Real-time monitoring and fault diagnosis and early warning method and system for multi-parameters of automated production lines

Abstract

This invention provides a method and system for real-time multi-parameter monitoring and fault diagnosis and early warning for automated production lines. This system involves acquiring multi-source sensor data, performing data quality assessment and correction, monitoring equipment component performance parameters, and calculating degradation rates. This system then identifies fault types through time-frequency domain feature analysis, and establishes fault propagation links based on physical connections to determine the source of the fault. This system can promptly detect equipment anomalies, accurately locate the fault source component, improve the accuracy and timeliness of fault diagnosis, and effectively reduce equipment maintenance costs.

Description

Technical Field

[0001] The present invention relates to automated production line monitoring technology, and in particular to a method and system for real-time multi-parameter monitoring and fault diagnosis and early warning of automated production lines. Background Art

[0002] With the continuous improvement of industrial automation, automated production lines are becoming increasingly popular in the manufacturing industry. To ensure stable production line operation, multi-source sensors are often used to monitor key equipment and process parameters in real time. However, due to factors such as sensor failure and data transmission anomalies, the reliability of collected data is difficult to guarantee, affecting the accuracy of fault diagnosis.

[0003] Existing fault diagnosis methods primarily analyze single devices or local parameters, failing to effectively identify fault correlations and propagation characteristics between device components. Furthermore, traditional methods suffer from incomplete feature extraction and inaccurate diagnostic results when processing multi-source heterogeneous data, making them difficult to meet the fault diagnosis needs of complex industrial scenarios.

[0004] To address the problem of fault diagnosis of equipment components in automated production lines, there is an urgent need for a comprehensive solution that can perform reliability assessment and correction on multi-source sensor data, analyze the fault propagation chain based on the physical connection relationship of equipment components, and accurately locate the fault source and identify the fault type. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for real-time monitoring of multiple parameters and fault diagnosis and early warning of an automated production line, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for real-time monitoring of multiple parameters and fault diagnosis and early warning of an automated production line is provided, comprising:

[0008] Acquire real-time operating data collected by multi-source sensors in an automated production line, calculate data quality characteristic values, generate data credibility, and when the data credibility is less than a first preset threshold, perform data correction based on historical data at adjacent moments to obtain corrected real-time operating data;

[0009] Grouping the corrected real-time operating data according to device components to generate operating parameter sequences corresponding to the device components, calculating performance parameters of the device components based on the operating parameter sequences, calculating a performance degradation rate based on the performance parameters, and determining the corresponding standby component as a component to be monitored when the performance degradation rate is greater than a second preset threshold;

[0010] Performing time-frequency domain feature analysis on the operating parameter sequence of the component to be monitored, extracting time-domain statistical features and frequency-domain energy features as fault features, inputting the fault features into a pre-trained fault recognition model, outputting a probability distribution of fault types based on a combination pattern of the fault features, and selecting the fault type with the highest probability as the fault determination result of the component to be monitored;

[0011] According to the physical connection relationship between the component to be monitored and the adjacent components, the performance degradation rate of the adjacent components is obtained, and a fault propagation link is established based on the temporal change sequence of the performance degradation rate. The fault source component is determined, and a fault diagnosis report containing the location information of the fault source component and the fault judgment result is generated and sent to the production management system.

[0012] In an optional embodiment,

[0013] Acquire real-time operation data collected by multi-source sensors in an automated production line, calculate data quality characteristic values, and generate data credibility. When the data credibility is less than a first preset threshold, perform data correction based on historical data at adjacent moments. The corrected real-time operation data includes:

[0014] Acquire real-time operating data collected by multi-source sensors in an automated production line and segment it according to time windows to obtain a segmented data sequence;

[0015] Calculate the signal fluctuation, value variation, and sampling integrity of the segmented data sequence respectively and combine them to obtain data quality characteristic values; calculate the information entropy of the data quality characteristic values ​​based on the probability distribution characteristics of the data in the segmented data sequence; generate a dynamic weight coefficient based on the information entropy; and perform a weighted combination of the dynamic weight coefficient and the corresponding data quality characteristic value to obtain data credibility;

[0016] When the data credibility is less than a first preset threshold, the data to be corrected is determined, and a historical data sequence of adjacent moments of N time windows before and after the time window corresponding to the data to be corrected is extracted;

[0017] Clustering the historical data sequences at adjacent moments according to the similarity of data change trends to obtain multiple subsequence groups, calculating the time correlation between each subsequence in each subsequence group and the data to be corrected, and constructing a recursive optimization model with differentiated weights based on the time correlation, where the weights are positively correlated with the time correlation;

[0018] A recursive optimization model is used to iteratively correct the data to be corrected. In each round of iteration, the fitting error between the correction result and the subsequence whose correlation is greater than a preset correlation threshold is calculated. When the fitting error is less than a dynamic threshold, the correction result is output as the corrected real-time running data.

[0019] In an optional embodiment,

[0020] Calculating the information entropy of the data quality characteristic value based on the probability distribution characteristics of the data in the segmented data sequence, and generating a dynamic weight coefficient according to the information entropy includes:

[0021] Constructing a probability density estimation function, calculating the standard deviation and the interquartile range based on the segmented data sequence, and dynamically adjusting the bandwidth coefficient of the probability density estimation function according to the ratio of the standard deviation to the interquartile range to obtain a continuous probability density curve for the segmented data sequence;

[0022] Performing piecewise integration on the continuous probability density curve to obtain a cumulative probability value for each segment of data, dividing the segmented data sequence into a first distribution area and a second distribution area according to the cumulative probability value, and calculating a cumulative probability ratio of the first distribution area to the second distribution area as a data distribution characteristic ratio;

[0023] The overlapping sliding window step size is set using the data distribution characteristic ratio, and a first local information entropy is obtained by sliding calculation within the first distribution area; the non-overlapping sliding window width is set using the inverse of the data distribution characteristic ratio, and a second local information entropy is obtained by calculation within the second distribution area;

[0024] Taking the cumulative probability values ​​of the first distribution area and the second distribution area as weight coefficients, weighting the first local information entropy and the second local information entropy to obtain a regional combined entropy;

[0025] The global information entropy is calculated for the continuous probability density curve, the regional combined entropy and the global information entropy are normalized to obtain the final information entropy of the data quality characteristic value, and a dynamic weight coefficient is generated according to the final information entropy, where the dynamic weight coefficient is inversely proportional to the final information entropy.

[0026] In an optional embodiment,

[0027] Calculating performance parameters of the device components based on the operating parameter sequence, calculating a performance degradation rate based on the performance parameters, and determining the corresponding standby component as a component to be monitored when the performance degradation rate is greater than a second preset threshold includes:

[0028] Performing time-frequency analysis on the operating parameter sequence to obtain energy distribution characteristics, selecting a characteristic parameter combination based on the energy distribution characteristics, and constructing the characteristic parameter combination into a comprehensive performance status indicator;

[0029] Analyze the temporal variation pattern of the comprehensive performance status indicator, establish a performance status transfer matrix, calculate the change rate of the performance parameter based on the performance status transfer matrix, and correct the change rate in combination with the performance status transfer matrix to obtain the performance degradation rate;

[0030] Establish a performance benchmark value for the normal operation of the component, determine the performance fluctuation tolerance range based on the performance benchmark value combined with historical operation data, start the cumulative time count when it is detected that the performance degradation rate exceeds the performance fluctuation tolerance range, and when the cumulative time reaches the preset period, determine the corresponding equipment component as the component to be monitored.

[0031] In an optional embodiment,

[0032] Establishing a performance baseline for a component during normal operation and determining the performance fluctuation tolerance based on the performance baseline combined with historical operating data include:

[0033] Obtaining historical operating data of the device components, identifying normal operating periods and abnormal operating periods from the historical operating data, performing comparative analysis on the normal operating periods and abnormal operating periods, extracting false positives and false negatives, and constructing a reward function;

[0034] Inputting the reward function and the performance data of the normal operation period into a reinforcement learning framework, obtaining a performance benchmark value through reinforcement learning training, and generating an initial tolerance adjustment strategy based on the performance benchmark value;

[0035] Collect real-time operating status data of components, match and analyze the real-time operating status data with the initial tolerance adjustment strategy, calculate the status warning probability under different tolerance settings, and generate risk prediction results within the tolerance range;

[0036] Obtaining the component's location information in the production process, process parameter requirements for the process, and historical maintenance cost data, and performing a multi-objective optimization calculation on the location information, process parameter requirements, and historical maintenance cost data with the risk prediction results to generate a tolerance optimization objective function;

[0037] The initial tolerance adjustment strategy is optimized based on the tolerance optimization objective function to obtain a performance fluctuation tolerance range, the performance fluctuation tolerance range is combined with a performance benchmark value to form a monitoring threshold, and the state data corresponding to the monitoring threshold is fed back to the reinforcement learning framework for continuous optimization training.

