the fault risk of the feeder is classified into risk levels, and a response strategy is matched according to the risk level.

[0033] Based on a comprehensive comparison of the cost-benefit of each response strategy, including downtime costs, repair costs, and risk costs, a response strategy is selected and synchronously transmitted to the feeder.

[0034] As an optional implementation, the fault threshold update logic includes:

[0035] The feeder is controlled based on a multi-level response mechanism, and the execution effect of the feeder is monitored. The fault threshold is adjusted according to whether the fault of the feeder is eliminated.

[0036] Real-time monitoring of the feeder's operating status, as well as the material's density and moisture content, allows for dynamic adjustment of fault thresholds.

[0037] After each adjustment of the fault threshold, a fault is simulated and injected to verify whether the adjusted fault threshold can trigger the multi-level response mechanism. If the verification fails, the fault threshold is rolled back and readjusted.

[0038] As an optional implementation, the logic for updating the edge weights includes:

[0039] The feeder is controlled based on a multi-level response mechanism. The execution effect of the feeder is monitored, and the actual fault propagation path is compared with the identified fault propagation path to obtain the path judgment result.

[0040] When the path determination result indicates that the paths are consistent, increase the weight of the edges of the fault propagation path;

[0041] When the path determination result indicates that the paths are inconsistent, the weights of the edges in the fault propagation path are reduced, and the weights of the edges in the graph neural network are updated periodically according to different time scales.

[0042] Secondly, this application provides a fault detection and handling method for a feeder. The method includes: acquiring multi-source data of the feeder in real time, including vibration signal, acoustic signal and motor current, processing the multi-source data through wavelet packet decomposition algorithm, and extracting and verifying multi-source features.

[0043] The fault threshold is dynamically corrected based on the density and humidity of the material. When the multi-source characteristics are greater than or equal to the fault threshold, the fault type and fault level are output.

[0044] When the multi-source features are less than the fault threshold, the multi-source features are processed by a graph neural network to generate the fault propagation probability of the feeder.

[0045] Configure an early warning threshold. When the probability of fault propagation is greater than or equal to the early warning threshold, output the fault propagation path and the probability of fault propagation.

[0046] A multi-level response mechanism is generated based on the output fault type, fault level, fault propagation path, and fault propagation probability, and the feeder is controlled based on the multi-level response mechanism.

[0047] Compared with existing technologies, the beneficial effects of this application are as follows: By acquiring multi-source data from the feeder in real time and extracting multi-source features, the operating status of the feeder can be perceived in real time; the multi-source features are compared with determined fault thresholds to determine the fault type and fault level, and the multi-source features are processed through a graph neural network to generate the fault propagation probability of the feeder, and the fault propagation path and fault propagation probability are output based on the fault propagation probability, thereby accurately identifying the faults of the feeder; then a multi-level response mechanism is generated, and the execution effect of the feeder is monitored for closed-loop updates, thereby significantly improving the timeliness and accuracy of fault detection, reducing the risk of unplanned downtime of the feeder, reducing production losses and maintenance costs caused by faults, and realizing the high efficiency, stability and intelligent management of the feeder operation, providing a strong guarantee for the continuity and safety of industrial production.

[0048] By acquiring multi-source data such as vibration signals, acoustic signals, and motor current, and processing the multi-source data through an adaptive wavelet packet decomposition algorithm, it is possible to comprehensively capture subtle changes in the feeder's operation. Whether it is surface defects of the equipment, abnormal material collisions, or fault features such as motor harmonic distortion, they can all be effectively extracted, avoiding the limitations of single signal detection and greatly improving the comprehensiveness and accuracy of fault feature identification.

[0049] By adjusting fault thresholds based on operating conditions such as material density and humidity, this approach overcomes the shortcomings of traditional fixed thresholds that cannot adapt to complex operating conditions. It divides the material into grid cells representing its characteristics, configures fault thresholds based on historical fault data, and dynamically optimizes them according to the feeder's real-time operating status and data errors. This ensures that the fault thresholds accurately match the feeder's actual operating status under different conditions, effectively reducing false alarms and missed alarms, and improving the accuracy of fault diagnosis. Furthermore, by processing multi-source features through graph neural networks, it deeply analyzes fault propagation patterns from a spatiotemporal perspective, accurately identifying potential fault transmission paths and predicting fault propagation trends in advance.

[0050] A multi-level response mechanism is generated based on fault type, fault level, fault propagation path, and fault propagation probability. By classifying risk levels and matching targeted strategies, it avoids resource waste caused by over-maintenance and prevents fault deterioration caused by insufficient response, effectively balancing production efficiency and equipment maintenance needs. By monitoring the execution effect of the feeder and updating the fault threshold and the edge weights of the graph neural network, the dynamic adjustment of the fault threshold ensures that the detection standard always fits the actual state of the feeder, while the updating of the graph neural network weights continuously optimizes the fault propagation model. This allows the detection system's ability to predict and handle faults to continuously improve with the accumulation of feeder operating data, ensuring the long-term reliability and effectiveness of the detection system. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0052] Figure 1 A flowchart illustrating a fault detection and handling system for a feeder provided in an embodiment of this application;

[0053] Figure 2 This is a sub-logic diagram of the fault threshold correction of a fault detection and processing system for a feeder provided in an embodiment of this application;

[0054] Figure 3 A sub-logic diagram for generating the fault propagation probability of a fault detection and handling system for a feeder provided in an embodiment of this application;

[0055] Figure 4 This is a flowchart illustrating a fault detection and handling method for a feeder provided in an embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0057] Example 1

[0058] like Figure 1 The diagram shown is a system flowchart of a fault detection and handling system for a feeder provided in this application embodiment. The system includes a sensor fusion module, a fault detection module, and an execution processing module.

[0059] The sensor fusion module is used to acquire multi-source data of the feeder in real time. The multi-source data includes vibration signals, acoustic signals and motor current. The multi-source data is processed by wavelet packet decomposition algorithm, and multi-source features are extracted and verified. The multi-source features include equipment surface defects, material collision acoustics and motor harmonic distortion.

[0060] Specifically, the extraction logic for multi-source features includes:

[0061] The acquired vibration signal, acoustic signal, and motor current are normalized, and the number of decomposition layers is dynamically adjusted according to the information entropy change rate of each multi-source data using an adaptive wavelet packet decomposition algorithm to obtain the decomposed frequency band signal.

[0062] Calculate the energy entropy of each frequency band signal after decomposition to generate energy entropy matrices for vibration signal, acoustic signature signal and motor current respectively;

[0063] Weights are assigned to vibration signals, acoustic signals, and motor current using an attention mechanism. The energy entropy matrices of vibration signals, acoustic signals, and motor current are then fused using tensors to form a feature tensor. This feature tensor is then processed by a convolutional neural network to extract multi-source features.

