A real-time anomaly monitoring method and system for an industrial production line

Abstract

The present application relates to the field of data processing, in particular to a real-time abnormality monitoring method and system for industrial production line production, the method comprising: collecting equipment state data to form a multi-dimensional time series and dividing it into period data blocks aligned with the production rhythm; determining a reference period length and a period jitter standard deviation based on offline health data; establishing an analysis matrix for the current period data block with rhythm-adaptive adjustment and singular value decomposition; calculating the spectral smoothness of each independent component, identifying components below the critical smoothness threshold as a trend component set; calculating the parameter drift degree and rhythm instability weight, fusing the two to obtain a non-steady-state weighted drift strength, and performing abnormality determination. The present application solves the problems of phase shift caused by rhythm fluctuation and difficulty in extracting weak fault features in traditional methods through rhythm-adaptive alignment, spectral smoothness screening and multi-index fusion evaluation, significantly improving the accuracy and robustness of monitoring.

Description

Technical Field

[0001] This invention relates to the field of data processing. More specifically, this invention relates to a method and system for real-time anomaly monitoring in industrial production lines. Background Technology

[0002] In discrete manufacturing fields (such as die casting and injection molding), monitoring the operating status of production equipment is crucial to ensuring product quality and production continuity. Current industrial monitoring methods mostly adopt fixed time window analysis or judgment strategies based on signal energy thresholds.

[0003] However, these traditional methods have fundamental flaws when dealing with scenarios where production cycles fluctuate significantly. The operating rhythm of equipment naturally changes with factors such as the state of the hydraulic system, temperature variations, and mechanical wear, making it difficult for a fixed window to synchronize with these changes. This leads to phase shifts in the signal processing, misinterpreting normal process fluctuations as abnormalities, or drowning out genuine early fault signals due to timing misalignment interference. Consequently, these methods fail to meet the robustness requirements of monitoring systems in complex industrial settings. Summary of the Invention

[0004] To address the technical problem that traditional methods struggle to meet the robustness requirements of monitoring systems in scenarios with significant fluctuations in production cycle time, this invention provides solutions in the following aspects.

[0005] In the first aspect, a method for real-time anomaly monitoring in industrial production lines includes:

[0006] Data is collected at a fixed frequency by a sensor network deployed at key locations on the equipment to form a multidimensional time series matrix, which is then divided into multiple periodic data blocks aligned with the production cycle.

[0007] Based on the baseline cycle data block collected offline under healthy conditions, the baseline cycle duration and cycle jitter standard deviation are determined; for the current cycle data block, an analysis matrix that adaptively adjusts with the actual production cycle is established through a cycle synchronization mechanism, and singular value decomposition is performed on the analysis matrix of the current cycle data block to obtain multiple independent components.

[0008] Calculate the spectral smoothness of each independent component of the analysis matrix of the current periodic data block, the spectral smoothness being determined based on the product of the differential energy value of the independent component and a pre-calculated periodic compensation factor;

[0009] Independent components with spectral smoothness below a preset critical smoothness threshold are marked as trend components, forming a set of trend components;

[0010] The pure trend signal is reconstructed based on the set of trend components, and the parameter drift of the pure trend signal is calculated. The beat instability weight of the data block in the current period is calculated based on the difference between the actual duration of the current period and the duration of the base period and the standard deviation of the period jitter.

[0011] The parameter drift degree is multiplied by the cycle instability weight to obtain the unsteady weighted drift intensity of the equipment in the current cycle; based on the comparison result of the unsteady weighted drift intensity and the preset threshold, the abnormal operation status of the equipment in the current cycle is determined.

[0012] Optionally, in offline acquisition, multiple consecutive confirmed healthy periodic data blocks are acquired as the reference periodic data blocks, and the average period duration of all reference periodic data blocks is taken as the reference period duration, and the standard deviation of the period duration of all reference periodic data blocks is taken as the standard deviation of period jitter.

[0013] Optionally, the beat synchronization mechanism includes:

[0014] Using multiple offline-acquired reference period data blocks, the information entropy of each reference period data block is calculated through trajectory matrix construction and singular value decomposition, and the arithmetic mean of the information entropy is used as a signal characteristic index.