[0038] In an optional embodiment,

[0039] The fault characteristics are input into the pre-trained fault identification model. The fault type probability distribution is output based on the combination pattern of the fault characteristics. The fault type with the highest probability is selected as the fault judgment result of the component to be monitored. The results include:

[0040] Inputting the fault features into a pre-trained fault recognition model, decomposing the fault features into common features and individual features through a feature decoupling network, and applying orthogonal constraints to the common features and individual features to obtain decoupled features;

[0041] The decoupled features are constructed as an attribute graph, the feature dimensions are set as graph nodes, and the feature association strength is set as graph edges. Local and global dependencies are extracted through graph convolution and multi-head attention mechanisms to obtain a combination pattern of fault features.

[0042] The combination pattern of fault features and decoupling features are input into the variational inference module, and a state transfer matrix is ​​established based on the feature correlation strength and orthogonal constraints to obtain the initial fault type probability;

[0043] The initial fault type probability is subjected to Bayesian modeling, and probability calibration is performed based on the local dependency and global dependency in the combination pattern of the fault characteristics to generate a calibrated fault type probability distribution. The fault type with the highest probability is selected from the calibrated fault type probability distribution as the fault judgment result of the component to be monitored.

[0044] In an optional embodiment,

[0045] Based on the physical connection relationship between the component to be monitored and its adjacent components, the performance degradation rate of the adjacent components is obtained. Based on the temporal change sequence of the performance degradation rate, a fault propagation link is established to determine the fault source component, including:

[0046] Obtain the mechanical coupling relationship, energy transfer relationship, and physical topological relationship between the component to be monitored and its adjacent components, and establish a component connection relationship matrix that characterizes the spatial connection strength between components;

[0047] Acquiring performance parameters of the component to be monitored and adjacent components within a continuous time window, calculating time derivatives of the performance parameters to obtain a performance degradation rate, and performing a Hilbert transform on the performance degradation rate to obtain an instantaneous phase;

[0048] According to the spatial connection strength and instantaneous phase variation trend in the component connection relationship matrix, the phase synchronization strength between adjacent components is calculated, and a network order parameter reflecting the dynamic coupling relationship between components is constructed;

[0049] monitoring the dynamic changes of the network order parameters, obtaining the mutation time and mutation amplitude of the network order parameters, and constructing a directed fault propagation graph based on the mutation time, mutation amplitude, and phase synchronization strength, wherein adjacent components are used as nodes, the time difference of the mutation time is used as the edge weight, and the phase synchronization strength is used as the edge direction;

[0050] Calculating the downstream impact range of each component based on the edge weights and edge directions in the directed fault propagation graph, determining the cascading failure scale of each component based on the downstream impact range, and combining the cascading failure scale with the mutation moment to obtain a failure impact factor;

[0051] According to the dynamic evolution law of the directed fault propagation graph, the mutation time characteristics and cascading failure scale in the failure impact factors are weighted, and the component with the largest failure impact factor and the earliest mutation time after configuration is determined as the fault source component.

[0052] According to a second aspect of the embodiments of the present invention,

[0053] Provides a multi-parameter real-time monitoring and fault diagnosis and early warning system for an automated production line, including:

[0054] The first unit is configured to obtain real-time operation data collected by multi-source sensors in an automated production line, calculate data quality characteristic values, generate data credibility, and, when the data credibility is less than a first preset threshold, perform data correction based on historical data at adjacent moments to obtain corrected real-time operation data;

[0055] a second unit, configured to group the corrected real-time operating data according to equipment components, generate operating parameter sequences corresponding to the equipment components, calculate performance parameters of the equipment components based on the operating parameter sequences, calculate a performance degradation rate based on the performance parameters, and determine the corresponding standby component as a component to be monitored when the performance degradation rate is greater than a second preset threshold;

[0056] The third unit is configured to perform time-frequency domain feature analysis on the operating parameter sequence of the component to be monitored, extract time-domain statistical features and frequency-domain energy features as fault features, input the fault features into a pre-trained fault recognition model, output a probability distribution of fault types based on a combination pattern of the fault features, and select the fault type with the highest probability as the fault determination result of the component to be monitored;

[0057] The fourth unit is used to obtain the performance degradation rate of adjacent components based on the physical connection relationship between the component to be monitored and the adjacent components, establish a fault propagation link based on the time series change sequence of the performance degradation rate, determine the fault source component, generate a fault diagnosis report containing the location information of the fault source component and the fault judgment result, and send it to the production management system.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] According to a fourth aspect of the embodiments of the present invention,

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0065] In this embodiment, by performing quality assessment and correction on the data collected by multi-source sensors on the automated production line, the reliability of the data is ensured, misjudgments due to data anomalies are avoided, and the accuracy and credibility of fault diagnosis are improved. Fault monitoring is performed based on the performance parameters and degradation rates of equipment components, and combined with time-frequency domain feature analysis methods, fault features are extracted from multiple dimensions. Precise identification of fault types is achieved through pre-trained fault identification models, improving the efficiency and accuracy of fault diagnosis. By analyzing the physical connection relationships between equipment components and the temporal variation patterns of performance degradation rates, a fault propagation link is established, the source of the fault is accurately located, and misjudgments due to fault propagation are effectively avoided. This provides a reliable decision-making basis for equipment maintenance and fault handling, and improves the operational reliability and maintenance efficiency of the production line. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 Schematic diagram of the process of the method for real-time monitoring of multiple parameters and fault diagnosis and early warning of an automated production line according to an embodiment of the present invention;

[0067] Figure 2 This is a schematic diagram of the probability distribution area division according to an embodiment of the present invention;

[0068] Figure 3 This is a comparison chart of early warning accuracy indicators of different methods in embodiments of the present invention. DETAILED DESCRIPTION

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0070] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0071] Figure 1 FIG. 1 is a flow chart of a method for real-time monitoring of multiple parameters and fault diagnosis and early warning of an automated production line according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0072] Acquire real-time operating data collected by multi-source sensors in an automated production line, calculate data quality characteristic values, generate data credibility, and when the data credibility is less than a first preset threshold, perform data correction based on historical data at adjacent moments to obtain corrected real-time operating data;

[0073] Grouping the corrected real-time operating data according to device components to generate operating parameter sequences corresponding to the device components, calculating performance parameters of the device components based on the operating parameter sequences, calculating a performance degradation rate based on the performance parameters, and determining the corresponding standby component as a component to be monitored when the performance degradation rate is greater than a second preset threshold;

[0074] Performing time-frequency domain feature analysis on the operating parameter sequence of the component to be monitored, extracting time-domain statistical features and frequency-domain energy features as fault features, inputting the fault features into a pre-trained fault recognition model, outputting a probability distribution of fault types based on a combination pattern of the fault features, and selecting the fault type with the highest probability as the fault determination result of the component to be monitored;

[0075] According to the physical connection relationship between the component to be monitored and the adjacent components, the performance degradation rate of the adjacent components is obtained, and a fault propagation link is established based on the temporal change sequence of the performance degradation rate. The fault source component is determined, and a fault diagnosis report containing the location information of the fault source component and the fault judgment result is generated and sent to the production management system.

[0076] In an optional embodiment,

[0077] Acquire real-time operation data collected by multi-source sensors in an automated production line, calculate data quality characteristic values, and generate data credibility. When the data credibility is less than a first preset threshold, perform data correction based on historical data at adjacent moments. The corrected real-time operation data includes:

[0078] Acquire real-time operating data collected by multi-source sensors in an automated production line and segment it according to time windows to obtain a segmented data sequence;

[0079] Calculate the signal fluctuation, value variation, and sampling integrity of the segmented data sequence respectively and combine them to obtain data quality characteristic values; calculate the information entropy of the data quality characteristic values ​​based on the probability distribution characteristics of the data in the segmented data sequence; generate a dynamic weight coefficient based on the information entropy; and perform a weighted combination of the dynamic weight coefficient and the corresponding data quality characteristic value to obtain data credibility;

[0080] When the data credibility is less than a first preset threshold, the data to be corrected is determined, and a historical data sequence of adjacent moments of N time windows before and after the time window corresponding to the data to be corrected is extracted;

[0081] Clustering the historical data sequences at adjacent moments according to the similarity of data change trends to obtain multiple subsequence groups, calculating the time correlation between each subsequence in each subsequence group and the data to be corrected, and constructing a recursive optimization model with differentiated weights based on the time correlation, where the weights are positively correlated with the time correlation;

[0082] A recursive optimization model is used to iteratively correct the data to be corrected. In each round of iteration, the fitting error between the correction result and the subsequence whose correlation is greater than a preset correlation threshold is calculated. When the fitting error is less than a dynamic threshold, the correction result is output as the corrected real-time running data.