[0064] The raw data of vibration signals, acoustic signals, and motor current have different dimensions, and their amplitude fluctuations vary greatly under different operating conditions such as no-load and heavy-load. Furthermore, different faults exhibit different characteristics in the signal frequency band, making a fixed data decomposition method unsuitable for all situations. This method acquires multi-source data from the feeder in real time using sensors, including vibration signals, acoustic signals, and motor current. First, the acquired three types of signals are normalized to eliminate interference from dimensional differences. Then, an adaptive wavelet packet decomposition algorithm is used to continuously analyze the information entropy changes of the three types of signals after each level of decomposition. When repeated analyses show that the information entropy changes tend to stabilize, it indicates that the three types of signals have reached a stable state. The signal features have been sufficiently decomposed. At this point, the decomposition is stopped and the optimal number of decomposition layers is determined. In specific applications, when a feeder is conveying high-density materials, the increased motor load makes the current motor more complex. The adaptive wavelet packet decomposition algorithm automatically increases the number of decomposition layers to meticulously capture abnormal features in the low-frequency band. When running under no-load conditions, the vibration signal is relatively simple, and the number of decomposition layers is automatically reduced to improve data processing efficiency. Compared with a fixed decomposition method, the adaptive wavelet packet decomposition algorithm can dynamically adjust according to the actual situation of the three types of signals. It will not miss key frequency band signals, nor will it generate redundant data, ensuring that fault-sensitive features are completely preserved under different operating conditions.

[0065] Traditional signal analysis focuses only on a single dimension, amplitude or frequency, making it difficult to comprehensively reflect fault characteristics. Energy entropy can quantify the stability of energy distribution in different frequency bands. For early, minor faults, subtle changes in energy entropy are more sensitive than amplitude fluctuations, and can uncover hidden fault information. The energy entropy of each frequency band signal after decomposition is calculated and arranged according to frequency band order and signal type, generating energy entropy matrices for vibration signals, acoustic signatures, and motor currents. Taking vibration signals as an example, each element in the energy entropy matrix represents the energy entropy of a specific frequency band. When early wear occurs in equipment bearings, the energy entropy in the high-frequency band of the vibration signal will show abnormal fluctuations, and this trend can be intuitively observed through the energy entropy matrix. This transforms the original signal into a more fault-sensitive feature form, enhancing the fault characterization ability of the three types of signals and enabling the detection of early fault signs that are easily overlooked by traditional methods. The structured energy entropy matrix provides a standardized data format for tensor fusion, facilitating the subsequent extraction of effective multi-source features.

[0066] Vibration signals, acoustic signatures, and motor currents exhibit varying sensitivities to different types of faults. Vibration signals are sensitive to mechanical structural faults, while motor currents are sensitive to electrical faults. Weighting is crucial to highlight key information, while tensor fusion integrates the spatial relationships of multi-source data, and convolutional neural networks further extract deep abstract features. Through an attention mechanism, dynamic weights are assigned to vibration signals, acoustic signatures, and motor currents based on the contribution of various signal features in historical fault data to fault diagnosis. Preferably, vibration signals are given higher weight when diagnosing surface defects, while current signals are weighted more heavily when diagnosing motor faults. The three energy entropy matrices are then fused using tensors to form a feature tensor. This feature tensor is then subjected to multi-layer convolution and pooling operations by a convolutional neural network to gradually extract multi-source features, including information on surface defects, material collision acoustic signatures, and motor harmonic distortion. This effectively integrates the advantages of each multi-source signal, avoiding the limitations of single-signal diagnosis. Dynamic weighting and deep feature extraction comprehensively enhance fault feature recognition capabilities. High-quality multi-source feature vectors provide accurate input data for the fault detection module, supporting subsequent processing.

[0067] Specifically, the verification logic for multi-source features includes:

[0068] The correlation between any two of the following is calculated: surface defects of equipment, sound patterns of material collisions, and harmonic distortion of motor. A correlation matrix is ​​generated, a correlation threshold is configured, and the correlation matrix is ​​compared with the correlation threshold to obtain the verification result of modal consistency.

[0069] Set a sliding time window to monitor in real time the difference between the mean of multi-source features in the current time window and the mean of multi-source features under historical normal conditions, and obtain time series verification results based on the difference.

[0070] The physical constraints are determined based on the mechanical principle of the feeder, and the monitored multi-source features are compared with the physical constraints to obtain the verification results of the physical constraints.

[0071] There is an inherent physical correlation between multi-source signals. When material collisions generate acoustic signals, they often cause changes in vibration signals. If the correlation between multi-source features is abnormal, it indicates sensor failure, signal transmission interference, or feature extraction errors. Therefore, consistency verification is required. The correlation coefficient is calculated for any two of the multi-source features, such as equipment surface defects, material collision acoustic signals, and motor harmonic distortion. A correlation matrix is ​​generated. Through statistical analysis of a large amount of normal operating data, a correlation threshold is preset. When the correlation coefficient in the correlation matrix is ​​less than the correlation threshold, the correlation coefficient between the vibration signal and the acoustic signal decreases significantly in specific applications. This indicates a problem with the feature combination and triggers a sensor status check process, including checking whether the sensor power supply and communication lines are normal. At the same time, the signal re-acquisition and feature extraction programs are started to prevent erroneous data from entering the subsequent processing flow. This enables timely detection of sensor failures or abnormal feature extraction, reduces the risk of misjudgment due to data errors, and ensures that the feature data input to the fault detection module is true and reliable, providing a solid foundation for accurately judging the fault type and fault level.

[0072] Equipment failures develop continuously over time. Early failures often manifest only as subtle changes in characteristics. By comparing current multi-source characteristics with those under historical normal conditions, these abnormal trends can be identified, enabling early warning of failures. A sliding time window is set to continuously acquire multi-source characteristics within the current time window. By analyzing the differences between the current multi-source characteristics and the average values ​​of multi-source characteristics under historical normal conditions, it is determined whether there are timing anomalies. Simultaneously, the judgment criteria are dynamically adjusted in conjunction with the feeder's start-up and shutdown patterns. In specific applications, a larger fluctuation range of current harmonic distortion is allowed during the start-up phase. Once the feeder starts, if the motor current remains higher than the fluctuation range of the normal start-up phase, an anomaly is determined, and an early overload warning is issued. This allows for the capture of subtle changes in the early stages of failure, early detection of potential failure risks, and provision of time-dimensional verification information for the fault detection module. This helps to more accurately determine the stage of failure development and buys time for developing a reasonable response strategy.

