A color masterbatch production equipment fault detection method and system

Abstract

This invention relates to the field of intelligent monitoring and fault diagnosis technology for industrial equipment, and discloses a method and system for fault detection in color masterbatch production equipment. The method includes: collecting multi-source sensor data and preprocessing it to construct a multi-dimensional sequence; performing principal component analysis to reduce dimensionality and extract low-dimensional degradation vectors; calculating time-frequency features to generate a degradation index sequence; identifying nonlinear acceleration stages and extracting potential fault signals; obtaining a health state probability distribution through a support vector machine model; determining the dominant degradation mode based on this distribution and generating early warning scheduling instructions; and fusing these instructions with historical data to determine preventative maintenance intervention points and update the reliability model. This invention achieves accurate early warning and predictive maintenance decision support for early-stage faults in color masterbatch production equipment through multi-source data fusion, nonlinear degradation trajectory identification, and probabilistic health assessment.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and fault diagnosis technology for industrial equipment, and in particular to a fault detection method and system for color masterbatch production equipment. Background Technology

[0002] Currently, color masterbatch production equipment is a key piece of equipment in the plastics product industry chain. Its reliability and stability directly determine the color consistency, batch quality, and production efficiency of the final product, and have a crucial impact on ensuring continuous production. Therefore, achieving accurate early warning of early performance degradation and potential failures is a core technological requirement for ensuring the smooth execution of production plans and avoiding losses from unplanned downtime. It is also crucial for implementing advanced fault prediction and health management concepts in this industrial field.

[0003] One existing technology primarily employs a passive equipment monitoring scheme based on manual periodic inspections and single-point static threshold alarms. This scheme typically involves deploying a single type of sensor at key locations on the equipment to collect operational data and setting fixed alarm thresholds for each monitoring parameter based on historical experience. When the monitored data exceeds its corresponding threshold, the system triggers a fault alarm. This method is essentially a static monitoring mechanism that relies on human experience and reactive response; its analysis and early warning logic are independent of the dynamic degradation process caused by the coupling of multiple factors during actual equipment operation.

[0004] This monitoring method, which relies on manual inspection and static threshold alarms, has inherent technical limitations. Because color masterbatch production equipment operates under complex conditions such as high temperature, high pressure, and high shear for extended periods, the wear, fatigue, and corrosion of its core components is a progressive, nonlinear evolutionary process involving the coupling and reinforcement of multiple factors, including mechanical, thermal, and material factors. Static thresholds cannot adapt to the weakness and time-varying nature of early degradation signals under such multi-factor interaction. This makes it difficult for the system to identify potential risks in the early stages of degradation accumulation, before the threshold is triggered, and alarms are only issued when the fault has developed to a significant stage, resulting in a severe delay in early warning. The fundamental reason is that this method can only perform threshold comparisons of isolated data points, lacking the ability to deeply fuse and analyze multi-source heterogeneous sensor data, as well as the ability to dynamically model and quantify the nonlinear degradation trajectory of equipment health status in its early stages.

[0005] Therefore, the core technical problem faced by existing technologies lies in how to solve the problem of early warning lag caused by relying on single-point static thresholds for the complex nonlinear degradation process exhibited by color masterbatch production equipment under multi-factor coupled operating conditions. That is, how to use online multi-source sensor data to construct a dynamic model that can quantitatively characterize the continuous evolution trajectory of the equipment from health and early minor wear to accelerated degradation under the condition of insufficient fault samples, and realize early and quantitative early warning of fault risks based on this model, thereby providing accurate time windows and basis for predictive maintenance decisions. Summary of the Invention

[0006] This invention provides a fault detection method and system for color masterbatch production equipment to solve the technical problems of existing technologies that rely on manual inspection and static threshold alarms, resulting in difficulty in quantifying the early nonlinear degradation trajectory of color masterbatch production equipment under the coupling effect of multiple factors, serious delay in early warning, and inability to achieve predictive maintenance decisions.

[0007] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for fault detection in color masterbatch production equipment, comprising: Collect multi-source sensor data during the operation of the color masterbatch production equipment, preprocess the multi-source sensor data, and construct a multi-dimensional data sequence; Principal component analysis is performed on the multidimensional data sequence to reduce its dimensionality and extract a low-dimensional degradation vector that represents the multi-factor coupled degradation trajectory of the device. Based on the low-dimensional degradation vector, the time-domain statistical features and frequency-domain energy distribution features within the sliding time window are calculated and concatenated to form a time-frequency joint feature vector. The deviation of the time-frequency joint feature vector from the equipment health benchmark is calculated to generate a degradation index sequence that quantifies the early wear degree of the equipment. A first-order difference operation is performed on the degradation index sequence to obtain a growth rate sequence. If the growth rate sequence exceeds a preset acceleration threshold, it is determined that the equipment degradation has entered a nonlinear acceleration stage, and potential fault signals are extracted from it. Based on the potential fault signals, and combined with a preset support vector machine model, the probability distribution vectors of the equipment belonging to different health states are analyzed. The dominant degradation mode and mode confidence are determined based on the probability distribution vector. The corresponding standard curve of degradation evolution is retrieved based on the dominant degradation mode, and the remaining effective operating time of the equipment is estimated by correcting the standard curve using the mode confidence. If the remaining effective running time is less than a preset time threshold, then instruction information containing the corresponding warning level and maintenance scheduling is generated; Based on the instruction information and degradation trajectory data from the historical operation of equipment of the same model, preventive maintenance intervention points are determined, and the instruction information is executed at the preventive maintenance intervention points.