[0083] For example, in the fault diagnosis process of an automated production line, real-time operational data collected by multiple sensors must first be preprocessed and quality assessed. The data is segmented by setting a fixed-length time window. The length of the time window is determined based on the dynamic response characteristics of the equipment components. For example, for a robotic welding station, the time window can be set to one welding cycle to ensure the integrity of the welding process. This ensures that the segmented data sequence fully reflects the changes in the equipment's operating status.

[0084] The calculation of data quality characteristic values ​​includes three aspects: signal volatility reflects the stability of the data sequence, which is specifically characterized by calculating the degree of dispersion of all numerical values ​​in the data sequence, that is, first calculating the average value of the entire sequence, then calculating the deviation of each data point from the average value, and finally obtaining the ratio of the standard deviation to the mean; numerical variability represents the dynamic change characteristics of the data, which is measured by calculating the numerical difference between adjacent sampling points, specifically calculating the difference between two adjacent data points, obtaining the difference sequence and then calculating its root mean square value; sampling completeness reflects the integrity of the data, which is measured by the ratio of the actual number of sampling points to the number of sampling points that should be obtained according to the sampling frequency theory. The closer the value is to one, the more complete the data.

[0085] To achieve a reasonable assessment of data quality characteristics, a dynamic weight assignment mechanism based on information entropy is introduced. First, a statistical analysis of the segmented data sequence is performed, dividing the data range into several equally spaced intervals. The frequency of occurrence of data points within each interval is counted, and a probability distribution is calculated. This basis is then used to calculate the information entropy of the data distribution. Information entropy reflects the uncertainty of the data distribution. This calculation is performed by substituting the probability values ​​of each interval into the information entropy calculation formula. Larger information entropy values ​​indicate a more dispersed data distribution and higher uncertainty, and the weight of the corresponding feature should be reduced.

[0086] In the process of generating dynamic weight coefficients based on information entropy values, the information entropy values ​​corresponding to each quality feature are first calculated and normalized. The initial weight is then subtracted from the normalized information entropy value to obtain the initial weight. Finally, the initial weight is normalized to obtain the final dynamic weight coefficient. The dynamic weight coefficient is multiplied by the corresponding data quality feature value and the sum is calculated to obtain a comprehensive data credibility index.

[0087] When data credibility falls below a preset threshold, the data needs to be corrected. First, the data segment to be corrected is identified, and historical data for the time window before and after that segment is extracted. The amount of historical data extracted is determined by the cyclical nature of the equipment's operation, typically selecting a time range that covers the entire equipment's operating cycle.

[0088] When performing cluster analysis on extracted historical data series, it's necessary to extract trend features. These include the distribution of rising, falling, and plateau segments, as well as the slope characteristics of each segment. Based on these features, similarity is calculated between different data series. This similarity calculation considers both the degree of matching and the magnitude differences in trend features. A density-based clustering algorithm is then used to group highly similar series together.

[0089] When calculating the temporal correlation between the subsequence and the data to be corrected, a weighted approach based on time decay is employed. Specifically, the time interval between the subsequence and the data to be corrected is calculated and then substituted into a decreasing function to obtain a temporal weight. The larger the time interval, the smaller the weight. Simultaneously, the trend similarity between the subsequence and the data to be corrected is calculated, and the final temporal correlation is obtained by multiplying the time weight by the trend similarity.

[0090] When constructing a recursive optimization model, temporal correlation is used as the weight coefficient for each subsequence's reference value. The model's optimization goal is to ensure that the corrected data maintains a consistent trend with highly correlated historical data while also satisfying the data's physical constraints. In each iteration, a weighted error is calculated between the current corrected value and each reference sequence, while also checking whether the corrected value meets constraints such as numerical range and rate of change.

[0091] The dynamic threshold is determined by taking into account data fluctuations and calculating the fluctuation range of the data within a group of highly correlated subsequences. When the fitting error between the correction result and the highly correlated subsequences is less than the dynamic threshold, and the correction value changes little over multiple iterations, the correction process is considered converged and the current correction result is output as the final correction data.

[0092] This embodiment achieves real-time quality assessment and dynamic correction of multi-source sensor data, improving the accuracy and reliability of automated production line data. Data quality is comprehensively assessed using signal volatility, value variability, and sampling integrity. Information entropy is introduced to dynamically adjust feature weights, ensuring that the assessment results adapt to diverse data distribution characteristics. For low-reliability data, clustering is performed using trend characteristics of historical data sequences, and a recursive optimization model is constructed based on temporal correlation. This ensures that correction results align with actual operating conditions, reduces errors and noise interference, and effectively improves the stability and accuracy of fault diagnosis and process control.

[0093] In an optional embodiment,

[0094] Calculating the information entropy of the data quality characteristic value based on the probability distribution characteristics of the data in the segmented data sequence, and generating a dynamic weight coefficient according to the information entropy includes:

[0095] Constructing a probability density estimation function, calculating the standard deviation and the interquartile range based on the segmented data sequence, and dynamically adjusting the bandwidth coefficient of the probability density estimation function according to the ratio of the standard deviation to the interquartile range to obtain a continuous probability density curve for the segmented data sequence;

[0096] Performing piecewise integration on the continuous probability density curve to obtain a cumulative probability value for each segment of data, dividing the segmented data sequence into a first distribution area and a second distribution area according to the cumulative probability value, and calculating a cumulative probability ratio of the first distribution area to the second distribution area as a data distribution characteristic ratio;

[0097] The overlapping sliding window step size is set using the data distribution characteristic ratio, and a first local information entropy is obtained by sliding calculation within the first distribution area; the non-overlapping sliding window width is set using the inverse of the data distribution characteristic ratio, and a second local information entropy is obtained by calculation within the second distribution area;

[0098] Taking the cumulative probability values ​​of the first distribution area and the second distribution area as weight coefficients, weighting the first local information entropy and the second local information entropy to obtain a regional combined entropy;

[0099] The global information entropy is calculated for the continuous probability density curve, the regional combined entropy and the global information entropy are normalized to obtain the final information entropy of the data quality characteristic value, and a dynamic weight coefficient is generated according to the final information entropy, where the dynamic weight coefficient is inversely proportional to the final information entropy.

[0100] For example, a kernel density estimation function is first constructed to evaluate the data distribution characteristics. For each segmented data sequence collected by each sensor, such as the speed data of the robot joint and the current data of the motor, its statistical characteristics are calculated. The standard deviation is calculated by summing the square root of the difference between each value in the data sequence and the mean; the interquartile range is calculated by sorting the data sequence by size and finding the difference between the 75th percentile and the 25th percentile. The standard deviation is divided by the interquartile range to obtain a ratio, which is used to adjust the bandwidth coefficient of the kernel density estimation function. When the ratio is greater than two, it indicates that the overall fluctuation of the data is significantly greater than the fluctuation in the middle area, and the bandwidth coefficient needs to be increased by 50%; when the ratio is less than 0.5, it indicates that the data distribution is concentrated, and the bandwidth coefficient is reduced by 30%.

[0101] Using the adjusted bandwidth coefficient, kernel density estimation is performed on the data sequence, resulting in a continuous probability density curve reflecting the data distribution. The probability density curve represents the density of data at different locations, with higher values ​​indicating denser data. The entire data range is divided into multiple intervals, and the probability density curve is integrated within each interval to obtain the cumulative probability value for that interval.

[0102] The data series is divided into regions based on the cumulative probability values. The cumulative probability values ​​are sorted from largest to smallest. The set of intervals where the sum of the probability values ​​reaches 70% is divided into the first distribution region, representing the main distribution area of ​​the data. The remaining intervals are divided into the second distribution region, representing the sparse distribution area of ​​the data. The ratio of the cumulative probability value of the first distribution region to the cumulative probability value of the second distribution region is calculated to obtain the data distribution characteristic ratio.

[0103] The local information entropy is calculated within the first distribution region using overlapping sliding windows. The window step size is set to the product of the data distribution characteristic ratio and the base step size, with the base step size being 5 percent of the data sequence length. For example, when the data distribution characteristic ratio is 2.5, the sliding step size is 2.5 times the base step size. Within each sliding window, the probability distribution of the data points is measured to calculate the first local information entropy.