[0073] The operation of each physical component of the feeder follows physical laws. A quantitative relationship exists between the conveyor belt speed, motor current, and material conveying capacity. Physical constraints can verify the rationality of multi-source characteristics and eliminate false multi-source characteristics that do not conform to actual conditions. Based on the mechanical principles of the feeder, the physical constraints that fault characteristics should meet are determined. Specifically, the acoustic frequency of material collision has a specific transmission ratio relationship with the dominant frequency of the vibration signal, and the motor current is positively correlated with the load weight. The monitored multi-source characteristics are compared with the physical constraints. If the actual monitored multi-source characteristics do not conform to the physical constraints, or if the ratio of acoustic frequency to dominant vibration frequency is abnormal in specific applications, it is necessary to check whether the sensor installation position and parameter settings are correct. If hardware problems are ruled out, the multi-source characteristics are filtered and corrected or re-extracted. This ensures that the multi-source characteristics conform to the physical laws of the feeder's actual operation, improves data reliability, reduces erroneous diagnosis caused by abnormal data, provides effective characteristics that conform to actual working conditions for the fault detection module, enhances the accuracy of fault diagnosis, and avoids misjudgments and omissions.

[0074] The fault detection module dynamically adjusts the fault threshold based on the density and moisture content of the material. When the multi-source features are greater than or equal to the fault threshold, it outputs the fault type and fault level. When the multi-source features are less than the fault threshold, it processes the multi-source features through a graph neural network to generate the fault propagation probability of the feeder. The nodes of the graph neural network represent the physical components of the feeder, and the edge weights are generated by training with historical fault data. At the same time, a warning threshold is configured. When the fault propagation probability is greater than or equal to the warning threshold, it outputs the fault propagation path and the fault propagation probability.

[0075] Furthermore, such as Figure 2 As shown, the fault threshold correction sub-logic includes:

[0076] Extract the correlation patterns between material density and moisture content and feeder malfunctions from historical fault data, divide the material density and moisture content into several intervals, and form grid cells by intersection. Each grid cell corresponds to a combination of density and moisture content.

[0077] The multi-source characteristics of historical faults in each grid cell are statistically analyzed, and fault thresholds are configured based on the correlation patterns between material density and moisture content and feeder faults.

[0078] The system monitors the feeder's operating status in real time, adjusts the fault threshold based on the feeder's operating status, monitors the measurement error of multi-source data to determine the confidence interval for the fault threshold adjustment, and expands the fluctuation range of the fault threshold based on the confidence interval for the fault threshold adjustment in order to correct the fault threshold.

[0079] The density and moisture content of materials significantly affect the operating status of the feeder. Materials with different densities and moisture contents will subject the feeder to different loads and wear levels, resulting in differences in the multi-source characteristics of faults. A uniform fault threshold cannot accurately determine the operating status of the feeder under different material conditions. Therefore, it is necessary to classify material characteristics to provide a basis for setting differentiated fault thresholds. First, historical fault data is acquired, and the correlation patterns between material density and moisture content and feeder faults are analyzed. Based on the analysis results, the material density is divided into multiple intervals, including low density, medium density, and high density, and the material moisture content is also divided into several intervals, including low moisture content, medium moisture content, and high moisture content. Then, the density intervals and moisture content intervals are cross-combined to form grid cells. Each grid cell corresponds to a specific combination of material density and moisture content. In specific applications, one grid cell corresponds to the material condition of high density and high moisture content. This allows for a structured classification of complex and variable material conditions, enabling the setting of fault thresholds to better reflect the actual operating status of the feeder under different conditions and improving the targeting of fault diagnosis.

[0080] Under different combinations of material characteristics, the multi-source characteristics of feeder failures vary. Only by understanding the actual distribution of fault characteristics under each condition can fault thresholds be set appropriately, avoiding misjudgments or missed detections due to improper threshold settings. This involves statistically analyzing the multi-source characteristics of historical faults within each grid cell. In high-density and high-humidity grid cells, the intensity range of vibration signals, frequency characteristics of acoustic signatures, and fluctuations in motor current are analyzed from historical fault data. Combined with the previously analyzed correlation patterns between material density and humidity and feeder faults, and considering the feeder's design parameters and operational experience, fault thresholds are configured for the multi-source characteristics of each grid cell. Preferably, under high-density and high-humidity combinations, the normal fluctuation range of motor current increases due to the larger feeder load, thus correspondingly increasing the fault threshold for motor current. This allows the fault thresholds to accurately reflect the boundary conditions of feeder failures under different material conditions, improving fault detection accuracy and reducing false alarms and missed alarms.

[0081] The feeder's operating status is dynamic, including load variations and runtime. Furthermore, measurements from multiple data sources can introduce errors, and fixed fault thresholds cannot adapt to these changes, easily leading to inaccurate fault diagnosis. Therefore, it is necessary to monitor the feeder's operating status and measurement errors in real time, dynamically adjusting the fault thresholds to ensure reliable fault detection. Real-time monitoring of the feeder's operating parameters, including motor speed, conveyor belt speed, and equipment vibration amplitude, is crucial for determining the equipment's operating status. When increased load or excessively long operating times are detected, the fault threshold is automatically adjusted. If the feeder runs continuously for several hours, considering component fatigue and wear, the fault threshold for surface defects is appropriately lowered to improve the sensitivity of potential fault detection. Simultaneously, continuous monitoring of measurement errors from multiple data sources is essential. In cases of poor performance, when a large measurement error is detected, a confidence interval for adjusting the fault threshold is determined based on the degree of error. The fluctuation range of the fault threshold is then expanded based on this confidence interval. If the measurement error of a certain sensor exceeds a certain proportion, the fault threshold range corresponding to the multi-source feature will be fluctuated up or down by a certain margin to avoid false alarms due to inaccurate data. This allows the fault threshold to dynamically change with the actual operating status of the feeder and the data measurement situation, effectively reducing the risk of misjudgment and missed judgment caused by operating condition fluctuations or data errors, and improving the adaptability and stability of the fault detection system. The corrected fault threshold is directly applied to the fault detection module to determine whether the multi-source feature is greater than or equal to the fault threshold, thereby deciding whether to output the fault type and fault level, and whether further calculation of the fault propagation probability and other subsequent operations are needed.

[0082] When a multi-source feature is greater than or equal to a fault threshold, the fault type and fault level are output. If any multi-source feature is greater than or equal to the fault threshold, it indicates that a component of the feeder has a problem. That is, if the equipment surface defect is greater than or equal to the corresponding fault threshold, it indicates that the equipment surface is damaged. If the material collision sound is greater than or equal to the corresponding fault threshold, it indicates that the material conveying is abnormal. If the motor harmonic distortion is greater than the corresponding fault threshold, it indicates that the motor is faulty. When any multi-source feature is less than the fault threshold, the corresponding multi-source feature is processed by a graph neural network to generate the fault propagation probability of the feeder. That is, if only the motor harmonic distortion is less than the corresponding fault threshold, while the equipment surface defect and the material collision sound are greater than or equal to the corresponding fault threshold, it indicates that the equipment surface is damaged and the material conveying is abnormal. However, it is necessary to further determine the fault propagation probability at the feeder motor. The comparison between the multi-source features and the fault threshold requires comparing the equipment surface defect, the material collision sound, and the motor harmonic distortion with the fault threshold respectively.