[0008] Secondly, the present invention provides a fault detection system for color masterbatch production equipment, comprising: The data processing module is used to collect multi-source sensor data during the operation of the color masterbatch production equipment, preprocess the multi-source sensor data, and construct a multi-dimensional data sequence. The feature dimensionality reduction module is used to perform principal component analysis on the multidimensional data sequence to reduce dimensionality and extract a low-dimensional degradation vector that characterizes the multi-factor coupled degradation trajectory of the device. The index generation module is used to calculate the time-domain statistical features and frequency-domain energy distribution features within the sliding time window based on the low-dimensional degradation vector, and splice them to form a time-frequency joint feature vector. It also calculates the deviation of the time-frequency joint feature vector from the equipment health benchmark to generate a degradation index sequence that quantifies the early wear degree of the equipment. The trend recognition module is used to perform a first-order difference operation on the degradation index sequence to obtain a growth rate sequence. If the growth rate sequence exceeds a preset acceleration threshold, it is determined that the equipment degradation has entered a nonlinear acceleration stage, and potential fault signals are extracted from it. The state classification module is used to analyze the probability distribution vector of the device belonging to different health states based on the potential fault signals and in combination with a preset support vector machine model. The pattern prediction module is used to determine the dominant degradation pattern and the pattern confidence based on the probability distribution vector, retrieve the corresponding degradation evolution standard curve based on the dominant degradation pattern, and use the pattern confidence to correct the standard curve to estimate the remaining effective operating time of the equipment. The instruction generation module is used to generate instruction information containing the corresponding warning level and maintenance schedule if the remaining effective running time is less than a preset time threshold. The intervention positioning module is used to determine preventive maintenance intervention points based on the instruction information and degradation trajectory data from the historical operation of equipment of the same model, and to execute the instruction information at the preventive maintenance intervention points.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a feature extraction mechanism for complex nonlinear degradation processes by performing principal component analysis to reduce the dimensionality of multi-source sensor data and extracting low-dimensional degradation vectors to characterize the equipment degradation trajectory under the coupling effect of multiple factors. The technical derivation of this method lies in the fact that existing single-threshold methods cannot characterize the complex state evolution of multi-physics interaction. This invention integrates multi-dimensional information such as vibration, temperature, and pressure through data dimensionality reduction, and can extract the essential features characterizing the overall health status evolution of equipment from massive data, thereby solving the problem of difficulty in effectively quantifying early, weak, and coupled degradation, and providing reliable input for subsequent accurate analysis.

[0010] (2) This invention calculates the growth rate of the degradation index sequence and identifies the nonlinear acceleration stage. It then uses a support vector machine model to classify potential fault signals, achieving an intelligent diagnostic closed loop from quantified degradation to probabilistic assessment of health status. The technical derivation of this method lies in the fact that simple trend extrapolation is insufficient to adapt to the nonlinear abrupt changes in the degradation process. This invention captures the degradation acceleration inflection point through phase space reconstruction and dynamic analysis, and uses a machine learning model to map continuous signals into discrete health status probability distributions. This accurately distinguishes between different stages such as normal wear, accelerated degradation, and adjacent faults, thereby significantly improving the early identification capability and warning timeliness of potential faults.

[0011] (3) This invention generates scheduling instructions that integrate early warning levels and specific maintenance resources based on the probability distribution of health status, and dynamically optimizes maintenance intervention points and reliability models by combining historical data, thus constructing a predictive maintenance support system that integrates "status monitoring, intelligent diagnosis, and decision optimization". The technical derivation of this method lies in the fact that traditional alarms only indicate abnormalities and lack linkage decision-making with maintenance resources. This invention directly links the diagnostic results to the maintenance resource library and production plan, dynamically calculates the optimal intervention time and continuously updates the evaluation model, which can transform fault early warning information into executable maintenance work orders and model optimization basis, thereby realizing the transformation from passive response to active prediction and from isolated alarms to system decision-making, effectively reducing the risk of unplanned downtime and operation and maintenance costs. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the fault detection method for color masterbatch production equipment provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the fault detection method for color masterbatch production equipment provided in the second embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 The first embodiment of the present invention provides a method for fault detection in color masterbatch production equipment, comprising the following steps: S11, Collect multi-source sensor data during the operation of the color masterbatch production equipment, preprocess the multi-source sensor data, and construct a multi-dimensional data sequence; S12, perform principal component analysis to reduce the dimensionality of the multidimensional data sequence and extract a low-dimensional degradation vector that characterizes the multi-factor coupled degradation trajectory of the device; S13, Based on the low-dimensional degradation vector, calculate the time-domain statistical features and frequency-domain energy distribution features within the sliding time window, and splice them to form a time-frequency joint feature vector. Then, calculate the deviation of the time-frequency joint feature vector from the equipment health benchmark to generate a degradation index sequence that quantifies the early wear degree of the equipment. S14, perform a first-order difference operation on the degradation index sequence to obtain a growth rate sequence. If the growth rate sequence exceeds a preset acceleration threshold, it is determined that the equipment degradation has entered a nonlinear acceleration stage, and potential fault signals are extracted from it. S15, Based on the potential fault signals, and combined with a preset support vector machine model, analyze the probability distribution vectors of the equipment belonging to different health states; S16. Determine the dominant degradation mode and mode confidence based on the probability distribution vector, retrieve the corresponding degradation evolution standard curve based on the dominant degradation mode, and use the mode confidence to correct the standard curve to estimate the remaining effective operating time of the equipment. S17, if the remaining effective running time is less than a preset time threshold, then generate instruction information containing the corresponding warning level and maintenance scheduling; S18. Based on the instruction information and degradation trajectory data from the historical operation of the same type of equipment, determine the preventive maintenance intervention point, and execute the instruction information at the preventive maintenance intervention point.