[0104] For the second distribution region, a non-overlapping sliding window is used. The window width is set to the data sequence length divided by the data distribution characteristic ratio to ensure reasonable division of the sparse region. For example, if the data distribution characteristic ratio is 2, the window width is half the data sequence length. The probability distribution of the data within each window is calculated to obtain the second local information entropy.

[0105] Perform a weighted combination of the first and second local information entropies. The weight coefficients are based on the cumulative probability values ​​of their respective regions. For example, if the cumulative probability value of the first distribution region is 70%, the weight of the first local information entropy is 0.7. Add the two weighted entropy values ​​to obtain the regional combined entropy.

[0106] The global information entropy is calculated for the entire probability density curve, covering the entire range of the data sequence. The regional and global information entropies are normalized by dividing each by its maximum possible value, ensuring that both entropies are within the range of zero to one. The normalized regional and global information entropies are combined with the global information entropy in a ratio of six to four to obtain the final information entropy.

[0107] Finally, the final information entropy is converted into a dynamic weight coefficient. By subtracting the final information entropy from one and then normalizing, the weight coefficient is inversely proportional to the information entropy. For example, when the final information entropy of a data quality feature is 0.8, its corresponding initial weight is 0.2. The initial weights of all features are then normalized to obtain the final dynamic weight coefficient. These weight coefficients are used in subsequent calculations of data credibility, providing an objective basis for data quality assessment.

[0108] Figure 2 This is a schematic diagram of the probability distribution area division of an embodiment of the present invention, as shown in FIG. Figure 2 The figure shows the region partitioning and information entropy calculation strategy for bimodal data. The horizontal axis represents data values ​​(0-125), and the vertical axis represents probability density (0-0.04). It can be observed that the data exhibits a distinct bimodal distribution, with probability density peaks around values ​​25 and 75, respectively. Based on the region partitioning principle of the proposed method, the system divides the data into five regions: two primary distribution regions (located around data values ​​25 and 75) and three secondary distribution regions (located in the data ranges 0-15, 35-65, and 90-125). By piecewise integration of the probability density curve, the cumulative probabilities in the primary regions are calculated to be 30% and 40%, respectively, for a total of 70%. The cumulative probabilities in the secondary regions are calculated to be 5%, 15%, and 10%, respectively, for a total of 30%. Different information entropy calculation strategies are used for different distribution regions: an overlapping sliding window method is used in the primary distribution region (data-dense area) to achieve refined local information entropy calculation; a non-overlapping window method is used in the secondary distribution region (data-sparse area) to ensure effective probability distribution characteristics even in sparse data. This adaptive region division and calculation strategy based on data distribution characteristics fully considers the characteristics of complex distribution data and can accurately capture the information entropy characteristics of each region. Compared with the traditional global information entropy calculation method, it has obvious advantages, especially for unconventional distribution data such as multi-peak distribution and skewed distribution.

[0109] Prior art approaches to quality assessment of multi-source sensor data typically combine multiple feature indices using fixed weights, failing to consider the impact of data distribution characteristics on quality assessment, resulting in inaccurate assessment results. Furthermore, during data correction, historical data from adjacent moments are often directly interpolated or fitted, failing to fully exploit the temporal correlation of the data. This significantly impacts the correction effect due to abnormal data. This application introduces a dynamic weight allocation mechanism based on data distribution characteristics, adaptively adjusts the bandwidth coefficient of the kernel density estimate according to the data's probability distribution characteristics, and combines regional and global information entropy to achieve more objective and reasonable data quality assessment. During data correction, a recursive optimization model that considers temporal correlation is employed to cluster and optimize weights of historical data sequences, improving the reliability of the correction results. This technical solution accurately identifies anomalies in multi-source sensor data and effectively corrects them, improving the reliability of data acquisition for automated production lines. Furthermore, the combination of the dynamic weight allocation mechanism and the recursive optimization model enhances the solution's adaptability to different types of abnormal data, providing a more reliable data foundation for production line condition monitoring and fault diagnosis.

[0110] In an optional embodiment,

[0111] Calculating performance parameters of the device components based on the operating parameter sequence, calculating a performance degradation rate based on the performance parameters, and determining the corresponding standby component as a component to be monitored when the performance degradation rate is greater than a second preset threshold includes:

[0112] Performing time-frequency analysis on the operating parameter sequence to obtain energy distribution characteristics, selecting a characteristic parameter combination based on the energy distribution characteristics, and constructing the characteristic parameter combination into a comprehensive performance status indicator;

[0113] Analyze the temporal variation pattern of the comprehensive performance status indicator, establish a performance status transfer matrix, calculate the change rate of the performance parameter based on the performance status transfer matrix, and correct the change rate in combination with the performance status transfer matrix to obtain the performance degradation rate;

[0114] Establish a performance benchmark value for the normal operation of the component, determine the performance fluctuation tolerance range based on the performance benchmark value combined with historical operation data, start the cumulative time count when it is detected that the performance degradation rate exceeds the performance fluctuation tolerance range, and when the cumulative time reaches the preset period, determine the corresponding equipment component as the component to be monitored.

[0115] When monitoring the performance of equipment components on automated production lines, the first step is to perform time-frequency analysis on the collected operating parameter sequences. Taking vibration signals as an example, a short-time Fourier transform (SFT) is used to convert the time-domain signal into the time-frequency domain, revealing the frequency characteristics that reflect the signal's temporal variations. Energy statistics are then performed on the time-frequency plot to calculate the energy distribution ratios in different frequency bands and obtain energy distribution characteristics. These energy distribution characteristics include information such as the energy percentages in each frequency band and the energy concentration in the main frequency bands.

[0116] When selecting characteristic parameter combinations based on energy distribution characteristics, we first analyze the correlation between the energy of each frequency band and the device's performance status. For frequency bands with a high energy contribution and a significant correlation with performance status, we extract characteristic parameters such as average energy value and energy fluctuation. We also consider the energy transfer characteristics between frequency bands and calculate the energy ratio of adjacent frequency bands as a supplementary feature. The selected characteristic parameters are weighted and combined according to their importance to construct a comprehensive indicator reflecting the overall performance status of the device components.

[0117] When analyzing the temporal changes in the performance status composite indicator, we first divide the indicator values ​​into multiple status levels based on their magnitude. By counting the transitions between status levels at adjacent moments, we construct a transition matrix that describes the state transition probabilities. Each element in the transition matrix represents the probability of transitioning from one status level to another, and the eigenvalues ​​of the matrix reflect the stability of the state transitions.

[0118] When calculating the rate of change of performance parameters, the directionality of state transitions must be considered. A greater probability of transitions toward performance degradation indicates a faster rate of performance degradation. The initial rate of change is calculated by calculating the weighted sum of the elements in the transition matrix that reflect performance degradation. The eigenvalues ​​of the transition matrix are also analyzed, and the rate of change is modified based on their distribution characteristics to obtain a more accurate performance degradation rate.

[0119] When establishing performance baselines, we select historical data from equipment components during normal operation and calculate the statistical characteristics of comprehensive performance indicators. By analyzing a large amount of normal operation data, we obtain the average level of the indicators as the baseline value. Based on the fluctuations in historical operating data, we calculate the standard deviation of the indicators, and use the range of multiples of the standard deviation as the tolerance range for performance fluctuations.

[0120] During real-time monitoring, if the performance degradation rate is detected to be outside the tolerance range, a cumulative time counter is activated. This cumulative time count is used to determine whether the performance degradation persists, avoiding misjudgments due to momentary fluctuations. When the cumulative time reaches the preset monitoring period, the device component is confirmed to have a performance anomaly and is marked as a component to be monitored, entering the key monitoring state.

[0121] For example, sensors collect vibration signal sequences from bearing components on an automated production line. Time-frequency analysis of these vibration signals reveals significant differences in energy distribution among low-, mid-, and high-frequency bands. During normal operation, the mid-frequency band accounts for the highest energy share, approximately half of the total energy. As bearing performance begins to deteriorate, the proportion of energy in the high-frequency band gradually increases.

[0122] The proportion of energy in the mid-frequency band, the proportion of energy in the high-frequency band, and their ratio were selected as characteristic parameters to construct a comprehensive performance status index. The index values ​​were divided into five status levels, and a state transition matrix was established using continuous monitoring data. Analysis revealed that when bearing performance begins to deteriorate, the probability of the status level shifting to a worse state increases significantly. Combining the eigenvalue distribution of the transition matrix, a corrected performance degradation rate was derived.