[0083] For equipment surface defects, image recognition technology is used to monitor the equipment surface. The acquired equipment surface images are preprocessed, including noise reduction and contrast enhancement, to extract key features such as edges, textures, and colors. Target detection algorithms are then used to identify defect features including cracks, wear, and deformation. The extracted defect features are compared with pre-set fault thresholds for equipment surface defects. The setting of fault thresholds needs to take into account factors such as equipment type, service life, and operating environment. For example, for feeders with a long service life, the allowable range for surface defects is relatively lenient, but it still needs to be dynamically adjusted based on actual operating conditions. If the area ratio, length, and depth of the identified defects are greater than or equal to the corresponding fault threshold, it indicates that there is damage to the equipment surface.

[0084] Material collision acoustic signatures require time-domain and frequency-domain analysis of the acquired acoustic signature signals. Time-domain analysis can obtain features such as the mean, variance, and peak value of the signal, while frequency-domain analysis uses methods such as Fourier transform to obtain the spectral characteristics of the signal, including the dominant frequency and harmonic distribution. These analyzed acoustic signature features are compared with a fault threshold set for material collision acoustic signatures. The determination of the fault threshold requires a large amount of acoustic signature data under normal operation and fault conditions and the use of statistical analysis methods. For example, the mean and standard deviation of the acoustic signature features under normal conditions are calculated, and the mean plus a certain multiple of the standard deviation is used as the fault threshold. If the actual acoustic signature features obtained from the analysis are greater than or equal to the corresponding fault threshold, it means that there is an abnormality in material conveying.

[0085] Motor harmonic distortion is determined by acquiring motor current through power monitoring equipment and then calculating the motor's harmonic distortion rate using a Fast Fourier Transform algorithm. This involves converting the motor current from the time domain to the frequency domain, analyzing the content of each harmonic, and thus obtaining the harmonic distortion rate. The calculated motor harmonic distortion rate is then compared with a fault threshold set based on the motor's normal operating characteristics and industry standards. The motor fault threshold varies depending on factors such as motor type, power, and rated operating parameters, and needs to be dynamically adjusted according to operating conditions such as material density and humidity. For example, when the material density is high, the increased motor load leads to an increase in the harmonic distortion rate, and the corresponding fault threshold should be adjusted accordingly. If the motor harmonic distortion rate is greater than or equal to the adjusted fault threshold, a motor fault is determined to exist.

[0086] Furthermore, such as Figure 3 As shown, the sub-logic for generating the fault propagation probability includes:

[0087] The physical components of the feeder are mapped to nodes of a graph neural network. The multi-source features of the feeder characterize the node features. The probability of fault propagation between physical components is calculated based on historical fault data and used as the weight of the edges of the graph neural network.

[0088] The node features are propagated according to the edge weights through spatiotemporal convolution operations to obtain the fault propagation relationship;

[0089] In a graph neural network, fault injection of nodes is simulated, and the changes in node characteristics of other nodes are observed to quantify the causal effect of the fault. By comparing the changes in node characteristics in the normal state and the fault injection state, the fault propagation path can be identified.

[0090] By combining fault propagation relationships, fault causal effects, and fault propagation paths, the probability distribution of individual physical components is generated, and the probability distribution of the physical components of the feeder is integrated to generate the fault propagation probability.

[0091] The physical components of a feeder do not operate independently but are interconnected. A failure in one component can trigger a chain reaction in others. Traditional analysis methods struggle to fully describe this complex relationship. Graph neural networks, however, can intuitively observe the relationships between physical components and the fault propagation path through nodes and edges. Therefore, it is necessary to map the physical components of the feeder into a graph neural network. First, the physical components of the feeder, including the motor, conveyor belt, bearings, and reducer, are mapped one-to-one to nodes in the graph neural network. Then, multi-source features extracted by the sensor fusion module, including information such as equipment surface defects, material collision sound patterns, and motor harmonic distortion, are assigned to each node as node features, enabling the nodes to characterize the corresponding material. The system first analyzes the operating status of the components, then, based on historical fault data, analyzes the sequence of faults and the degree of correlation between the physical components. In specific applications, it reviews historical fault data and counts the number of times other physical components such as conveyor belts and motors fail within a certain period of time after a bearing failure. Based on these statistical results, it calculates the probability of fault propagation between physical components, thereby determining the weight of edges in the graph neural network. The higher the probability of fault propagation between two physical components, the greater the weight of the edge between their corresponding nodes. This transforms the physical components of the feeder and the fault correlation into a computer-processable graph neural network, providing a structured data foundation and analytical framework for further in-depth analysis of fault propagation patterns.

[0092] The propagation of faults in a feeder involves not only spatial diffusion from one physical component to another but also temporal characteristics. That is, the fault gradually develops and spreads over time. Simple spatial relationship analysis cannot accurately capture the dynamic evolution of the fault; therefore, a comprehensive analysis combining spatiotemporal dimensions is necessary to obtain the true fault propagation relationship. Through spatiotemporal convolution operations, node features are propagated in a graph neural network according to edge weights. In the temporal dimension, the changes in node features of each physical component over time are continuously observed to analyze how fault features affect other physical components at different times. This includes focusing on the changes in abnormal motor current over a period of time, and... This study investigates how various changes gradually affect the vibration characteristics and operating status of the conveyor belt. In the spatial dimension, the direction and extent of fault propagation between different physical components are determined based on the weight of the edges. Edges with higher weights indicate that faults are more likely to propagate along this path and have a greater impact on the target component. Through multiple spatiotemporal convolution operations, the study continuously explores and refines the fault propagation relationships between physical components, gradually forming a clear understanding of the path and patterns of fault propagation in the feeder. This approach comprehensively considers the spatiotemporal characteristics of fault propagation and, compared to single-dimensional analysis methods, can more accurately and meticulously describe the diffusion process of faults in the feeder, discover more potential fault propagation paths, and improve the ability to grasp the patterns of fault propagation.

[0093] Understanding fault propagation relationships alone is insufficient to assess the specific impact of a fault on each component or accurately identify the fault propagation path. By simulating fault injection and observing changes in other physical components, the causal effects of the fault among physical components can be quantified, thereby more accurately determining the fault propagation path and providing more detailed information for fault risk assessment. In the constructed graph neural network, a specific node is selected to simulate fault injection, i.e., the component represented by that node is artificially set to fail. Then, the changes in the node characteristics of other nodes are observed in real time. Preferably, a fault is simulated at a motor node, and vibration signals and acoustic signatures from nodes such as conveyor belts and bearings are monitored. How the characteristics change is determined by comparing the changes in multi-source characteristics of each node under normal conditions with those after fault injection. This quantifies the causal effect of the fault on other physical components, assessing the impact of the fault on the operating status of other physical components. Based on the order and extent of changes in multi-source characteristics, the actual fault propagation path is identified, clarifying how the fault spreads from the injection node to other physical components. By simulating real fault scenarios, the causal effect of the fault is quantified, making the analysis of fault propagation closer to reality and enabling more accurate identification of fault propagation paths. This provides a strong basis for developing targeted fault prevention and handling strategies.