[0015] In step S11, multi-source sensor data is collected during the operation of the color masterbatch production equipment. The multi-source sensor data is preprocessed to construct a multi-dimensional data sequence, including: Acquire multi-source sensing data output from multiple sensors, wherein the multi-source sensing data includes vibration, temperature, pressure, and shear force signals; The multi-source sensor data are subjected to analog-to-digital conversion and digital filtering respectively to obtain denoised digital signals; Perform a high-precision timestamp alignment operation on the denoised digital signal to generate a synchronized multidimensional data stream; The multidimensional data stream is subjected to data integrity verification and missing value imputation to construct the multidimensional data sequence.

[0016] First, multi-source sensor data is acquired from a multi-source sensor network deployed at key monitoring nodes of the masterbatch production equipment. Specifically, vibration acceleration sensors are installed at the motor bearing housings driving the extruder or internal mixer to collect vibration signals reflecting mechanical wear and impact; thermocouples or resistance temperature detectors (RTDs) are installed in the material mixing chamber and the housing of transmission components to collect temperature signals reflecting frictional heat and process heat distribution; pressure transmitters are installed in hydraulic system pipelines or material conveying pipelines to collect pressure signals reflecting system load and material flow status; and strain gauge sensors are installed in the agitator or screw to indirectly measure or calculate shear force signals reflecting material mixing and shear intensity. These sensors convert physical quantities into continuous analog electrical signals; for example, the vibration signal amplitude ranges from 0-5V, and the temperature signal corresponds to a 4-20mA current loop, collectively forming a multi-source sensor data source characterizing the equipment's operating status.

[0017] Subsequently, the analog electrical signals from each channel are subjected to analog-to-digital conversion and digital filtering to obtain a clean digital signal stream. The sampling frequency setting for the analog-to-digital conversion needs to be determined based on the highest effective frequency component of the monitored physical quantity. For vibration signals, to fully capture the impact response caused by early faults in rotating components such as bearings, the effective fault characteristic frequency components are typically distributed in the frequency band from tens of Hz to approximately 2000 Hz. To ensure distortion-free sampling, according to the sampling theorem, the sampling frequency of the vibration signal is set to no less than 5000 Hz to ensure distortion-free capture of potential early fault impact components. For temperature and pressure signals, which change relatively slowly, the focus is mainly on trend changes, and the sampling frequency is set to 10 Hz. The resolution of the analog-to-digital converter is at least 16 bits to ensure sufficient dynamic range and quantization accuracy. The converted digital signal stream contains high-frequency noise introduced by electromagnetic interference, line noise, etc., therefore digital filtering is required. For vibration signals, a Butterworth low-pass filter with a cutoff frequency of 2000 Hz is used to retain the fault characteristic frequency band while filtering out high-frequency noise. For temperature and pressure signals, a Butterworth low-pass filter with a cutoff frequency of 2Hz is used for smoothing to filter out high-frequency measurement noise. After the above processing, a multi-channel denoised digital signal stream is obtained.

[0018] Next, a high-precision timestamp alignment operation is performed on the multi-channel denoised digital signal stream to generate a strictly synchronized multidimensional data stream. Since there are slight differences in the sampling start time, sampling frequency, and the time it takes for data to reach the acquisition terminal via different transmission paths from different sensors, direct splicing will lead to data misalignment, affecting subsequent multivariate correlation analysis. This implementation uses a hardware clock synchronization mechanism based on the IEEE 1588 precise time protocol or similar to assign a unified absolute timestamp with an accuracy better than 1 millisecond to each data sample from different sensors. During alignment, the highest sampling frequency, 5000Hz for the vibration signal in this example, is used as the reference time axis. Linear interpolation or spline interpolation algorithms are used to resample the temperature and pressure signals with lower sampling frequencies onto a 5000Hz time grid. The interpolation algorithm, based on the assumption of signal continuity over a short period, can accurately calculate the signal value at non-sampling times, thus ensuring that at any given time point, the data in the four dimensions of vibration, temperature, pressure, and shear force strictly correspond to the device state at the same physical moment, forming a synchronized multidimensional data stream.

[0019] Then, the synchronized multidimensional data stream is subjected to data integrity verification and missing value imputation to construct a final usable multidimensional data sequence. Data integrity verification includes checking whether there is continuous data loss or outliers in the data stream due to transmission interruptions, sensor transient failures, etc. A reasonable time window is set, the length of which is determined based on the data sampling interval and the minimum continuous change period of the device state, for example, set to 10 to 100 times the sampling interval, to ensure that a statistically significant segment of device operation state is covered without over-smoothing transient anomalies. The number of data points in each channel within the window is checked to see if it matches the expected number of samples, and the presence of singular values ​​exceeding the physical range is detected. For short-term data gaps, linear interpolation based on valid data before and after is used for imputation; for longer-term gaps or uncorrectable outlier segments, the data segment is marked as invalid and excluded from subsequent analysis. After verification and filling, the processed synchronization data is arranged and stored in chronological order. Each row of data contains values ​​of multiple dimensions such as vibration, temperature, pressure, and shear force at the same timestamp, thus constructing a well-structured and reliable multidimensional data sequence.

[0020] In step S12, principal component analysis is performed on the multidimensional data sequence to reduce its dimensionality, and a low-dimensional degradation vector representing the multi-factor coupled degradation trajectory of the device is extracted, including: The multidimensional data sequence is subjected to zero-mean standardization to obtain a standardized data matrix, and the covariance matrix between each dimension is calculated based on the standardized data matrix. The covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvectors and eigenvalues. The principal component directions are selected according to the eigenvalues ​​in descending order to form the projection transformation matrix. Multiply the standardized data matrix by the projection transformation matrix to obtain the principal component score matrix; The principal component score matrix is ​​reconstructed into a low-dimensional degenerate vector in chronological order.