[0123] Based on extensive historical operating data, a performance baseline for bearings operating normally is established. The fluctuation range of the comprehensive performance indicator is typically within plus or minus 15 percent of the baseline value. If the performance degradation rate continuously exceeds this range, and the cumulative duration exceeds two preset operating cycles, the bearing assembly is placed on hold for monitoring and prompt inspection and maintenance.

[0124] In this embodiment, the energy distribution characteristics are extracted through time-frequency analysis to establish a dynamic characteristic parameter combination, thereby achieving a comprehensive evaluation of the performance status of the equipment. The analysis method based on the performance state transfer matrix can accurately capture the gradual characteristics of the performance parameters and improve the prediction accuracy of the performance degradation trend. By establishing an adaptive performance fluctuation tolerance range based on historical data, the adaptability of performance monitoring is improved. The cumulative time counting mechanism effectively filters the impact of instantaneous fluctuations and ensures the stability and reliability of performance degradation judgment. Accurate evaluation and timely warning of the performance status of equipment components are achieved, providing a reliable basis for preventive maintenance decisions and improving the operational reliability and maintenance efficiency of the automated production line.

[0125] In an optional embodiment,

[0126] Establishing a performance baseline for a component during normal operation and determining the performance fluctuation tolerance based on the performance baseline combined with historical operating data include:

[0127] Obtaining historical operating data of the device components, identifying normal operating periods and abnormal operating periods from the historical operating data, performing comparative analysis on the normal operating periods and abnormal operating periods, extracting false positives and false negatives, and constructing a reward function;

[0128] Inputting the reward function and the performance data of the normal operation period into a reinforcement learning framework, obtaining a performance benchmark value through reinforcement learning training, and generating an initial tolerance adjustment strategy based on the performance benchmark value;

[0129] Collect real-time operating status data of components, match and analyze the real-time operating status data with the initial tolerance adjustment strategy, calculate the status warning probability under different tolerance settings, and generate risk prediction results within the tolerance range;

[0130] Obtaining the component's location information in the production process, process parameter requirements for the process, and historical maintenance cost data, and performing a multi-objective optimization calculation on the location information, process parameter requirements, and historical maintenance cost data with the risk prediction results to generate a tolerance optimization objective function;

[0131] The initial tolerance adjustment strategy is optimized based on the tolerance optimization objective function to obtain a performance fluctuation tolerance range, the performance fluctuation tolerance range is combined with a performance benchmark value to form a monitoring threshold, and the state data corresponding to the monitoring threshold is fed back to the reinforcement learning framework for continuous optimization training.

[0132] For example, when acquiring historical operating data for equipment components, it primarily includes performance parameter data, operating condition parameter data, and maintenance record data. Performance parameter data includes time series of physical quantities such as vibration, temperature, and pressure; operating condition parameter data includes operating status information such as equipment speed and load; and maintenance record data includes historical fault information, repair records, and component replacement records.

[0133] A multi-layered data analysis approach is used to identify periods of normal and abnormal operation. First, statistical analysis is performed on performance parameter data, calculating statistics such as mean, standard deviation, skewness, and kurtosis. Data is then stratified based on operating parameters, clustering data under similar operating conditions. Known periods of abnormality are identified based on expert experience and historical maintenance records. A density clustering algorithm is used to detect anomalies in unlabeled data. This algorithm identifies outliers by calculating the local density of data points and their distance from high-density points.

[0134] During the comparative analysis process, test results are compared with maintenance records to count false positives and missed negatives. A false positive refers to a situation where the test result appears abnormal but is actually normal, while a missed negative refers to a situation where an actual abnormality is not detected. The causes of these misjudgments are analyzed based on information such as the time distribution of false positives and missed negatives and the characteristics of the operating conditions.

[0135] The reward function is constructed based on multiple evaluation dimensions. Accurate warnings receive positive rewards proportional to the lead time. False alarms and missed warnings receive negative rewards, with the value of the negative rewards proportional to the impact. The impact is quantified using factors such as equipment downtime and repair costs. The impact of operating conditions is also taken into account, and the penalty for false alarms is appropriately reduced during operating condition transitions.

[0136] The reinforcement learning framework uses a deep Q-network structure. The state space contains the current performance parameter values, recent trend characteristics, and operating condition parameters. The action space is a selectable range of performance benchmark values, and continuous values ​​are converted into a finite number of discrete actions through discretization. The Q-network uses a multi-layer neural network structure. The input layer receives state information, the hidden layer uses a fully connected layer with batch normalization, and the output layer corresponds to the expected reward value for each action. The training process uses an experience replay mechanism to store historical experience in an experience pool. Each training session randomly selects a batch of experience samples, which contain state transition sequences and corresponding reward values. The target Q value is calculated using a temporal difference algorithm, and the network parameters are updated by minimizing the mean squared error between the predicted Q value and the target Q value. To balance exploration and exploitation, a decaying ε-greedy strategy is used to select actions.

[0137] During the matching and analysis of real-time operating status data, the data is first preprocessed, including denoising and standardization. The changing characteristics of performance parameters at different time scales are then calculated, including short-term fluctuations and long-term trends. For each candidate tolerance range, the warning trigger probability is calculated using a sliding window analysis. Multiple indicators are considered comprehensively when calculating the status warning probability. The short-term warning probability is calculated based on the frequency and magnitude of parameter deviations from the tolerance range; the long-term warning probability takes into account the parameter's changing trend and cumulative effect. The short-term and long-term warning probabilities are weighted and combined, with the weights dynamically adjusted based on different operating conditions.

[0138] Analysis of process flow location information includes analysis of the upstream and downstream correlations of equipment. Statistical analysis is performed to determine the impact of equipment failures on related processes, including direct impacts and chain reactions. Process parameter requirements are converted into specific performance indicator constraints, such as tolerance limits corresponding to accuracy levels. Maintenance cost data includes spare parts costs, labor costs, downtime losses, etc., and a mapping relationship is established between costs and failure types and maintenance times. Multi-objective optimization calculations use the Pareto optimization method. Objectives such as early warning timeliness, reliability, and maintenance costs are constructed as optimization objective functions. A set of candidate solutions for the tolerance adjustment strategy is generated through a genetic algorithm, and the performance of each solution on each objective is evaluated. The Pareto optimal solution set is selected through non-dominated sorting, and the final solution is selected based on actual needs.

[0139] The monitoring threshold is continuously optimized using an online learning mechanism. New monitoring data and maintenance records are regularly collected, and samples in the experience pool are updated. The reinforcement learning model is incrementally trained using these new samples, adjusting network parameters. Furthermore, monitoring effectiveness is regularly evaluated, and the reward function parameters are dynamically adjusted based on the evaluation results.

[0140] This adaptive threshold optimization method, based on reinforcement learning, adapts to dynamic changes in equipment operating conditions through continuous learning and dynamic adjustment, improving monitoring accuracy and reliability. Furthermore, multi-objective optimization ensures a balanced monitoring solution between practicality and cost-effectiveness, providing effective decision support for preventive maintenance of equipment.

[0141] Figure 3 This is a comparison chart of early warning accuracy indicators of different methods in the embodiment of the present invention. Figure 3 The figure shows a comparison of five different equipment condition monitoring methods across five key performance indicators. The data in the table clearly demonstrates that this technical solution significantly outperforms existing technologies across all indicators. In terms of early warning accuracy, this technical solution achieved 94.2%, 14.7 percentage points higher than the machine learning method and 51.9 percentage points higher than the fixed threshold method. In terms of false alarm rate, this technical solution was only 4.2%, a 60.0% reduction compared to the machine learning method and 89.0% lower than the fixed threshold method. In terms of missed alarm rate, this technical solution was only 3.7%, a 76.6% reduction compared to the machine learning method and 88.6% lower than the fixed threshold method. Particularly noteworthy is the early warning lead time indicator, where this technical solution can provide a 46-hour lead time, 1.92 times that of the machine learning method (24 hours) and 5.75 times that of the fixed threshold method (8 hours). This has significant practical value for maintenance decision-making.

[0142] In terms of the operating condition adaptability index, this technical solution achieved 92.3%, far exceeding other methods, demonstrating its excellent adaptability in complex and changing industrial environments. This series of experimental results fully demonstrates that the adaptive threshold optimization method based on reinforcement learning has significant overall performance advantages over traditional technologies, especially in reducing false alarms and missed alarms and increasing early warning lead time, providing more reliable technical support for industrial equipment condition monitoring and preventive maintenance.