[0015] The ratio of the signal characteristic index to the theoretical maximum information entropy is determined as the time proportionality coefficient of the device;

[0016] The analysis window length of the current period data block is obtained by multiplying the product of the actual duration of the current period data block and the sampling frequency by the time scaling factor.

[0017] Based on the analysis window length, a two-dimensional analysis matrix is ​​constructed for each sensor channel.

[0018] Optionally, the calculation of the spectral smoothness includes:

[0019] The differential energy value of the independent component is obtained by summing the squared differences between adjacent elements in the independent component.

[0020] The cycle compensation factor is obtained by calculating the ratio of the baseline cycle length to the actual current cycle length.

[0021] The spectral smoothness is obtained by multiplying the differential energy value by the period compensation factor and then taking the square root.

[0022] Optionally, the setting of the critical smoothing threshold includes:

[0023] Singular value decomposition is performed on the baseline periodic data blocks under multiple healthy states, and the spectral smoothness of each independent component is calculated.

[0024] Extract several independent components with the lowest spectral smoothness from each reference period data block as candidate components;

[0025] Calculate the mean and standard deviation of the spectral smoothness of the candidate components;

[0026] The critical smoothing threshold is the mean of the spectral smoothness minus a certain number of times the standard deviation of the spectral smoothness.

[0027] Optionally, the calculation of the clock instability weight includes:

[0028] Calculate the absolute value of the difference between the actual duration of the current cycle and the duration of the base cycle;

[0029] The absolute value of the difference is divided by the standard deviation of the period jitter, and an exponential operation is performed with the natural constant as the base to obtain the beat instability weight.

[0030] Optionally, the calculation of the parameter drift degree includes:

[0031] Obtain the difference between the value of the pure trend signal at each moment in the current period and the health benchmark value;

[0032] The square of the difference between the value of the pure trend signal at each moment in the current period and the healthy baseline value is integrated over the actual duration of the current period, divided by the actual duration of the current period, and then the square root is taken to obtain the degree of drift of the parameter.

[0033] The health benchmark value is the global average of the pure trend signals corresponding to the benchmark period data blocks under multiple health states.

[0034] Optionally, the anomaly determination includes:

[0035] A health baseline distribution is constructed based on the non-steady-state weighted drift intensity corresponding to the baseline periodic data blocks under multiple health states;

[0036] Calculate the mean and standard deviation of the health baseline distribution and set monitoring thresholds for multiple levels;

[0037] The current period's unsteady weighted drift intensity is compared with monitoring thresholds at multiple levels to generate corresponding anomaly detection signals.

[0038] In a second aspect, a real-time anomaly monitoring system for industrial production lines includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the real-time anomaly monitoring method for industrial production lines described in any one of the claims is implemented.

[0039] The present invention has the following beneficial effects:

[0040] 1. This invention solves the problem of phase shift caused by clock fluctuations and difficulty in extracting weak fault features under strong periodic interference in traditional methods by constructing a technical architecture of "clockwise adaptive alignment - spectral smoothness screening - multi-index fusion evaluation". First, an analysis matrix that adaptively adjusts with the actual production clockwise is established through a dynamic clockwise synchronization mechanism, keeping the analysis window synchronized with the physical actions of the equipment to eliminate phase misalignment. Second, spectral smoothness, independent of signal energy, is introduced as a screening criterion. Utilizing the inherent characteristic of slow trend signal changes, weak trend components reflecting early equipment degradation are accurately identified from extremely low-energy components. Finally, the degree of parameter drift and the clockwise instability weight are fused into a non-steady-state weighted drift intensity, automatically amplifying the sensitivity to parameter drift when the clockwise is disordered, achieving a comprehensive quantitative assessment of the equipment's health status. This technical architecture can extract true pre-fault signals from noisy industrial data, providing a reliable foundation for predictive maintenance. Its monitoring sensitivity and robustness are significantly superior to existing methods.