[0021] First, zero-mean standardization is performed on the multidimensional data sequence constructed in step S11. The multidimensional data sequence is a matrix whose data points are time points and whose columns represent different sensor signal dimensions, such as vibration, temperature, pressure, and shear force. Since each signal has different physical meanings and dimensions, direct mathematical operations will be biased towards larger numerical dimensions due to differences in numerical scale. Therefore, Z-score standardization is required for each column of signal data to ensure that the mean of each dimension is zero and the standard deviation is one, thereby eliminating the influence of dimensions and bringing all dimensions to the same order of magnitude. After all columns are standardized, a standardized data matrix is ​​formed.

[0022] Subsequently, the covariance matrix of each dimension is calculated based on the standardized data matrix. Covariance is used to measure the degree of linear correlation between two variables, and its purpose is to quantitatively reveal the statistical relationship patterns of mutual influence and coupling of physical quantities such as vibration, temperature, pressure, and shear force during equipment operation.

[0023] Next, eigenvalue decomposition is performed on the calculated covariance matrix. This decomposition yields a series of eigenvalues ​​and their corresponding eigenvectors. The magnitude of the eigenvalue represents the amount of original data variance carried by the direction of its corresponding eigenvector, while the eigenvector defines the new orthogonal coordinate axis directions, i.e., the principal component directions. The direction with the largest variance represents the most prevalent pattern of data variation.

[0024] Then, based on the eigenvalues ​​sorted from largest to smallest, the top [values] are selected. The principal component directions are used to construct the projection transformation matrix. The number of principal components selected... This is a key parameter, and its specific value can be determined by setting a cumulative contribution rate threshold, i.e., selecting the minimum number of... This makes the former The cumulative variance contribution rate of each principal component reaches or exceeds this threshold, such as 85%. The selected first... The eigenvectors are arranged column-wise to form the projection transformation matrix. Then, the standardized data matrix is ​​multiplied by the projection transformation matrix to perform a linear projection transformation, yielding the principal component score matrix. Each row of this matrix represents a time point in the new... The coordinates in the principal component space are called principal component scores. They are linear combinations of the original sensing signals and are sorted in descending order of the amount of information they carry.

[0025] Finally, the principal component score matrix is ​​reconstructed into the low-dimensional degenerate vector in chronological order. Specifically, each column of the score matrix is ​​arranged in chronological order to form a time series; Such a combination of time series constitutes a A dimensional, time-evolving sequence of vectors, which is the low-dimensional degenerate vector.

[0026] In step S13, based on the low-dimensional degradation vector, the time-domain statistical characteristics and frequency-domain energy distribution characteristics within the sliding time window are calculated and concatenated to form a time-frequency joint feature vector. The deviation of the time-frequency joint feature vector from the equipment health baseline is then calculated to generate a degradation index sequence that quantifies the early wear degree of the equipment, including: A series of time window slices are obtained by truncating the low-dimensional degenerate vector using a sliding window of fixed length. For each time window slice, calculate the statistical mean vector and the statistical variance vector; Perform a Fast Fourier Transform on the time window slices to extract the energy percentage within the preset frequency band and obtain the spectral energy feature vector; The statistical mean vector, the statistical variance vector, and the spectral energy feature vector are concatenated to form a time-frequency joint feature vector; Based on the data collected during the initial health operation phase of the equipment, the average value of the time-frequency joint feature vector calculated through the same process is used to predetermine the equipment health baseline vector; Calculate the Euclidean distance between the time-frequency joint feature vector and a predetermined device health baseline vector to generate a degradation index sequence.

[0027] First, a fixed-length sliding window is used to truncate the low-dimensional degradation vector obtained in step S12. The low-dimensional degradation vector is a multi-dimensional vector sequence arranged chronologically. To analyze the local evolution characteristics of the equipment state, it needs to be divided into continuous, potentially overlapping time periods. The length of the sliding window is a key parameter, and its determination requires a trade-off between feature stability and sensitivity to change. A window that is too long will smooth out subtle changes caused by early degradation, while a window that is too short is susceptible to random fluctuations. In this embodiment, the window length is set based on the main operating cycle of the equipment and the time scale of fault evolution. For example, for a batch-produced masterbatch extruder, a complete production cycle is approximately 2 hours. Setting the window length to 1 hour, or half a cycle, ensures that each window covers a relatively complete stage of equipment operation while capturing cross-cycle trend changes. The window sliding step size is set to one-quarter of the window length, such as 15 minutes, to generate overlapping window slices, ensuring that no state transitions occurring near the window boundaries are missed.

[0028] Subsequently, for each time window slice, its statistical mean vector and statistical variance vector are calculated. For a window slice containing several time points, each dimension is calculated independently. The statistical mean vector is obtained by taking the arithmetic mean of the values ​​of all time points within the window across each dimension, while the statistical variance vector is obtained by calculating the variance of the values ​​across each dimension. A shift in the mean vector may indicate a baseline drift in performance, while an increase in the variance vector may reflect an increase in operational instability. These two vectors characterize the operating characteristics of the device over a short period of time from the perspectives of central tendency and dispersion.