[0143] Existing equipment performance monitoring methods often use fixed thresholds or simple statistical models for judgment. These methods are unable to adapt to the performance fluctuations of equipment under different operating conditions and fail to fully consider the impact of false positives and false negatives on the monitoring system, resulting in insufficient reliability of monitoring results. Furthermore, traditional methods fail to consider comprehensive factors such as equipment location, process requirements, and maintenance costs, making it difficult to achieve cost-effective optimization of monitoring strategies. This application improves the monitoring system's accuracy in identifying equipment performance changes by constructing a reward function based on false positives and false negatives, combined with a reinforcement learning framework to achieve adaptive optimization of performance benchmarks. A multi-objective optimization approach comprehensively considers equipment location, process requirements, and maintenance costs, enabling dynamic adjustment of tolerance ranges, ensuring monitoring reliability while reducing maintenance costs. Through a continuous optimization training mechanism, the monitoring system is able to continuously learn and adapt to dynamic changes in equipment status. This solution significantly improves the accuracy and reliability of equipment performance monitoring, reduces the occurrence of false positives and false negatives, optimizes preventive maintenance decisions, and improves the economic benefits of equipment management. Furthermore, the adaptive capabilities of the monitoring system are enhanced, enabling it to better adapt to monitoring needs under different operating conditions.

[0144] In an optional embodiment,

[0145] The fault characteristics are input into the pre-trained fault identification model. The fault type probability distribution is output based on the combination pattern of the fault characteristics. The fault type with the highest probability is selected as the fault judgment result of the component to be monitored. The results include:

[0146] Inputting the fault features into a pre-trained fault recognition model, decomposing the fault features into common features and individual features through a feature decoupling network, and applying orthogonal constraints to the common features and individual features to obtain decoupled features;

[0147] The decoupled features are constructed as an attribute graph, the feature dimensions are set as graph nodes, and the feature association strength is set as graph edges. Local and global dependencies are extracted through graph convolution and multi-head attention mechanisms to obtain a combination pattern of fault features.

[0148] The combination pattern of fault features and decoupling features are input into the variational inference module, and a state transfer matrix is ​​established based on the feature correlation strength and orthogonal constraints to obtain the initial fault type probability;

[0149] The initial fault type probability is subjected to Bayesian modeling, and probability calibration is performed based on the local dependency and global dependency in the combination pattern of the fault characteristics to generate a calibrated fault type probability distribution. The fault type with the highest probability is selected from the calibrated fault type probability distribution as the fault judgment result of the component to be monitored.

[0150] For example, in the process of diagnosing equipment faults on automated production lines, the first step is to identify and analyze fault characteristics. Fault characteristics include multi-source sensor data such as vibration signals, temperature, pressure, and current, as well as equipment operating and process parameters. These characteristics are preprocessed and then fed into a pretrained fault recognition model. This model utilizes a deep learning architecture, comprising core components such as a feature decoupling network, a graph convolutional network, and a variational inference module.

[0151] The feature decoupling network uses an encoder-decoder structure to map input fault features into a latent space through a multi-layer neural network. Within the latent space, the features are decomposed into two subspaces: common features and individual features. Common features reflect common characteristic patterns across different fault types, such as the basic characteristics of equipment vibration; individual features characterize the unique manifestations of each fault type, such as the impact signature specific to bearing faults. To ensure the effectiveness of the feature decomposition, an orthogonality constraint is imposed between the two subspaces to ensure their independence. This orthogonality constraint is achieved by minimizing the inner product of the eigenvectors of the two subspaces.

[0152] The decoupled features are constructed as an attribute graph. Nodes in the graph represent features of different dimensions, such as vibration amplitude and frequency. Edges between nodes indicate the strength of the association between features, calculated through feature correlation analysis. Graph convolution is applied to the attribute graph, aggregating information from adjacent nodes through a message passing mechanism to capture local dependencies between features. A multi-head attention mechanism is also introduced to calculate attention weights between different feature nodes and extract global dependencies between features. This dual dependency extraction yields a more complete combination of fault feature patterns.

[0153] The variational inference module receives the combined pattern of fault features and decoupled features as input. It constructs a state transition matrix based on the strength of feature correlations, with the matrix elements representing the transition probabilities between feature states. It also considers the orthogonality constraints of the decoupled features to ensure the rationality of state transitions. Variational inference calculates the initial probability distribution for each fault type.

[0154] Bayesian modeling is performed on the initial fault type probabilities, incorporating prior knowledge. This prior knowledge includes the distribution of fault types in historical fault data and the patterns of fault occurrence under different operating conditions. The initial probabilities are calibrated by combining local and global dependencies extracted from the feature combination pattern. Local dependencies are used to adjust the fault probability distribution between related features, while global dependencies consider the synergistic effects of the features as a whole. Ultimately, a calibrated fault type probability distribution is obtained, and the type with the highest probability is selected as the fault determination result.

[0155] Taking bearing fault diagnosis as an example, we first collect multi-source data, such as bearing vibration signals and temperature, as fault features. Using a feature decoupling network, these features are decomposed into common features reflecting the basic vibration characteristics of the bearing and individual features representing specific fault modes. In an attribute graph, frequency-domain features and time-domain features are set as nodes, and the correlation coefficients between features are set as edge weights.

[0156] Graph convolution operations extract local correlation patterns between vibration frequency features, such as the mutual influence of energy in different frequency bands. A multi-head attention mechanism captures long-range dependencies between vibration features and temperature characteristics. These dependencies together form a combined pattern of fault features, reflecting the overall characteristics of fault development.

[0157] The variational inference module calculates initial fault probabilities based on feature combination patterns, such as the probability distribution of inner race faults and outer race faults. It incorporates historical fault statistics through Bayesian modeling and calibrates probabilities based on feature dependencies. The module ultimately outputs a calibrated probability distribution of fault types, accurately determining the specific bearing fault type.

[0158] In this embodiment, by introducing a feature decoupling network, the fault features are decomposed into common features and individual features, which improves the accuracy and discrimination of feature expression and realizes the effective identification of fault features. The attribute graph structure constructed based on graph convolution and multi-head attention mechanism can simultaneously capture the local dependency and global dependency between features, and enhance the expression ability of the fault feature combination pattern. A probabilistic inference method combining variational inference and Bayesian modeling is adopted, and probability calibration is performed in combination with feature correlation strength and orthogonal constraints, which improves the accuracy and reliability of fault type determination. It can accurately identify various types of faults of equipment components, reduce the false positive rate and missed judgment rate, and provide reliable decision support for equipment preventive maintenance. At the same time, it has strong generalization ability, can adapt to the fault diagnosis needs under different working conditions, and improve the practicality and reliability of the fault diagnosis system.

[0159] In an optional embodiment,

[0160] Based on the physical connection relationship between the component to be monitored and its adjacent components, the performance degradation rate of the adjacent components is obtained. Based on the temporal change sequence of the performance degradation rate, a fault propagation link is established to determine the fault source component, including:

[0161] Obtain the mechanical coupling relationship, energy transfer relationship, and physical topological relationship between the component to be monitored and its adjacent components, and establish a component connection relationship matrix that characterizes the spatial connection strength between components;

[0162] Acquiring performance parameters of the component to be monitored and adjacent components within a continuous time window, calculating time derivatives of the performance parameters to obtain a performance degradation rate, and performing a Hilbert transform on the performance degradation rate to obtain an instantaneous phase;

[0163] According to the spatial connection strength and instantaneous phase variation trend in the component connection relationship matrix, the phase synchronization strength between adjacent components is calculated, and a network order parameter reflecting the dynamic coupling relationship between components is constructed;

[0164] monitoring the dynamic changes of the network order parameters, obtaining the mutation time and mutation amplitude of the network order parameters, and constructing a directed fault propagation graph based on the mutation time, mutation amplitude, and phase synchronization strength, wherein adjacent components are used as nodes, the time difference of the mutation time is used as the edge weight, and the phase synchronization strength is used as the edge direction;

[0165] Calculating the downstream impact range of each component based on the edge weights and edge directions in the directed fault propagation graph, determining the cascading failure scale of each component based on the downstream impact range, and combining the cascading failure scale with the mutation moment to obtain a failure impact factor;

[0166] According to the dynamic evolution law of the directed fault propagation graph, the mutation time characteristics and cascading failure scale in the failure impact factors are weighted, and the component with the largest failure impact factor and the earliest mutation time after configuration is determined as the fault source component.

[0167] When analyzing fault propagation in automated production lines, the first step is to determine the connectivity relationships between components. Mechanical coupling relationships describe the physical connections between components, such as the fit between bearings and shafts, or the meshing of gears. Energy transfer relationships represent the energy flow paths between components, including the transfer of mechanical and thermal energy. Physical topological relationships reflect the relative spatial relationships between components. Based on these relationships, a component connectivity matrix is ​​constructed. Matrix elements represent the strength of the spatial connections between components, which is calculated using physical parameters such as contact area and connection stiffness.