[0041] 2. This invention achieves adaptive adjustment of the analysis window through a device time ratio coefficient based on information entropy, enabling the monitoring method to automatically adapt to the process complexity of different equipment. For complex equipment such as injection molding machines, the signal information entropy is large and the time ratio coefficient is close to 1, allowing the analysis window to cover the complete multi-stage process. For simple equipment such as die casting machines, the signal information entropy is small and the time ratio coefficient is much less than 1, allowing the analysis window to focus on key change stages. Simultaneously, this mechanism can respond in real-time to changes in production cycle time: when the cycle time slows down, the analysis window is automatically extended to capture slow degradation trends; when the cycle time abnormally speeds up, the analysis window is automatically shortened to enhance sensitivity to transient changes, making it suitable for different types of production equipment and process scenarios. Attached Figure Description

[0042] Figure 1 This is a flowchart of steps S1-S3 in a real-time anomaly monitoring method for industrial production lines according to an embodiment of the present invention.

[0043] Figure 2 This is a schematic diagram illustrating the method for feature extraction from multi-dimensional time-series data aligned with the production cycle in a real-time anomaly monitoring method for industrial production lines according to an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of the structure of a real-time anomaly monitoring system for industrial production lines according to an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0046] Reference Figure 1 A method for real-time anomaly monitoring in industrial production lines includes steps S1-S3, as detailed below:

[0047] S1: Collect equipment status data and preprocess it to construct multi-dimensional time-series data aligned with the production cycle.

[0048] In discrete manufacturing processes, changes in equipment status are initially manifested as numerical fluctuations in multidimensional sensor signals. However, raw industrial data often contains noise and outliers, and the time series lengths between different cycles vary due to cycle time fluctuations. If the system directly uses such data for analysis, signal phase misalignment and noise interference will severely distort the true representation of the equipment status. Therefore, the system first needs to establish a unified, clean data foundation that is strictly synchronized with the physical movements of the equipment.

[0049] First, the system performs multi-dimensional data synchronous acquisition, which means that through a sensor network deployed at key locations on the equipment (such as synchronously acquiring injection pressure, clamping force, hydraulic oil temperature, and mold temperature on a die-casting machine), data is collected at a fixed frequency to form a multi-dimensional time series matrix.

[0050] Meanwhile, the system uses a PLC (Programmable Logic Controller) to capture the rising edge of instruction signals defining the production cycle boundaries, such as "mold opening end," and uses this as a hard synchronization reference to precisely divide the aforementioned multi-dimensional time series matrix into multiple independent periodic data blocks. Thus, each sensor acquisition ultimately yields a corresponding periodic data block. This segmentation method ensures that each periodic data block encapsulates a complete and independent equipment operation unit.

[0051] Secondly, the system performs data cleaning and smoothing. During the offline initialization phase, when the device is in a confirmed healthy and stable state, the system continuously collects, for example, 100 baseline period data blocks. The average period duration of all baseline period data blocks is used as the baseline period duration, and the standard deviation of period jitter is obtained. The system uses the baseline period duration and the standard deviation of period jitter as the benchmark reference for all subsequent evaluations.

[0052] Furthermore, in the online real-time processing phase, the system uses the current periodic data block for each sensor channel... (in, This represents the mean. The standard deviation (represented by the standard deviation) is used to identify and eliminate outliers caused by transient sensor interference, and real-time correction is performed using the linear interpolation method between adjacent points to ensure data continuity. After correction, a filter is used for conventional data smoothing.

[0053] After the above processing, the system converts the periodic data blocks of each cycle into a clean, smooth, multi-dimensional time series matrix that is synchronized with the actual operating cycle of the equipment.

[0054] S2: Based on the constructed multidimensional time-series data aligned with the production cycle, feature extraction is performed through adaptive cycle synchronization and spectral smoothness analysis to obtain the set of trend components and non-steady-state weighted drift intensity that characterize the health status of the equipment.