[0094] Failure propagation in a feeder is a holistic phenomenon. A comprehensive assessment of the overall feeder failure risk requires considering the failure probability of each physical component and their interrelationships. Therefore, it's necessary to integrate the probability distributions of individual physical components to obtain the overall failure propagation probability of the feeder. Combining the failure propagation relationships, causal effects, and propagation paths obtained in previous steps, the probability of each physical component failing under different failure scenarios is analyzed, generating the probability distribution of individual physical components. For the conveyor belt component, the impact of different sources of failure, such as motor and bearing failures, is considered, calculating the probability of conveyor belt failure under various conditions. Then, the probability distributions of all physical components of the feeder are integrated. The probability distribution is integrated, and during the integration process, the fault propagation relationship and fault causal effect between physical components are fully considered. The probabilities of interrelated physical components are calculated and adjusted collaboratively to obtain a fault propagation probability that reflects the fault propagation trend and risk level of the entire feeder. The fault risk of the feeder is assessed from an overall perspective, providing maintenance personnel with comprehensive and accurate fault warning information, enabling them to take measures in advance to prevent fault propagation, reduce equipment downtime and maintenance costs. The generated fault propagation probability will serve as an important basis. When the fault propagation probability is greater than or equal to the warning threshold, the fault propagation path and fault propagation probability are output, triggering the execution processing module to take corresponding response measures.

[0095] Specifically, the output logic of the fault propagation path includes:

[0096] The fault propagation probability is transformed into the conditional probability of a Bayesian network, where the physical components of the feeder represent the nodes of the Bayesian network, and the fault propagation relationship represents the directed edges of the Bayesian network.

[0097] The reliability of each fault propagation path is evaluated based on conditional probability using a probabilistic reasoning algorithm.

[0098] The reliability score of each fault propagation path is calculated by combining the reliability of each fault propagation path. The fault propagation paths are then sorted in descending order according to their reliability scores to output the fault propagation paths.

[0099] Fault propagation probability reflects the likelihood of a fault spreading among the components of the feeder. However, this probabilistic information is rather abstract. Bayesian networks, in the form of nodes and directed edges, can intuitively and structurally express the probabilistic dependencies between physical components, transforming the fault propagation probability into the conditional probability of the Bayesian network. This facilitates subsequent systematic analysis and reasoning of the fault propagation path. The physical components of the feeder are mapped to nodes in the Bayesian network, with each node representing the operating state of a physical component, including normal and fault states. The fault propagation relationship obtained from graph neural network analysis is transformed into directed edges in the Bayesian network, where the direction of the directed edge indicates the direction of fault propagation. If the graph neural network shows a bearing fault... If a fault can cause conveyor belt malfunctions, a directed edge is constructed in the Bayesian network from the bearing node to the conveyor belt node. Then, based on the fault propagation probability between physical components calculated by the graph neural network, the conditional probability table of each node in the Bayesian network is filled. The conditional probability table records the probability that the child node is in a fault state when the parent node is in different states. In the conditional probability table of the bearing node, the probability of the conveyor belt malfunctioning when the bearing is in different fault states is recorded. This transforms the complex fault propagation probability information into a structured and visualized Bayesian network, making the fault propagation relationship clearer and providing an intuitive and easy-to-analyze logical framework for subsequent evaluation of the reliability of the fault propagation path.

[0100] The feeder has multiple potential fault propagation paths, each with varying probabilities and impacts. To prioritize these paths and avoid indiscriminate troubleshooting, the reliability of each path needs to be assessed to identify the most likely and damaging faults. A probabilistic reasoning algorithm, based on the conditional probability of a Bayesian network, is used to analyze each path. Starting from the fault source node, the joint probability of each node along the directed edge is calculated under given conditions. This includes analyzing faults such as bearing failure leading to conveyor belt malfunction and motor overload. On the transmission path, based on the conditional probability of the bearing node, the conditional probability of the bearing failure causing the conveyor belt abnormality, and the conditional probability of the conveyor belt abnormality causing the motor overload, the joint probability of the overall failure of the fault transmission path is calculated. In this way, all possible fault transmission paths are calculated to obtain the probability of failure of each fault transmission path, thereby assessing the reliability of the path. Through quantitative analysis, the reliability of each fault transmission path can be objectively and accurately assessed, the risk level of different fault transmission paths can be distinguished, and scientific decision-making basis can be provided for operation and maintenance personnel to avoid misjudgment or omission due to experience judgment.

[0101] After assessing the reliability of each fault propagation path, to more intuitively compare the merits of different paths and clarify the priority of fault handling, the reliability needs to be converted into a comprehensive reliability score. Paths are then ranked according to their scores, with the most critical paths prioritized to improve fault handling efficiency. A reliability score is calculated for each fault propagation path, taking into account its reliability, the importance of the components involved, and the difficulty of repair. Paths involving critical production components and with high reliability receive higher reliability scores, while those with high reliability but involving components with low repair difficulty and minimal impact on production receive lower scores. After calculation, all fault propagation paths are then adjusted based on their reliability scores. The fault propagation path is ranked and the highest-scoring fault propagation path is output as the primary fault propagation path. The output includes information such as the names of the physical components involved in the fault propagation path and the propagation relationship between the physical components. In a specific application, the output fault propagation path is the bearing pointing to the conveyor belt and then to the motor, suggesting that the bearing assembly should be checked first. Through reliability scoring and ranking, a clear and specific fault handling priority is provided to maintenance personnel, enabling them to quickly locate the source of the fault and the affected components, reduce fault investigation time, improve fault handling efficiency, and reduce losses caused by equipment downtime. The output fault propagation path will directly guide the execution processing module to formulate corresponding response strategies, including determining the maintenance sequence and preparing spare parts, so as to achieve rapid fault handling and timely equipment recovery.

[0102] The execution processing module generates a multi-level response mechanism based on the output fault type, fault level, fault propagation path, and fault propagation probability. It controls the feeder based on the multi-level response mechanism and monitors the feeder's execution effect to update the fault threshold and the weights of the edges in the graph neural network.

[0103] Specifically, the multi-level response mechanism includes:

[0104] Based on the output fault type and fault level, as well as the fault propagation path and fault propagation probability, the fault risk of the feeder is classified into risk levels, and a response strategy is matched according to the risk level.