[0168] Within a set continuous time window, performance parameters of the component to be monitored and its adjacent components are collected. These parameters include physical quantities such as vibration, temperature, and pressure. The performance degradation rate is calculated by calculating the rate of change of these parameters over time, reflecting the rate of component degradation. The performance degradation rate is then subjected to a Hilbert transform, a signal processing method that extracts the instantaneous phase information of a signal. This instantaneous phase represents the phase angle of the signal at each instant and reflects the dynamic characteristics of performance degradation.

[0169] The phase synchronization strength between adjacent components is calculated based on the spatial connection strength and instantaneous phase variation in the component connection matrix. Phase synchronization strength indicates the degree of coordination between the performance degradation processes of two components; higher synchronization strength indicates greater correlation between the performance degradations of the two components. A network order parameter is constructed based on the phase synchronization strength. The network order parameter describes the overall dynamic characteristics of a complex network and is used to characterize the coevolutionary characteristics of a group of components.

[0170] Dynamically monitor network order parameters to identify their mutation moments and magnitudes. Mutation moments refer to the point in time when a network order parameter undergoes a significant change, while magnitudes indicate the degree of change. This information is combined with phase synchronization strength to construct a directed fault propagation graph. In this graph, components serve as nodes, and the fault propagation relationships between components are represented by directed edges. Edge weights are determined by the order of mutation moments, with components that mutate first pointing to components that mutate later. Edge direction is determined by phase synchronization strength, with the direction with greater synchronization strength being considered the primary direction of fault propagation.

[0171] The downstream impact of each component is analyzed based on the directed fault propagation graph. The downstream impact refers to the number of other components that may be affected by a component failure. By analyzing the connectivity of directed edges and calculating the number of nodes reachable from each component, the potential cascading failure scale of that component is determined. The cascading failure scale is combined with information about the moment of the mutation to construct a failure impact factor. The failure impact factor comprehensively considers the timing characteristics and impact scale of the component failure.

[0172] The dynamic evolution patterns of directed fault propagation graphs are analyzed, including the changing characteristics of network topology and the evolutionary patterns of fault propagation paths. Based on these patterns, failure impact factors are weighted, with weight coefficients assigned based on the characteristics of the mutation moment and the importance of the cascading failure scale. Ultimately, the component with the largest configured failure impact factor and the earliest mutation moment is identified as the fault source component.

[0173] For example, a production line drive system consists of key components such as the motor, reducer, and spindle. First, a connection matrix is ​​established for these components. For example, the motor and reducer are connected via a coupling, and the strength of the connection is determined by the stiffness of the coupling. Within a one-hour window, the vibration signals of each component are monitored, and the rate of change of the vibration amplitude is calculated to determine the performance degradation rate.

[0174] Using a Hilbert transform to extract the instantaneous phase of each component's vibration signal, analysis revealed strong synchronization between the phase changes of the reducer and the main shaft, while the motor and reducer exhibited weaker phase synchronization. The constructed network order parameters revealed that at a certain moment, the reducer experienced a sudden change first, followed by anomalies in the main shaft and motor.

[0175] In the directed fault propagation graph, the edge weight from the reducer to the spindle is large, indicating that the fault is likely to propagate from the reducer to the spindle. Analysis shows that the reducer has the largest downstream impact, and its failure could cause a chain reaction in the spindle and motor. Taking into account the timing of the sudden change and the scale of the cascading failure, the reducer was ultimately identified as the fault source component.

[0176] Existing fault tracing methods mainly rely on the fault feature analysis of a single component or simple correlation analysis, failing to fully consider the complex physical connection relationship between components and the dynamic characteristics of fault propagation, making it difficult to accurately identify the fault source component. At the same time, traditional methods lack quantitative analysis of the cascade effect during fault propagation and cannot effectively evaluate the impact range and propagation path of the fault. This application comprehensively describes the mechanical coupling, energy transfer and physical topological relationship between components by constructing a component connection relationship matrix, providing a reliable physical basis for fault propagation analysis. The introduction of Hilbert transform and phase synchronization analysis methods achieves an accurate characterization of the dynamic characteristics of fault propagation. Based on the analysis framework of directed fault propagation graphs and failure influencing factors, a quantitative evaluation of fault propagation paths and cascade effects is achieved, which significantly improves the accuracy of fault source identification, can effectively predict the propagation range and impact of faults, and provides a reliable basis for equipment maintenance decisions. At the same time, it has strong adaptability and can handle fault tracing problems of different types of equipment systems, thereby enhancing the practical value of the fault diagnosis system.

[0177] A second aspect of the present invention provides a multi-parameter real-time monitoring and fault diagnosis and early warning system for an automated production line, the system comprising:

[0178] The first unit is configured to obtain real-time operation data collected by multi-source sensors in an automated production line, calculate data quality characteristic values, generate data credibility, and, when the data credibility is less than a first preset threshold, perform data correction based on historical data at adjacent moments to obtain corrected real-time operation data;

[0179] a second unit, configured to group the corrected real-time operating data according to equipment components, generate operating parameter sequences corresponding to the equipment components, calculate performance parameters of the equipment components based on the operating parameter sequences, calculate a performance degradation rate based on the performance parameters, and determine the corresponding standby component as a component to be monitored when the performance degradation rate is greater than a second preset threshold;

[0180] The third unit is configured to perform time-frequency domain feature analysis on the operating parameter sequence of the component to be monitored, extract time-domain statistical features and frequency-domain energy features as fault features, input the fault features into a pre-trained fault recognition model, output a probability distribution of fault types based on a combination pattern of the fault features, and select the fault type with the highest probability as the fault determination result of the component to be monitored;

[0181] The fourth unit is used to obtain the performance degradation rate of adjacent components based on the physical connection relationship between the component to be monitored and the adjacent components, establish a fault propagation link based on the time series change sequence of the performance degradation rate, determine the fault source component, generate a fault diagnosis report containing the location information of the fault source component and the fault judgment result, and send it to the production management system.

[0182] According to a third aspect of the embodiments of the present invention,

[0183] An electronic device is provided, comprising:

[0184] processor;

[0185] a memory for storing processor-executable instructions;

[0186] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0187] According to a fourth aspect of the embodiments of the present invention,

[0188] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0189] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time monitoring and fault diagnosis and early warning of multiple parameters in an automated production line, characterized in that: include: Acquire real-time operating data collected by multi-source sensors in an automated production line, calculate data quality characteristic values, generate data credibility, and when the data credibility is less than a first preset threshold, perform data correction based on historical data at adjacent moments to obtain corrected real-time operating data; Grouping the corrected real-time operating data according to device components to generate operating parameter sequences corresponding to the device components, calculating performance parameters of the device components based on the operating parameter sequences, calculating a performance degradation rate based on the performance parameters, and determining the corresponding standby component as a component to be monitored when the performance degradation rate is greater than a second preset threshold; Performing time-frequency domain feature analysis on the operating parameter sequence of the component to be monitored, extracting time-domain statistical features and frequency-domain energy features as fault features, inputting the fault features into a pre-trained fault recognition model, outputting a probability distribution of fault types based on a combination pattern of the fault features, and selecting the fault type with the highest probability as the fault determination result of the component to be monitored; Based on the physical connection relationship between the component to be monitored and its adjacent components, the performance degradation rate of the adjacent components is obtained. Based on the temporal change sequence of the performance degradation rate, a fault propagation link is established to determine the fault source component. A fault diagnosis report containing the location information of the fault source component and the fault determination result is generated and sent to the production management system. Acquire real-time operation data collected by multi-source sensors in an automated production line, calculate data quality characteristic values, and generate data credibility. When the data credibility is less than a first preset threshold, perform data correction based on historical data at adjacent moments. The corrected real-time operation data includes: Acquire real-time operation data collected by multi-source sensors in the automated production line and segment it according to time windows to obtain segmented data sequences; Calculate the signal fluctuation, value variation, and sampling integrity of the segmented data sequence respectively and combine them to obtain data quality characteristic values; calculate the information entropy of the data quality characteristic values ​​based on the probability distribution characteristics of the data in the segmented data sequence; generate a dynamic weight coefficient based on the information entropy; and perform a weighted combination of the dynamic weight coefficient and the corresponding data quality characteristic value to obtain data credibility; When the data credibility is less than a first preset threshold, determining the data to be corrected, and extracting the historical data sequence of adjacent moments of N time windows before and after the time window corresponding to the data to be corrected; Clustering the historical data sequences at adjacent moments according to the similarity of data change trends to obtain multiple subsequence groups, calculating the time correlation between each subsequence in each subsequence group and the data to be corrected, and constructing a recursive optimization model with differentiated weights based on the time correlation, where the weights are positively correlated with the time correlation; The data to be corrected is iteratively corrected using a recursive optimization model. In each iteration, the fitting error between the correction result and the subsequence whose correlation is greater than a preset correlation threshold is calculated. When the fitting error is less than a dynamic threshold, the correction result is output as the corrected real-time running data. Based on the physical connection relationship between the component to be monitored and its adjacent components, the performance degradation rate of the adjacent components is obtained. Based on the temporal change sequence of the performance degradation rate, a fault propagation link is established to determine the fault source component, including: Obtain the mechanical coupling relationship, energy transfer relationship, and physical topological relationship between the component to be monitored and its adjacent components, and establish a component connection relationship matrix that characterizes the spatial connection strength between components; Acquiring performance parameters of the component to be monitored and adjacent components within a continuous time window, calculating time derivatives of the performance parameters to obtain a performance degradation rate, and performing a Hilbert transform on the performance degradation rate to obtain an instantaneous phase; According to the spatial connection strength and instantaneous phase variation trend in the component connection relationship matrix, the phase synchronization strength between adjacent components is calculated, and a network order parameter reflecting the dynamic coupling relationship between components is constructed; monitoring the dynamic changes of the network order parameters, obtaining the mutation time and mutation amplitude of the network order parameters, and constructing a directed fault propagation graph based on the mutation time, mutation amplitude, and phase synchronization strength, wherein adjacent components are used as nodes, the time difference of the mutation time is used as the edge weight, and the phase synchronization strength is used as the edge direction; Calculating the downstream impact range of each component based on the edge weights and edge directions in the directed fault propagation graph, determining the cascading failure scale of each component based on the downstream impact range, and combining the cascading failure scale with the mutation moment to obtain a failure impact factor; According to the dynamic evolution law of the directed fault propagation graph, the mutation time characteristics and cascading failure scale in the failure impact factors are weighted, and the component with the largest failure impact factor and the earliest mutation time after configuration is determined as the fault source component.