[0029] Next, a Fast Fourier Transform (FFT) is performed on the time window slice, and the energy percentage within the preset frequency band is extracted to obtain a spectral energy feature vector. The fluctuation pattern in the time domain manifests as a specific energy distribution in the frequency domain. For the time series of the vector in each dimension of the window slice, i.e., each principal component dimension, a FFT is performed to transform it from the time domain to the frequency domain, obtaining its power spectrum. The preset frequency band is divided based on the physical mechanism of the equipment failure. For example, for periodic failures caused by imbalance, asymmetry, or bearing damage in rotating parts, the vibration energy will be concentrated near a characteristic frequency that is a multiple of the rotational frequency; while wideband excitation caused by wear or loosening may cause energy to diffuse to higher frequencies. Therefore, the frequency band can be divided into low-frequency, mid-frequency, and high-frequency bands. The total energy of the signal power spectrum within each preset frequency band is calculated, and its percentage of the total energy is calculated to obtain a vector characterizing the energy distribution. For all principal component dimensions, calculations can be performed separately or comprehensively to ultimately form a spectral energy feature vector. This feature can capture periodic or resonant phenomena associated with specific failure modes that cannot be reflected by time-domain statistics.

[0030] Then, the statistical mean vector, statistical variance vector, and spectral energy feature vector calculated for the same time window are normalized to eliminate the dimensional influence caused by different physical meanings and numerical scales of different features. The normalization process uses the Z-score standardization method, which involves subtracting the mean of historical health data for each feature dimension from the mean of that dimension and then dividing by its standard deviation, thus converting the data for each dimension into a standard distribution with a mean of 0 and a standard deviation of 1. Subsequently, the normalized statistical mean vector, statistical variance vector, and spectral energy feature vector are concatenated to form a comprehensive time-frequency joint feature vector. Finally, the Euclidean distance between the normalized time-frequency joint feature vector corresponding to each time window and the equipment health baseline vector is calculated to generate a degradation index sequence. The equipment health baseline vector is the average value of the time-frequency joint feature vector calculated using the same process described above, based on data collected for a sufficient period during the initial stable and fault-free operation after factory commissioning or major overhaul. This baseline represents the feature profile of the equipment under ideal health conditions. For each window's time-frequency joint feature vector generated in real time during online monitoring, the Euclidean distance between it and this fixed health baseline vector is calculated. The larger the distance, the more significant the degradation. Arranging these distance values ​​in the order of time windows forms the degradation index sequence.

[0031] In step S14, a first-order difference operation is performed on the degradation index sequence to obtain a growth rate sequence. If the growth rate sequence exceeds a preset acceleration threshold, it is determined that the equipment degradation has entered a nonlinear acceleration stage, and potential fault signals are extracted from it, including: The degradation index sequence is subjected to a first-order difference operation to obtain an instantaneous growth rate sequence, and the instantaneous growth rate sequence is mapped to the reconstructed phase space to form a dynamic evolution trajectory; Calculate the local tangent slope at each point on the dynamic evolution trajectory. If the local tangent slope continuously exceeds the preset acceleration threshold, it is determined that the equipment degradation has entered the nonlinear acceleration stage. Local maxima of the degradation index sequence during the nonlinear acceleration phase are extracted, an evolution envelope is constructed based on the local maxima, and the divergence of the evolution envelope is calculated. The signal segments whose divergence exceeds a preset fluctuation threshold are marked as potential fault signals.

[0032] First, a first-order difference operation is performed on the degradation index sequence generated in step S13 to obtain a growth rate sequence characterizing its instantaneous rate of change. This sequence directly reflects the instantaneous speed of equipment health deterioration. To further reveal its inherent dynamic evolution pattern, especially the possible chaotic or nonlinear attractor characteristics, this time series needs to be mapped to a higher-dimensional phase space. The time delay is used to determine the interval between adjacent components in the reconstructed time series. Its determination needs to ensure that the new coordinate components contain new information and have a certain correlation with the original sequence. One method for determination is the autocorrelation function method. In specific implementation, the Pearson correlation coefficient between the growth rate sequence and its own sequence formed after different time offsets is first calculated to obtain the curve of the autocorrelation coefficient changing with the delay time. Then, the search is performed along the direction of increasing delay time, and the delay time corresponding to the first drop of the autocorrelation coefficient to zero or a certain preset small positive value is determined as the time delay.

[0033] The embedding dimension determines the dimension of the reconstructed phase space. Its value must be high enough to fully represent the intrinsic structure of the system's dynamics and avoid undue overlap of trajectories across different dimensions. One method for determining this is the spurious nearest neighbor method. In practice, the phase space is reconstructed starting with a relatively small initial embedding dimension. For each data point in the phase space, its nearest neighbor in the current dimension is found (i.e., the point with the closest Euclidean distance). Then, the embedding dimension is increased by 1, and the distance between these two points is recalculated in the new higher-dimensional space. If the increase in distance exceeds a preset threshold, the nearest neighbor is considered "spurious" in the current lower dimension; that is, their proximity is an illusion caused by insufficient dimension or projection, rather than a true dynamic proximity. By calculating the proportion of spurious nearest neighbors among all data points and gradually increasing the embedding dimension, when this proportion falls below a preset convergence threshold, the current embedding dimension is considered sufficient to effectively unfold the system's attractor, and this is determined as the final embedding dimension. By reconstructing a one-dimensional growth rate sequence into a series of points in a multi-dimensional phase space through time-delay embedding, the trajectory formed by these points in the high-dimensional space, i.e. the dynamic evolution trajectory, can more comprehensively reveal the geometric structure and dynamic characteristics of the system's state evolution.