[0055] After obtaining high-quality multidimensional time-series data aligned with the production cycle, the key to determining the sensitivity and accuracy of the monitoring system lies in how it extracts subtle trend signals that truly reflect equipment health degradation. Traditional methods often fail to extract features due to their inability to adapt to cycle fluctuations and strong periodic interference. In this embodiment of the invention, referring to... Figure 2 The specific sub-steps for feature extraction from multidimensional time-series data are as follows:

[0056] S20: Based on the periodic data blocks and their period duration obtained from the segmentation of multi-dimensional time-series data, a dynamic cycle synchronization mechanism is implemented to establish a unified analysis matrix for all sensor channels that adaptively adjusts with the actual production cycle.

[0057] First, the device-specific time ratio coefficient is determined by setting initial parameter configurations. After the initial system installation or significant process adjustments, multiple baseline cycle data blocks acquired offline in S1 are used, and each baseline cycle data block is processed as follows:

[0058] A trajectory matrix is ​​constructed using a conservative initial window length, which is half the product of the aforementioned reference period duration and the sampling frequency, rounded down to ensure coverage of half a period of signal during normal device operation. Subsequently, singular value decomposition is performed on the trajectory matrix to obtain a series of singular values. Each singular value reflects the energy magnitude of different feature components in the signal. The system normalizes each singular value by dividing it by the sum of all singular values. Based on these normalized singular values, the information entropy of the corresponding periodic data block signal is calculated. The larger the entropy value, the richer and more diverse the signal components within the periodic data block, and the more complex the process.

[0059] Furthermore, the arithmetic mean of the information entropy of all periodic data blocks is used as the signal characteristic index of the device to reflect the typical process complexity of the device in a healthy state.

[0060] Furthermore, since the maximum value of information entropy occurs when all singular values ​​are equal, the energy distribution of each component in the signal is most uniform at this time. Therefore, the natural logarithm of the number of columns in the above trajectory matrix is ​​taken as the theoretical maximum information entropy. Furthermore, the ratio of the signal characteristic index of the above device to the theoretical maximum information entropy is taken as the device-specific time scaling factor.

[0061] In actual industrial settings, the process signals of different equipment exhibit vastly different levels of complexity. For example, die-casting machines present a simple, pulse-type process, while injection molding machines present a complex, multi-stage process. Through the aforementioned operations, multiple healthy baseline cycle data blocks are collected. A trajectory matrix is ​​constructed for each cycle data block, and singular value decomposition is performed. The information entropy is calculated and averaged, thereby automatically quantifying the equipment's process characteristics. For equipment with complex signal changes, such as injection molding machines, the average information entropy is relatively large, and the time scale factor approaches 1. This means that the analysis matrix window needs a longer time range to capture multi-stage process characteristics. For equipment with simple signal changes, such as die-casting machines, the average information entropy is relatively small, and the time scale factor is much smaller than 1. This means that the analysis matrix window focuses on the key change stages.

[0062] Secondly, during the real-time operation phase, the actual duration of the current period's data block is multiplied by the sampling frequency, then multiplied by the aforementioned time scaling factor and rounded down to obtain the analysis window length for the current period's data blocks. A significant increase in the analysis window length may indicate a slower production cycle, suggesting potential issues such as insufficient hydraulic pressure, reduced cooling efficiency, or wear and tear on mechanical components. Conversely, a significant decrease in the analysis window length may indicate an abnormally fast production cycle, suggesting the equipment may be operating under overload conditions or that process parameters are improperly set, requiring attention to equipment safety and product quality stability.

[0063] All sensor channels use a uniform analysis window length, ensuring that different physical quantities such as pressure and temperature are analyzed on the same time scale, maintaining the physical correlation between the parameters.

[0064] Finally, based on a unified analysis window length, an analysis matrix is ​​independently constructed for each sensor channel. Specifically, for each sensor channel, the time-series data of that channel within the current period's data block is rearranged according to the previously calculated analysis window length to construct a two-dimensional analysis matrix. The analysis matrix is ​​constructed as follows: using the analysis window length as the number of rows, starting with the first data point of the current period's data block, data segments with a length equal to the analysis window length are sequentially extracted as the first column of the matrix; then, one data point is slid forward, and the next data segment with a length equal to the analysis window length is extracted as the second column of the matrix; this process continues until all data points of the entire current period are covered. The number of columns in the analysis matrix is ​​determined by subtracting the analysis window length from the total number of sampling points in the current period and then adding 1.