[0105] Based on a comprehensive comparison of the cost-benefit of each response strategy, including downtime costs, repair costs, and risk costs, a response strategy is selected and synchronously transmitted to the feeder.

[0106] The impact of different fault types and levels on the operation of the feeder varies significantly. A single factor cannot comprehensively assess the severity of a fault. By combining fault type, fault level, fault propagation path, and fault spread probability to classify risk levels, the potential scope and severity of the fault's impact can be more accurately identified, providing a foundation for developing differentiated response strategies. First, fault types are classified, including mechanical, electrical, and control faults. Then, the initial severity of the fault is assessed based on its fault level. Simultaneously, the complexity of the fault propagation path and the spread probability are analyzed. If a fault rapidly spreads to critical components through multiple propagation paths, it is classified as high-risk; otherwise, it is considered low-risk. When a bearing fault is detected with a high spread probability, affecting the conveyor belt and motor, the detection system classifies it as high-risk. Through multi-dimensional comprehensive assessment, a refined classification of fault risk is achieved, avoiding the limitations of single-factor assessments and allowing the risk level to more accurately reflect the actual severity of the fault.

[0107] Different risk levels require different response strategies, but the implementation of these strategies incurs downtime costs, repair costs, and risk costs. To minimize costs and maximize benefits while ensuring safe equipment operation, a cost-benefit analysis of each response strategy is necessary. Multiple response strategies are preset for each risk level, including emergency shutdown, degraded operation, and planned maintenance. For each strategy, the downtime costs, repair costs, and risk costs are comprehensively evaluated. Downtime costs include production interruption losses; repair costs include spare parts replacement and labor costs; and risk costs include further damage caused by untimely fault handling. For high-risk faults, while the emergency shutdown strategy incurs higher downtime costs, it effectively reduces risk costs. The degraded operation strategy reduces downtime costs but increases the risk of further damage to the feeder. The detection system compares the cost-benefit ratios of different response strategies and selects the optimal strategy. Through cost-benefit analysis, the selected response strategy ensures that the feeder's safety is guaranteed while maximizing economic benefits, avoiding over-maintenance or under-maintenance.

[0108] Effective execution of the response strategy is crucial for mitigating the impact of failures. Since feeders typically consist of multiple subsystems, the response strategy requires coordinated operation among these subsystems. Therefore, the response strategy needs to be synchronously transmitted to all relevant physical components of the feeder to ensure consistency and coordination. The selected response strategy is converted into executable control commands and synchronously transmitted to the feeder's control system and relevant physical components via a communication interface. If an emergency stop strategy is selected, the detection system sends a stop command to the motor controller and simultaneously sends status update information to the monitoring system. To ensure reliable transmission of control commands, the system employs redundant communication links and a command confirmation mechanism to ensure that each physical component correctly receives and executes the commands. By synchronously transmitting and executing the response strategy, coordinated action among the feeder's components is achieved, ensuring efficient and reliable fault handling and reducing processing delays or errors caused by information asynchrony. The effective execution of the response strategy directly impacts the fault handling effect and equipment recovery time, laying the foundation for the feeder's rapid resumption of operation.

[0109] Specifically, the logic for updating the fault threshold includes:

[0110] The feeder is controlled based on a multi-level response mechanism, and the execution effect of the feeder is monitored. The fault threshold is adjusted according to whether the fault of the feeder is eliminated.

[0111] Real-time monitoring of the feeder's operating status, as well as the material's density and moisture content, allows for dynamic adjustment of fault thresholds.

[0112] After each adjustment of the fault threshold, a fault is simulated and injected to verify whether the adjusted fault threshold can trigger the multi-level response mechanism. If the verification fails, the fault threshold is rolled back and readjusted.

[0113] If the fault persists after the response strategy is executed, it indicates that the current fault threshold is set too high, resulting in the fault characteristics not being effectively identified. If a false alarm occurs, it indicates that the fault threshold is too low. Therefore, the fault threshold needs to be dynamically adjusted based on the execution results to ensure the accuracy of fault judgment. After the response strategy is executed, the multi-source characteristics of the feeder are continuously monitored. If the multi-source characteristics still exceed the original fault threshold, it indicates that the fault threshold is too high. In this case, the fault threshold for the corresponding multi-source characteristics is reduced, including lowering the fault threshold for current harmonic distortion. If a response is triggered without detecting an actual fault, the fault threshold is increased, including widening the fluctuation range of the vibration signal, and the adjustment log is recorded. In one instance, the system was stopped due to a vibration signal but no fault was detected. The detection system then raised the vibration threshold and recorded the adjustment log. This allows the fault threshold to dynamically match the actual operating state of the feeder, reducing the false positive and false negative rates and improving the accuracy of fault detection. The adjusted fault threshold directly participates in the judgment of subsequent fault detection modules, affecting the fault level output and the generation of response strategies.

[0114] The normal parameter range of the feeder varies significantly under different operating conditions or material density and humidity. Fixed thresholds are prone to misjudgment, so it is necessary to dynamically adjust the fault threshold based on real-time operating conditions. The detection system acquires the feeder's operating parameters and the material's density and humidity in real time through sensors. When an increase in material density is detected, the upper limit of the current threshold is automatically increased considering the increased motor load. After the feeder has been running continuously for more than a few hours, the fault threshold for surface defects is reduced based on component fatigue laws to provide early warning. When the material humidity changes, the fault threshold range for material collision sound patterns is adjusted synchronously to avoid false alarms caused by changes in material viscosity. The fault threshold is dynamically adapted to the operating conditions, enhancing the detection system's adaptability to complex scenarios and improving the fault detection accuracy under varying operating conditions. The dynamically adjusted fault threshold provides a more accurate benchmark for calculating the fault propagation probability, affecting the graph neural network's prediction of the fault propagation path.

[0115] Arbitrarily adjusting the threshold can introduce new judgment biases, requiring verification to ensure the rationality of the adjustment and avoid system malfunctions. After each threshold adjustment, the detection system simulates and injects a standard fault signal in a virtual environment, i.e., the preset bearing wear vibration characteristics, and observes whether the adjusted fault threshold can trigger the corresponding response strategy. If the simulated fault does not trigger a response or is falsely triggered under normal operating conditions, the verification is deemed a failure, and the fault threshold is automatically rolled back to the pre-adjustment state, and a second adjustment process is initiated. Preferably, after adjusting the vibration threshold, if simulating slight wear does not trigger a warning, the fault threshold is rolled back and a small-scale readjustment is attempted. This ensures the reliability of the fault threshold adjustment, prevents instability of the detection system due to parameter drift, maintains the consistency of the fault detection logic, and officially puts the verified fault threshold into use to support the entire process of subsequent fault detection and response. The rollback operation prompts the detection system to re-optimize the adjustment strategy, forming a closed-loop improvement mechanism.