[0034] Subsequently, the local geometric characteristics of the dynamic evolution trajectory are analyzed, and the local tangent slope at each point on the trajectory is calculated. In phase space, the tangent direction of the trajectory represents the instantaneous evolution direction of the system state, and its slope quantifies the speed at which the state evolves in that direction. For a point on the trajectory, a local hyperplane or direction vector is fitted using the least squares method by selecting several of its nearest neighboring points. The magnitude of this direction vector or its component in the main evolution direction can be regarded as the local tangent slope at that point. The determination of this threshold is not subjective, but based on the statistical distinction between normal degradation and accelerated degradation modes. Specifically, a large amount of historical operating data is collected, including samples of the entire process from the slow evolution of normal wear to the final failure of accelerated degradation. In the samples, "acceleration start points" are marked based on clear performance inflection points, such as a sudden drop in efficiency or excessive vibration. The local tangent slope distribution of all samples before the acceleration start point and the slope distribution after the acceleration start point are calculated. Through statistical analysis, the statistical distance between the two distributions is calculated to determine a slope value that best distinguishes between "normal evolution rate" and "accelerated evolution rate," which is then set as the critical acceleration threshold. During online monitoring, if the local tangent slopes at multiple consecutive sampling points are all greater than this threshold, it is determined that the equipment degradation has entered the nonlinear acceleration stage. This continuous judgment can effectively filter out instantaneous slope spikes caused by accidental disturbances.

[0035] Next, for the identified nonlinear acceleration phase, the evolution of its internal fluctuation patterns is further analyzed. All local maxima of the original degradation index sequence within this phase are extracted. These maxima correspond to a series of peaks in the degradation index during the accelerated rise, potentially reflecting intermittently intensified wear events or shocks. Based on these discrete peaks, a smooth curve, the evolution envelope, is constructed using cubic spline interpolation. This envelope outlines the upper boundary of the degradation index fluctuation amplitude during the acceleration phase. To quantify the degree of disorder or diffusion trend of the fluctuation patterns, the divergence of this envelope is calculated. A common method is to perform a first-order difference on the envelope sequence to obtain its change, and then calculate the root mean square value of this change as the divergence. A larger divergence value indicates more severe and irregular fluctuations in the envelope, potentially signifying increased instability in the degradation process and a higher risk of failure.

[0036] Finally, the most noteworthy potential fault signal segments are screened and labeled based on divergence. This needs to be compared with a preset fluctuation threshold. The determination of the fluctuation threshold is also based on historical data learning. In the labeled historical acceleration phase samples, the divergence of each sample is calculated. Based on whether these samples eventually develop into faults requiring emergency shutdown in a short period of time, they are divided into two categories: "high-risk acceleration" and "non-high-risk acceleration". By analyzing the statistical distribution of the divergence values ​​of the two types of samples, for example, using receiver operating characteristic (ROC) curve analysis, to maximize classification accuracy or determine an acceptable false alarm rate level, an optimal fluctuation threshold is determined. In the online application, for each identified nonlinear acceleration phase, its envelope divergence is calculated. If the divergence is greater than the fluctuation threshold, it is considered that the acceleration phase is not only accelerating, but its fluctuation pattern also exhibits high-risk characteristics. Therefore, the original signal of this time period is explicitly marked as a "potential fault signal" to trigger subsequent higher-level diagnosis and warning. Conversely, if the divergence does not exceed the threshold, only an acceleration status is indicated, but the alarm level is not escalated.

[0037] In step S15, based on the potential fault signal and combined with a preset support vector machine model, the probability distribution vectors of the equipment belonging to different health states are analyzed, including: Extract time-domain statistical amplitude features from the potential fault signals to construct a multi-dimensional feature input vector; The multidimensional feature input vector is input into the support vector machine model to calculate the signed geometric distance from the multidimensional feature to the optimal classification hyperplane; The signed geometric distance is input into a preset probability calibration function to obtain the posterior probability that the potential fault signal belongs to various health states. The posterior probabilities output by multiple binary classifiers in the support vector machine model are aggregated to form a probability distribution vector.

[0038] First, a set of time-domain statistical amplitude features capable of effectively distinguishing different health states is extracted from the potential fault signals marked in step S14 to construct a multi-dimensional feature input vector for classification. The potential fault signal is a segment of time-series data representing a phase of nonlinear accelerated degradation of the equipment. The extracted features specifically include the signal's peak value, root mean square (RMS) value, kurtosis, and impulse factor. For example, for a 10-second segment of vibration acceleration potential fault signal, its peak value is calculated to be 5.2 m / s², its RMS value to be 1.8 m / s², its kurtosis to be 6.5, and its impulse factor to be 4.1. These feature values ​​are arranged in a fixed order to form a multi-dimensional feature input vector.

[0039] Subsequently, the constructed multidimensional feature input vector is fed into a pre-trained support vector machine (SVM) model. This SVM model is a classification model based on radial basis function kernels. Its training process includes collecting historical data of the color masterbatch production equipment operating under various known health states, such as normal, slight degradation, moderate acceleration, and severe degradation, extracting the same time-domain statistical amplitude features as those in the online monitoring phase, and forming a training sample set with clear category labels. The model is then trained using this sample set to find the optimal classification hyperplane that can separate samples of different categories with the maximum margin.

[0040] Subsequently, the constructed multidimensional feature input vector is fed into a pre-trained support vector machine (SVM) model. This SVM model is a classification model based on radial basis function kernels. Its training process includes collecting historical data of the color masterbatch production equipment operating under various known health states, such as normal, slight degradation, moderate acceleration, and severe degradation, extracting the same time-domain statistical amplitude features as those in the online monitoring phase, and forming a training sample set with clear category labels. The model is then trained using this sample set to find the optimal classification hyperplane that can separate samples of different categories with the maximum margin.