[0065] By constructing the data using the above method, one-dimensional time series data is transformed into a two-dimensional matrix, revealing the internal structural features and correlations of the signal. Since the analysis window length dynamically adjusts with the actual duration of the current production cycle, the number of rows and columns in the analysis matrix also automatically scales, ensuring that each column covers the critical stages of equipment operation. A larger analysis matrix indicates a slower production cycle, allowing the system to capture longer-term trends and facilitating the detection of slow equipment degradation. Conversely, a smaller analysis matrix indicates a faster production cycle, making the system more sensitive to instantaneous changes. In this case, it is necessary to combine other parameters for comprehensive judgment to avoid misinterpreting normal process fluctuations as equipment malfunctions.

[0066] In summary, all sensor channels share the same analysis window length, which means that the time series of different physical quantities such as pressure, temperature, and vibration are constructed into an analysis matrix under a unified time reference, thereby maintaining the inherent physical correlation between the parameters and providing a reliable data foundation for subsequent accurate extraction of equipment status characteristics and improvement of anomaly judgment accuracy.

[0067] S21: Based on the constructed analysis matrix and the corresponding period duration, the set of trend components characterizing the health status of the equipment is obtained using the spectral smoothness driving mechanism.

[0068] Early equipment failures, such as slight wear of valve cores or minor leaks in sealing rings, often result in trend changes whose energy is only one-thousandth or even less of the energy of normal periodic process operations. Traditional singularity spectrum analysis methods rely on component energy (singularity magnitude) for screening, which can easily misjudge high-energy periodic process operations as trends or drown out weak failure trends in periodic fluctuations. Therefore, it is necessary to introduce a "spectral smoothness" that is independent of signal energy, utilizing the inherent characteristic of trend signals changing slowly in the time domain to achieve accurate capture of weak features.

[0069] First, singular value decomposition is performed on the analysis matrix constructed in S20 above to obtain a series of left singular vectors. Each left singular vector corresponds to an independent component in the signal. For each independent component, its spectral smoothness is calculated. The spectral smoothness includes two parts: the first part is to calculate the difference energy value of the independent component, that is, starting from the first element of the left singular vector, the difference between two adjacent elements is calculated in turn, and the sum of the squares of each difference is obtained to obtain the difference energy value; the second part is to calculate the period compensation factor, that is, the ratio of the above reference period duration to the actual duration of the current period, which is used to eliminate the influence of changes in production cycle on the smoothness of the waveform.

[0070] Furthermore, the differential energy value of the above independent components is multiplied by the period compensation factor and then squared to obtain the spectral smoothness of the above independent components. When the spectral smoothness is close to zero, it indicates that the independent component is an extremely smooth, slowly varying signal, which usually corresponds to the long-term degradation trend of the equipment or zero-point drift. When the spectral smoothness is much greater than zero, it indicates that the independent component is a high-frequency oscillating signal, which usually corresponds to noise or periodic interference.

[0071] Secondly, during the initial parameter configuration phase, the system repeats the singular value decomposition and spectral smoothness calculation process for each reference periodic data block under each healthy state, extracting several independent components with the lowest spectral smoothness as candidate components. These candidate components are typically high-energy periodic process components, and their spectral smoothness represents the normal smoothness of the signal waveform under healthy conditions.

[0072] Furthermore, the mean and standard deviation of the spectral smoothness of these candidate components are calculated, and the mean minus 2 to 3 times the standard deviation is used as the critical smoothing threshold. This critical smoothing threshold represents the lower limit of signal smoothness in health devices.

[0073] During real-time operation, the system automatically identifies components in the current periodic data block whose spectral smoothness is below the critical smoothing threshold as trend components and adds them to the trend component set. When the spectral smoothness of an independent component is below the critical smoothing threshold, it indicates that the smoothness of that component has exceeded the normal range of a healthy device. This usually corresponds to a slow change in the physical state of the device, such as a continuous shift in the pressure reference value or a gradual drift of the zero temperature point. The system indexes and records these abnormally smooth independent components to accurately extract these subtle features reflecting early device degradation in subsequent steps, without them being masked by strong periodic process fluctuations.