[0116] Specifically, the logic for updating the edge weights includes:

[0117] The feeder is controlled based on a multi-level response mechanism. The execution effect of the feeder is monitored, and the actual fault propagation path is compared with the identified fault propagation path to obtain the path judgment result.

[0118] When the path determination result indicates that the paths are consistent, increase the weight of the edges of the fault propagation path;

[0119] When the path determination result indicates that the paths are inconsistent, the weights of the edges in the fault propagation path are reduced, and the weights of the edges in the graph neural network are updated periodically according to different time scales.

[0120] The fault propagation path identified by the graph neural network is based on historical fault data and model predictions, which may deviate from the actual situation. By comparing the actual fault propagation path with the fault propagation path identified by the graph neural network, the shortcomings of the graph neural network can be identified, and the edge weights can be updated accordingly to make the graph neural network more closely reflect the actual operating state of the feeder. After the multi-level response mechanism controls the feeder to execute the response strategy, the detection system continuously monitors the operating status of the feeder, and acquires multi-source data such as vibration signals, acoustic signals, and motor current in real time through sensors. Combined with equipment maintenance records and information such as the sequence of fault occurrence, the actual fault propagation path is identified. After a machine malfunctions, analysis reveals that the fault first propagates to the conveyor belt, causing it to malfunction and subsequently affecting the speed reducer. This establishes the fault propagation path: the motor affects the conveyor belt, which in turn affects the speed reducer. This path is then compared with the fault propagation path previously identified by the graph neural network to determine if they match. This allows for timely detection of discrepancies between the graph neural network's identification and the actual situation, providing an accurate basis for adjusting edge weights and preventing subsequent fault prediction errors due to graph neural network biases. The judgment result obtained from the fault propagation path comparison directly determines the direction of subsequent edge weight adjustments—whether to increase or decrease weights or perform other optimization operations.

[0121] If the actual fault propagation path matches the identified fault propagation path, it indicates that the graph neural network's prediction of this propagation relationship is accurate. Increasing the weight of the corresponding edge can strengthen the graph neural network's understanding of this propagation relationship. If they do not match, it indicates that the graph neural network has made an incorrect prediction. Reducing the weight of the edge on the incorrect path and correcting the graph neural network helps improve its accuracy. When the path judgment result is that the paths match, that is, both the actual and identified fault propagation paths involve the bearing affecting the conveyor belt, the detection system automatically increases the weight of the edge between the bearing node and the conveyor belt node, making the graph neural network more inclined to believe that there is a fault propagation relationship between these two physical components in subsequent predictions. When the paths do not match, that is, the actual fault... The fault propagation path is that the motor affects the reducer, while the identified fault propagation path is that the motor affects the conveyor belt. The detection system reduces the weight of the edge between the motor node and the conveyor belt node, while increasing the weight of the edge between the motor node and the reducer node. In this way, the fault propagation relationship in the graph neural network that does not conform to reality is gradually corrected. This allows the graph neural network to dynamically adjust according to the actual fault situation, enhances its ability to learn the fault propagation law, and improves the accuracy and reliability of fault prediction. The adjusted edge weights become the basis for the graph neural network's next prediction, affecting the calculation of subsequent fault propagation probability and the identification of fault propagation paths. At the same time, it also provides real-time optimized data accumulation for periodic weight updates at different time scales.

[0122] The operation of a feeder is a dynamic process. Factors such as component wear, process adjustments, and changes in material properties can slowly alter the fault propagation pattern. A single weight adjustment can only address the current situation and cannot adapt to long-term changes. Therefore, it is necessary to periodically update the edge weights according to different time scales to continuously optimize the graph neural network. The detection system has set update cycles of daily, monthly, and yearly time scales. Daily, it summarizes the fault handling records and feeder operation data of the day, and fine-tunes the weights of edges involved in the fault propagation path to quickly respond to small changes in feeder operation, including minor changes in fault propagation probability caused by slight component wear during daily operation. Monthly, it combines monthly equipment maintenance data and production process adjustments to further optimize the graph neural network. A comprehensive optimization is performed, with significant adjustments made to edges with large weight deviations to adapt to the feeder's operational changes within a month. Each year, based on the feeder's full lifecycle data and information from major process improvements, the graph neural network is retrained, and the edge weight relationships are reconstructed. This allows the graph neural network to adapt to changes in fault propagation patterns caused by equipment aging and process innovations. This ensures the graph neural network can continuously track changes in the feeder's state, constantly adapt to new fault propagation patterns, and maintain the accuracy and effectiveness of fault prediction over the long term. It avoids fault prediction failure due to graph neural network lag. The updated edge weights provide a more accurate fault propagation relationship for the fault detection module, ensuring the entire fault detection and handling system operates efficiently at all times.

[0123] Example 2

[0124] like Figure 4 The diagram shown is a flowchart of a method for fault detection and handling of a feeder, as provided in an embodiment of this application. The method includes:

[0125] The system acquires multi-source data from the feeder in real time, including vibration signals, acoustic signals, and motor current. It processes the multi-source data using a wavelet packet decomposition algorithm and extracts and verifies the multi-source features.

[0126] The fault threshold is dynamically corrected based on the density and humidity of the material. When the multi-source characteristics are greater than or equal to the fault threshold, the fault type and fault level are output.

[0127] When the multi-source features are less than the fault threshold, the multi-source features are processed by a graph neural network to generate the fault propagation probability of the feeder.

[0128] Configure an early warning threshold. When the probability of fault propagation is greater than or equal to the early warning threshold, output the fault propagation path and the probability of fault propagation.

[0129] A multi-level response mechanism is generated based on the output fault type, fault level, fault propagation path, and fault propagation probability, and the feeder is controlled based on the multi-level response mechanism.

[0130] Monitor the performance of the feeder to update the fault threshold and the weights of the edges in the graph neural network.

[0131] Since the principle of the method in this application embodiment is similar to that of the system described in this application embodiment, the implementation of the method is the same as that of the system, and the repeated parts will not be described again.

Claims

1. A fault detection and handling system for a feeder, characterized in that, include: The module consists of a sensor fusion module, a fault detection module, and an execution processing module. The sensor fusion module is used to acquire multi-source data of the feeder in real time. The multi-source data includes vibration signals, acoustic signals and motor current. The multi-source data is processed by wavelet packet decomposition algorithm, and multi-source features are extracted and verified. The multi-source features include equipment surface defects, material collision acoustics and motor harmonic distortion. The fault detection module is used to dynamically correct the fault threshold based on the density and humidity of the material. When the multi-source features are greater than or equal to the fault threshold, the fault type and fault level are output. When the multi-source features are less than the fault threshold, the multi-source features are processed by a graph neural network to generate the fault propagation probability of the feeder. The nodes of the graph neural network represent the physical components of the feeder, and the edge weights are generated by training from historical fault data. At the same time, a warning threshold is configured. When the fault propagation probability is greater than or equal to the warning threshold, the fault propagation path and fault propagation probability are output. The execution processing module is used to generate a multi-level response mechanism based on the output fault type and fault level, as well as the fault propagation path and fault propagation probability. Based on the multi-level response mechanism, the feeder is controlled, and the execution effect of the feeder is monitored to update the fault threshold and the weights of the edges of the graph neural network.