[0041] The specific training process is a supervised optimization process, consisting of two parts: first, maximizing the classification margin; and second, tolerating some classification errors by introducing slack variables and penalty parameters, thereby achieving a balance between model complexity and training error. This optimization problem is ultimately transformed into a convex quadratic programming problem. To efficiently solve this optimization problem, the implementation of this invention typically employs a sequential minimum optimization algorithm. This algorithm decomposes the large-scale optimization problem into a series of minimal, easily solvable bivariate subproblems and iteratively optimizes these subproblems to approximate the global optimum. In each iteration, the algorithm selects a pair of Lagrange multipliers for optimization according to heuristic rules and updates the model parameters. The convergence condition for training is set based on the state of the optimization process. A common strategy is to determine that the training process has converged when the decrease in the objective function value is less than a preset minimum threshold or when a preset maximum number of iterations is reached in consecutive iterations, stopping the iteration and outputting the model parameters obtained at this point, thus completing the model training.

[0042] Then, due to the standard SVM output decision value and geometric distance Instead of probability values, to obtain more intuitive posterior probabilities belonging to each health state, geometric distances need to be considered. A preset probability calibration function is input for mapping. This implementation uses the Pratt scaling method for probability calibration. This method assumes that the probability of a sample belonging to the positive class follows a Sigmoid function distribution. ,in and The parameters to be fitted are defined. The calibration process is performed using a separate validation dataset after model training. This validation dataset also originates from historical operating data, but its samples are not used in the training of the aforementioned model hyperplane; it is specifically used for calibration. From the historical operating data of the color masterbatch production equipment, a subset of samples with clear equipment health status labels, which were not used during model training, is selected. The same temporal statistical amplitude features are extracted from these samples to form a validation set specifically for probabilistic calibration. The validation set contains feature vectors and their true class labels. The geometric distance is calculated using all samples on this validation set. The optimal parameters are fitted using the maximum likelihood estimation method, along with the labels. and The trained calibration model can calculate the geometric distance to any new sample. The value is mapped to a probability value between 0 and 1, which is the posterior probability estimate of whether the sample belongs to "accelerated degradation" or a specific health state.

[0043] Finally, since the health status of color masterbatch production equipment is usually not limited to two categories, such as normal, slight degradation, moderate acceleration, and severe degradation, this invention employs a "one-to-one" or "one-to-the-others" strategy to construct multiple binary classification SVM sub-models. For example, for a four-class classification problem, the "one-to-one" strategy requires training six binary classifiers. Each binary classifier is trained for a pair of classes and outputs a posterior probability of belonging to its positive class for a new input feature vector. The aggregation of probabilities requires a robust strategy. For the "one-to-one" strategy, a combination of "voting" and "weighted probability averaging" is used. Each binary classifier casts a vote for its predicted class, and its output posterior probability is used as the weight of that vote. The votes of all binary classifiers are statistically analyzed. For each class, the average probability output by all binary classifiers predicting that class is calculated as the final probability that the sample belongs to that class. The final probabilities of all classes are arranged in a fixed order, such as normal, slight, moderate, and severe, forming a probability distribution vector, and the sum of all probabilities is 1. This vector clearly quantifies the probability of the device being in various health states, providing a probabilistic and interpretable basis for subsequent risk assessment and maintenance decisions.

[0044] In step S16, the dominant degradation mode and mode confidence are determined based on the probability distribution vector. A standard curve of the corresponding degradation evolution is retrieved based on the dominant degradation mode, and the remaining effective operating time of the equipment is calculated by correcting the standard curve using the mode confidence. This includes: Identify the component with the largest value in the probability distribution vector, determine the health state corresponding to the component with the largest value as the dominant degradation mode, and use the value of the component with the largest value as the mode confidence. According to the dominant degradation mode, the corresponding standard curve of degradation evolution is retrieved from the pre-established standard curve library of degradation evolution generated based on the statistical analysis of historical failure data of similar equipment. The standard curve is scaled and corrected using the pattern confidence level to obtain a corrected degradation trajectory, and the remaining effective operating time of the device is estimated along the corrected degradation trajectory.

[0045] First, the probability distribution vector from step S15 is received. This vector contains the posterior probability of the device belonging to each preset health state category, with the sum of all probabilities being 1. To transform the probability information into a clear decision-making basis, it is necessary to determine the most probable dominant degradation mode and its confidence level. By comparing the values ​​of each component in the probability distribution vector, the component with the largest probability value corresponding to the maximum value is identified. The health state corresponding to this component is determined as the dominant degradation mode, and this maximum probability value is directly used as the mode confidence level.

[0046] Subsequently, based on the determined dominant degradation pattern Retrieve and retrieve the associated standard curves of degradation and evolution from a pre-built knowledge base. This curve is derived through statistical regression fitting based on historical data of similar equipment under the same degradation mode, showing the entire process from a similar state to functional failure. It characterizes the typical trajectory of degradation indicators evolving over time. For example, for the "moderate wear of extruder screws" mode, the standard curve might describe the remaining useful life as decaying from its initial value to the failure threshold in approximately 400 hours. However, the standard curve reflects a group average; individual equipment will exhibit different degradation rates due to differences in operating load, maintenance history, and environmental conditions. Therefore, it is necessary to utilize mode confidence levels. The standard curve is customized to obtain a corrected degradation trajectory that better reflects the actual conditions of the current equipment. The core idea behind the correction is that confidence level not only reflects the certainty of pattern judgment but may also implicitly suggest how close the device's current state is to the standard pattern. A confidence-dependent time scaling factor is then used. Adjust the time axis of the standard curve: scaling factor It is about The monotonic function, whose specific form is learned through historical data, is given by... A feasible parameterized form is... ,in The adjustment coefficients are determined by minimizing the mean square error between the actual failure time of historical individual devices and the predicted time corrected based on their confidence level. hour, The time axis is compressed, indicating that the predicted degradation rate is accelerated; when hour, The timeline has been expanded, indicating a more conservative prediction. (Revised trajectory) This refers to the prediction of personalized degradation paths for the current device.