[0074] S22: Based on the obtained trend component set and the baseline period duration and period jitter standard deviation, the non-steady-state weighted drift intensity is evaluated by integrating parameter drift and time-domain stability as dual criteria to comprehensively quantify the equipment health status and obtain the non-steady-state weighted drift intensity.

[0075] After identifying the set of trend components, it is necessary to quantify their comprehensive impact on equipment health. In complex industrial environments, simple numerical drift may be caused by non-steady-state conditions such as environmental changes and preheating. Failure to distinguish between these factors can easily lead to false alarms. More importantly, the instability of the production cycle itself, such as random fluctuations in cycle length, is often a direct manifestation of equipment performance degradation. Therefore, it is necessary to integrate the two dimensions of parameter drift and cycle jitter to enable the system to construct a comprehensive evaluation index that is adaptive to changes in operating conditions.

[0076] First, using the set of trend components obtained in S21, the pure trend signal of the current periodic data block is reconstructed by methods such as diagonal averaging. This pure trend signal has eliminated periodic process fluctuations and high-frequency noise, and only retains the gradual trend of the physical state of the equipment.

[0077] Secondly, calculate the square of the difference between the value of the pure trend signal at each moment in the current period and the health baseline value of the device. Then, integrate it over the actual duration of the current period, divide it by the actual duration of the current period to obtain the average value, and finally take the square root to obtain the parameter drift degree, so as to quantify the overall level of deviation of the device's internal physical parameters from the healthy state. The larger the value, the more drastic the change in the physical state of the device.

[0078] Among them, the health baseline value is the pure trend signal of the corresponding period that is reconstructed from multiple healthy baseline period data blocks collected by S1 during the initial parameter configuration phase of the system. Then, the global average value of these pure trend signals at all times in the current period is calculated as the health baseline value of the device.

[0079] Then, calculate the absolute value of the difference between the actual duration of the current cycle and the duration of the base cycle, and divide it by the standard deviation of the cycle jitter (to prevent the denominator from being 0, a very small constant such as 0.001 can be added to the denominator and added to the standard deviation of the cycle jitter). Then, perform an exponential operation with the natural constant e as the base to obtain the beat instability weight.

[0080] Finally, the parameter drift degree is multiplied by the cycle instability weight to obtain the final unsteady-state weighted drift intensity. When the production cycle is stable, the absolute value of the difference between the actual duration of the current cycle and the baseline cycle duration is small, and the value of the cycle instability weight is close to 1. The system evaluation is mainly determined by the parameter drift degree. When abnormal fluctuations occur in the production cycle, the absolute value of the difference between the actual duration of the current cycle and the baseline cycle duration increases significantly. The cycle instability weight is rapidly amplified through an exponential function, and the system automatically increases its sensitivity to parameter drift. This means that if a slight parameter drift is detected while the cycle is in disarray, the system will treat it as a much more serious anomaly than if it occurs alone.

[0081] When the above-mentioned unsteady-state weighted drift intensity approaches zero, it indicates that the equipment is operating very stably and all parameters are within the healthy range. When the unsteady-state weighted drift intensity is significantly greater than zero, it usually indicates that the equipment has both significant parameter drift and cycle time instability, which may be a precursor to a serious failure.

[0082] S3: Based on the trend component set and the unsteady weighted drift intensity, anomaly determination is made for the current period's equipment operating status.

[0083] After calculating the quantitative indicators that comprehensively reflect the equipment status and identifying the specific trend components in S2 above, these indicators need to be transformed into clear and operable instructions so that relevant personnel can take corresponding measures for the equipment status in the current period.