2. The method according to claim 1, characterized in that Calculating the information entropy of the data quality characteristic value based on the probability distribution characteristics of the data in the segmented data sequence, and generating a dynamic weight coefficient according to the information entropy includes: Constructing a probability density estimation function, calculating the standard deviation and the interquartile range based on the segmented data sequence, and dynamically adjusting the bandwidth coefficient of the probability density estimation function according to the ratio of the standard deviation to the interquartile range to obtain a continuous probability density curve for the segmented data sequence; Performing piecewise integration on the continuous probability density curve to obtain a cumulative probability value for each segment of data, dividing the segmented data sequence into a first distribution area and a second distribution area according to the cumulative probability value, and calculating a cumulative probability ratio of the first distribution area to the second distribution area as a data distribution characteristic ratio; The overlapping sliding window step size is set using the data distribution characteristic ratio, and a first local information entropy is obtained by sliding calculation within the first distribution area; the non-overlapping sliding window width is set using the inverse of the data distribution characteristic ratio, and a second local information entropy is obtained by calculation within the second distribution area; Taking the cumulative probability values ​​of the first distribution area and the second distribution area as weight coefficients, weighting the first local information entropy and the second local information entropy to obtain a regional combined entropy; The global information entropy is calculated for the continuous probability density curve, the regional combined entropy and the global information entropy are normalized to obtain the final information entropy of the data quality characteristic value, and a dynamic weight coefficient is generated according to the final information entropy, where the dynamic weight coefficient is inversely proportional to the final information entropy.

3. The method according to claim 1, characterized in that Calculating performance parameters of the device components based on the operating parameter sequence, calculating a performance degradation rate based on the performance parameters, and determining the corresponding standby component as a component to be monitored when the performance degradation rate is greater than a second preset threshold includes: Performing time-frequency analysis on the operating parameter sequence to obtain energy distribution characteristics, selecting a characteristic parameter combination based on the energy distribution characteristics, and constructing the characteristic parameter combination into a comprehensive performance status indicator; Analyze the temporal variation pattern of the comprehensive performance status indicator, establish a performance status transfer matrix, calculate the change rate of the performance parameter based on the performance status transfer matrix, and correct the change rate in combination with the performance status transfer matrix to obtain the performance degradation rate; Establish a performance benchmark value for the normal operation of the component, determine the performance fluctuation tolerance range based on the performance benchmark value combined with historical operation data, start the cumulative time count when it is detected that the performance degradation rate exceeds the performance fluctuation tolerance range, and when the cumulative time reaches the preset period, determine the corresponding equipment component as the component to be monitored.

4. The method according to claim 3, characterized in that Establishing a performance baseline for a component during normal operation and determining the performance fluctuation tolerance based on the performance baseline combined with historical operating data include: Obtaining historical operating data of the device components, identifying normal operating periods and abnormal operating periods from the historical operating data, performing comparative analysis on the normal operating periods and abnormal operating periods, extracting false positives and false negatives, and constructing a reward function; Inputting the reward function and the performance data of the normal operation period into a reinforcement learning framework, obtaining a performance benchmark value through reinforcement learning training, and generating an initial tolerance adjustment strategy based on the performance benchmark value; Collect real-time operating status data of components, match and analyze the real-time operating status data with the initial tolerance adjustment strategy, calculate the status warning probability under different tolerance settings, and generate risk prediction results within the tolerance range; Obtaining the component's location information in the production process, process parameter requirements for the process, and historical maintenance cost data, and performing a multi-objective optimization calculation on the location information, process parameter requirements, and historical maintenance cost data with the risk prediction results to generate a tolerance optimization objective function; The initial tolerance adjustment strategy is optimized based on the tolerance optimization objective function to obtain a performance fluctuation tolerance range, the performance fluctuation tolerance range is combined with a performance benchmark value to form a monitoring threshold, and the state data corresponding to the monitoring threshold is fed back to the reinforcement learning framework for continuous optimization training.

5. The method according to claim 1, characterized in that The fault characteristics are input into the pre-trained fault identification model. The fault type probability distribution is output based on the combination pattern of the fault characteristics. The fault type with the highest probability is selected as the fault judgment result of the component to be monitored. The results include: Inputting the fault features into a pre-trained fault recognition model, decomposing the fault features into common features and individual features through a feature decoupling network, and applying orthogonal constraints to the common features and individual features to obtain decoupled features; The decoupled features are constructed as an attribute graph, the feature dimensions are set as graph nodes, and the feature association strength is set as graph edges. Local and global dependencies are extracted through graph convolution and multi-head attention mechanisms to obtain a combination pattern of fault features. The combination pattern of fault features and decoupling features are input into the variational inference module, and a state transfer matrix is ​​established based on the feature correlation strength and orthogonal constraints to obtain the initial fault type probability; The initial fault type probability is subjected to Bayesian modeling, and probability calibration is performed based on the local dependency and global dependency in the combination pattern of the fault characteristics to generate a calibrated fault type probability distribution. The fault type with the highest probability is selected from the calibrated fault type probability distribution as the fault judgment result of the component to be monitored.

6. A multi-parameter real-time monitoring and fault diagnosis and early warning system for an automated production line, used to implement the method according to any one of claims 1 to 5, characterized in that: include: The first unit is configured to obtain real-time operation data collected by multi-source sensors in an automated production line, calculate data quality characteristic values, generate data credibility, and, when the data credibility is less than a first preset threshold, perform data correction based on historical data at adjacent moments to obtain corrected real-time operation data; a second unit, configured to group the corrected real-time operating data according to equipment components, generate operating parameter sequences corresponding to the equipment components, calculate performance parameters of the equipment components based on the operating parameter sequences, calculate a performance degradation rate based on the performance parameters, and determine the corresponding standby component as a component to be monitored when the performance degradation rate is greater than a second preset threshold; The third unit is configured to perform time-frequency domain feature analysis on the operating parameter sequence of the component to be monitored, extract time-domain statistical features and frequency-domain energy features as fault features, input the fault features into a pre-trained fault recognition model, output a probability distribution of fault types based on a combination pattern of the fault features, and select the fault type with the highest probability as the fault determination result of the component to be monitored; The fourth unit is used to obtain the performance degradation rate of adjacent components based on the physical connection relationship between the component to be monitored and the adjacent components, establish a fault propagation link based on the time series change sequence of the performance degradation rate, determine the fault source component, generate a fault diagnosis report containing the location information of the fault source component and the fault judgment result, and send it to the production management system.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images