2. The fault detection and handling system for a feeder as described in claim 1, characterized in that, The extraction logic for the multi-source features includes: The acquired vibration signal, acoustic signal, and motor current are normalized, and the number of decomposition layers is dynamically adjusted according to the information entropy change rate of each multi-source data using an adaptive wavelet packet decomposition algorithm to obtain the decomposed frequency band signal. Calculate the energy entropy of each frequency band signal after decomposition to generate energy entropy matrices for vibration signal, acoustic signature signal and motor current respectively; Weights are assigned to vibration signals, acoustic signals, and motor current using an attention mechanism. The energy entropy matrices of vibration signals, acoustic signals, and motor current are then fused using tensors to form a feature tensor. This feature tensor is then processed by a convolutional neural network to extract multi-source features.

3. The fault detection and handling system for a feeder as described in claim 2, characterized in that, The verification logic for the multi-source features includes: The correlation between any two of the following is calculated: surface defects of equipment, sound patterns of material collisions, and harmonic distortion of motor. A correlation matrix is ​​generated, a correlation threshold is configured, and the correlation matrix is ​​compared with the correlation threshold to obtain the verification result of modal consistency. Set a sliding time window to monitor in real time the difference between the mean of multi-source features in the current time window and the mean of multi-source features under historical normal conditions, and obtain time series verification results based on the difference. The physical constraints are determined based on the mechanical principle of the feeder, and the monitored multi-source features are compared with the physical constraints to obtain the verification results of the physical constraints.

4. The fault detection and handling system for a feeder as described in claim 3, characterized in that, The output logic of the fault propagation path includes: The fault propagation probability is transformed into the conditional probability of a Bayesian network, where the physical components of the feeder represent the nodes of the Bayesian network, and the fault propagation relationship represents the directed edges of the Bayesian network. The reliability of each fault propagation path is evaluated based on conditional probability using a probabilistic reasoning algorithm. The reliability score of each fault propagation path is calculated by combining the reliability of each fault propagation path. The fault propagation paths are then sorted in descending order according to their reliability scores to output the fault propagation paths.

5. The fault detection and handling system for a feeder as described in claim 4, characterized in that, The fault threshold correction sub-logic includes: Extract the correlation patterns between material density and moisture content and feeder malfunctions from historical fault data, divide the material density and moisture content into multiple intervals, and form grid cells by intersection. Each grid cell corresponds to a combination of density and moisture content. The multi-source characteristics of historical faults in each grid cell are statistically analyzed, and fault thresholds are configured based on the correlation patterns between material density and moisture and feeder faults. The system monitors the feeder's operating status in real time, adjusts the fault threshold based on the feeder's operating status, monitors the measurement error of multi-source data to determine the confidence interval for the fault threshold adjustment, and expands the fluctuation range of the fault threshold based on the confidence interval for the fault threshold adjustment in order to correct the fault threshold.

6. The fault detection and handling system for a feeder as described in claim 5, characterized in that, The sub-logic for generating the fault propagation probability includes: The physical components of the feeder are mapped to nodes of a graph neural network. The multi-source features of the feeder characterize the node features. The probability of fault propagation between physical components is calculated based on historical fault data and used as the weight of the edges of the graph neural network. The node features are propagated according to the edge weights through spatiotemporal convolution operations to obtain the fault propagation relationship; In a graph neural network, fault injection of nodes is simulated, and the changes in node characteristics of other nodes are observed to quantify the causal effect of the fault. By comparing the changes in node characteristics in the normal state and the fault injection state, the fault propagation path can be identified. By combining fault propagation relationships, fault causal effects, and fault propagation paths, the probability distribution of individual physical components is generated, and the probability distribution of the physical components of the feeder is integrated to generate the fault propagation probability.

7. The fault detection and handling system for a feeder as described in claim 6, characterized in that, The multi-level response mechanism includes: Based on the output fault type and fault level, as well as the fault propagation path and fault propagation probability, the fault risk of the feeder is classified into risk levels, and a response strategy is matched according to the risk level. Based on a comprehensive comparison of the cost-benefit of each response strategy, including downtime costs, repair costs, and risk costs, a response strategy is selected and synchronously transmitted to the feeder.

8. The fault detection and handling system for a feeder as described in claim 7, characterized in that, The update logic for the fault threshold includes: The feeder is controlled based on a multi-level response mechanism, and the execution effect of the feeder is monitored. The fault threshold is adjusted according to whether the fault of the feeder is eliminated. Real-time monitoring of the feeder's operating status, as well as the material's density and moisture content, allows for dynamic adjustment of fault thresholds. After each adjustment of the fault threshold, a fault is simulated and injected to verify whether the adjusted fault threshold can trigger the multi-level response mechanism. If the verification fails, the fault threshold is rolled back and readjusted.

9. The fault detection and handling system for a feeder as described in claim 8, characterized in that, The update logic for the edge weights includes: The feeder is controlled based on a multi-level response mechanism. The execution effect of the feeder is monitored, and the actual fault propagation path is compared with the identified fault propagation path to obtain the path judgment result. When the path determination result indicates that the paths are consistent, increase the weight of the edges of the fault propagation path; When the path determination result indicates that the paths are inconsistent, the weights of the edges in the fault propagation path are reduced, and the weights of the edges in the graph neural network are updated periodically according to different time scales.

10. A method for fault detection and handling of a feeder, implemented based on a fault detection and handling system for a feeder according to any one of claims 1-9, characterized in that, include: The system acquires multi-source data from the feeder in real time, including vibration signals, acoustic signals, and motor current. It processes the multi-source data using a wavelet packet decomposition algorithm and extracts and verifies the multi-source features. The fault threshold is dynamically corrected based on the density and moisture content of the material. When the multi-source features are greater than or equal to the fault threshold, the fault type and fault level are output. When the multi-source features are less than the fault threshold, the multi-source features are processed by a graph neural network to generate the fault propagation probability of the feeder. An early warning threshold is configured. When the fault propagation probability is greater than or equal to the early warning threshold, the fault propagation path and fault propagation probability are output. A multi-level response mechanism is generated based on the output fault type, fault level, fault propagation path, and fault propagation probability. The feeder is controlled based on the multi-level response mechanism. The execution effect of the feeder is monitored to update the fault threshold and the weights of the edges of the graph neural network.

Images