[0084] Specifically, the system performs the full analysis of S2 on multiple baseline periodic data blocks collected in S1, obtaining the non-steady-state weighted drift intensity under multiple healthy states, forming a healthy baseline distribution. The mean and standard deviation of this healthy baseline distribution are then calculated, and thresholds for anomaly monitoring are set based on these. For example, the mean plus two standard deviations can be used as the lower limit threshold for states of interest, the mean plus four standard deviations as the lower limit threshold for abnormal states, and the mean plus six standard deviations as the lower limit threshold for severe abnormal states. These multiples can be configured according to the reliability requirements of different equipment and field application experience.

[0085] As a specific implementation example:

[0086] When the non-steady-state weighted drift intensity of the current period is less than the mean of the healthy baseline distribution plus twice the standard deviation, the system determines that the device is in normal condition for the current period, and only continuously monitors and updates data in the background without generating any alarms.

[0087] When the unsteady-state weighted drift intensity of the current cycle is greater than or equal to the mean plus two standard deviations but less than the mean plus four standard deviations, the system determines that the equipment in the current cycle is in a state of concern and generates a level one warning signal. This warning signal is pushed to the large-screen monitoring interface on the production site and the mobile terminals of front-line operators. The warning signal contains the sensor channels and parameter types that need attention, indicated by the trend component set, such as "injection pressure shows a slow drift trend," prompting operators to pay attention to the changing trends of relevant process parameters in advance.

[0088] As a preferred implementation example, the system can also be configured with secondary and tertiary monitoring mechanisms:

[0089] When the non-steady-state weighted drift intensity of the current period is greater than or equal to the mean plus 4 standard deviations but less than the mean plus 6 standard deviations, the system determines that the equipment is in an abnormal state and generates a level-two warning signal. This warning signal automatically generates a work order containing abnormal parameters and trend graphs, which is then pushed to the equipment maintenance team. The work order details the abnormal channels recorded in the trend component set, the spectral smoothness, and their comparison with the critical smoothing threshold, helping maintenance personnel quickly locate possible fault sources, such as micro-wear of hydraulic valve cores or micro-leakage of sealing rings.

[0090] When the non-steady-state weighted drift intensity of the current cycle is greater than or equal to the mean plus 6 times the standard deviation, the system determines that the equipment in the current cycle is in a severely abnormal state and generates a level-three emergency shutdown signal. This signal directly triggers the on-site audible and visual alarms and pops up a window in the central control system suggesting a production line shutdown. At the same time, the system pushes a document containing a detailed anomaly analysis report, reconstructed waveform spectra of all abnormal components in the trend component set, and troubleshooting steps suggested based on parameter drift characteristics to the technical supervisor and engineers.

[0091] This completes the real-time anomaly monitoring of the industrial production line.

[0092] This invention also provides a real-time anomaly monitoring system for industrial production lines. For example... Figure 3 As shown, the system includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement a real-time anomaly monitoring method for industrial production lines according to the first aspect of the present invention.

[0093] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0094] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for real-time anomaly monitoring in industrial production lines, characterized in that, include: By deploying a sensor network at key locations on the equipment, injection pressure, clamping force, hydraulic oil temperature, and mold temperature are collected at a fixed frequency to form a multidimensional time series matrix, which is then divided into multiple periodic data blocks aligned with the production cycle. Based on the baseline periodic data blocks collected offline under the health status, the baseline period duration and the standard deviation of period jitter are determined. For the current periodic data block, using multiple offline-collected reference periodic data blocks, the information entropy of each reference periodic data block is calculated through trajectory matrix construction and singular value decomposition. The arithmetic mean of the information entropy is used as the signal characteristic index. The ratio of the signal characteristic index to the theoretical maximum information entropy is determined as the time proportionality coefficient of the device. The analysis window length of the current period data block is obtained by multiplying the product of the actual duration of the current period data block and the sampling frequency by the time ratio coefficient. Based on the analysis window length, a two-dimensional analysis matrix is ​​constructed for each sensor channel. The construction of the two-dimensional analysis matrix includes: using the analysis window length as the number of rows in the matrix, starting with the first data point of the current period's data block, sequentially extracting data segments with a length equal to the analysis window length as the first column of the matrix; then sliding one data point backward, extracting the next data segment with a length equal to the analysis window length as the second column of the matrix; and so on, until all data points of the entire current period are covered. The number of columns in the analysis matrix is ​​determined by the total number of sampling points in the current period minus the analysis window length plus 1; singular value decomposition is performed on the two-dimensional analysis matrix to obtain a series of left singular vectors, each of which corresponds to an independent component in the signal. Calculate the spectral smoothness of each independent component of the analysis matrix of the current periodic data block, the spectral smoothness being determined based on the product of the differential energy value of the independent component and a pre-calculated periodic compensation factor; Independent components with spectral smoothness below a preset critical smoothness threshold are marked as trend components, forming a set of trend components; Based on the trend component set, the pure trend signal of the current period data block is reconstructed by diagonal averaging. This pure trend signal retains the gradual change trend of the physical state of the device. The parameter drift of the pure trend signal is calculated. Based on the difference between the actual duration of the current period and the duration of the reference period and the standard deviation of the period jitter, the clock instability weight of the current period data block is calculated. The parameter drift degree is multiplied by the cycle instability weight to obtain the unsteady weighted drift intensity of the equipment in the current cycle; based on the comparison result of the unsteady weighted drift intensity and the preset threshold, the abnormal operation status of the equipment in the current cycle is determined.

2. The real-time anomaly monitoring method for industrial production lines according to claim 1, characterized in that, In offline acquisition, multiple consecutive confirmed healthy periodic data blocks are collected as the reference periodic data blocks, and the average period duration of all reference periodic data blocks is used as the reference period duration. The standard deviation of the period duration of all reference periodic data blocks is used as the standard deviation of period jitter.

3. The method for real-time anomaly monitoring in industrial production lines according to claim 1, characterized in that, The calculation of the spectral smoothness includes: The differential energy value of the independent component is obtained by summing the squared differences between adjacent elements in the independent component. The cycle compensation factor is obtained by calculating the ratio of the baseline cycle length to the actual current cycle length. The spectral smoothness is obtained by multiplying the differential energy value by the period compensation factor and then taking the square root.

4. The real-time anomaly monitoring method for industrial production lines according to claim 1, characterized in that, The setting of the critical smoothing threshold includes: Singular value decomposition is performed on the baseline periodic data blocks under multiple healthy states, and the spectral smoothness of each independent component is calculated. Extract several independent components with the lowest spectral smoothness from each reference period data block as candidate components; Calculate the mean and standard deviation of the spectral smoothness of the candidate components; The critical smoothing threshold is the mean of the spectral smoothness minus a certain number of times the standard deviation of the spectral smoothness.

5. The method for real-time anomaly monitoring in industrial production lines according to claim 1, characterized in that, The calculation of the clock instability weight includes: Calculate the absolute value of the difference between the actual duration of the current cycle and the duration of the base cycle; The absolute value of the difference is divided by the standard deviation of the period jitter, and an exponential operation is performed with the natural constant as the base to obtain the beat instability weight.

6. The method for real-time anomaly monitoring in industrial production lines according to claim 1, characterized in that, The calculation of the parameter drift degree includes: Obtain the difference between the value of the pure trend signal at each moment in the current period and the health benchmark value; The square of the difference between the value of the pure trend signal at each moment in the current period and the healthy baseline value is integrated over the actual duration of the current period, divided by the actual duration of the current period, and then the square root is taken to obtain the degree of drift of the parameter. The health benchmark value is the global average of the pure trend signals corresponding to the benchmark period data blocks under multiple health states.

7. The real-time anomaly monitoring method for industrial production lines according to claim 1, characterized in that, The anomaly determination includes: A health baseline distribution is constructed based on the non-steady-state weighted drift intensity corresponding to the baseline periodic data blocks under multiple health states; Calculate the mean and standard deviation of the health baseline distribution and set monitoring thresholds for multiple levels; The current period's unsteady weighted drift intensity is compared with monitoring thresholds at multiple levels to generate corresponding anomaly detection signals.

8. A real-time anomaly monitoring system for industrial production lines, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the real-time anomaly monitoring method for industrial production line production according to any one of claims 1-7.

Images