[0047] Next, along the corrected degradation trajectory Estimate the remaining effective operating time of the equipment The remaining effective runtime is defined as the time remaining from the current moment. This process continues until the equipment's health indicators first reach or exceed the preset failure threshold. The predicted time elapsed. Mathematically, this requires solving an equation. .like For analytical functions, the solution can be obtained directly. For discrete sequences or nonparametric models, estimation can be performed using linear interpolation or extrapolation. Simultaneously, a Bayesian update framework is employed to update the latest degradation index observations monitored online. As new evidence, the parameters of the trajectory model are continuously updated. The posterior distribution provides a dynamic, uncertain range of remaining lifetime predictions.

[0048] In step S17, if the remaining effective running time is less than a preset time threshold, instruction information containing the corresponding warning level and maintenance scheduling is generated, including: When the remaining effective running time is less than the preset emergency time threshold, a high-risk warning level is determined; Based on the high-risk warning level and the current maintenance resource status, instruction information including specific maintenance time, spare parts list and personnel arrangement is generated.

[0049] First, the dynamically updated remaining effective operating time is compared with a preset emergency time threshold to determine the warning level. Setting the emergency time threshold is a multi-objective decision-making process. The method involves first statistically analyzing the average time required to complete a single preventative maintenance cycle for the dominant degradation mode, including spare parts allocation, personnel arrangement, downtime application, on-site work, and testing and recovery. Then, a safety factor greater than 1 is introduced to address uncertainties and potential emergencies during task execution. Finally, the product of this average time and the safety factor is set as the emergency time threshold. The specific value of this safety factor is determined based on factors such as the volatility of historical maintenance task execution times, the stability of critical spare parts supply, and the rigidity of production plans. For example, by analyzing the distribution of deviations between actual and planned maintenance times in historical data, the safety factor is set to a value that covers the vast majority, such as 95% of cases, for example, between 1.2 and 1.5. When the predicted remaining effective operating time is less than or equal to this emergency time threshold, the equipment is determined to have entered a high-risk state, and a high-risk warning level must be triggered immediately. Multiple levels of early warning can be set up, with the following specific rules: when the predicted remaining effective running time is less than or equal to the emergency time threshold, it is determined to be a high-risk level and a red alert is triggered; when the remaining effective running time is greater than the emergency time threshold but less than twice it, it is determined to be a medium-risk level and a yellow alert is triggered; when the remaining effective running time is greater than twice the emergency time threshold, it is determined to be a low-risk or observation level and a blue alert is triggered or only the status trend is recorded.

[0050] Finally, based on the determined high-risk warning level, the system automatically generates structured, executable maintenance scheduling instructions by linking the maintenance resource database. The maintenance resource database stores real-time information such as spare parts inventory status, maintenance personnel skill matrix and scheduling table, and potential downtime windows in the equipment production plan. The instruction generation process is an optimization decision involving resource and time matching, specifically including the following three steps. First, it receives warning information, with the core constraint that the maintenance task must be started and completed within a time window of [current time + current time + remaining effective running time]. Simultaneously, it reads the real-time status of the maintenance resource database to obtain the following information: the real-time quantity and location of spare parts inventory, the available time slots for maintenance personnel with the corresponding skills, and potential downtime windows extracted from the production plan. Next, with minimizing the impact of equipment failure as the optimization objective, and under the premise of satisfying the above hard time constraints, spare parts availability constraints, and personnel skill and time constraints, a feasible maintenance execution plan is generated through search and optimization algorithms.

[0051] Specifically, an optimization strategy combining heuristic rules and local search can be adopted: First, based on spare parts inventory and personnel scheduling, all potential candidate time periods that meet resource constraints are initially screened. Then, for each candidate time period, its corresponding comprehensive cost is calculated, including production losses due to downtime and potential additional maintenance costs. Next, the candidate time period with the lowest comprehensive cost is selected as the optimal maintenance time window, and the corresponding spare parts requisition plan and personnel assignment plan are determined. Finally, based on the optimization decision, structured maintenance scheduling instructions are automatically generated. The instructions specifically include: the optimal maintenance time window, the required spare parts list and quantity and requisition location, the list of assigned technicians or teams, and recommended maintenance operation guidelines or historical case references. The generated instructions are packaged in a standard format and automatically pushed to the Manufacturing Execution System, the computerized maintenance management system, or the mobile terminals of relevant personnel via the Enterprise Service Bus or API interface, thereby completing an intelligent closed loop from fault warning to maintenance task creation.

[0052] In step S18, preventive maintenance intervention points are determined based on the instruction information and degradation trajectory data from the historical operation of equipment of the same model, and the instruction information is executed at the preventive maintenance intervention points.

[0053] Specifically, based on the instruction information and degradation trajectory data from the historical operation of equipment of the same model, preventive maintenance intervention points are determined, including: Taking the planned maintenance time points and historical degradation trajectory data in the instruction information as input, perform time axis alignment and data fusion operations to output a multi-dimensional fused dataset. Calculate the deviation between the current degraded trajectory and historical similar trajectories in the multidimensional fusion dataset, and calculate the risk evolution rate based on the deviation. The remaining life prediction value of the equipment is dynamically updated based on the risk evolution rate. The updated remaining life forecast is matched with the resource gaps in the production plan to determine the optimal preventative maintenance intervention point.

[0054] First, the system receives the maintenance schedule instruction information generated in step S17. This instruction information already contains a planned maintenance time point determined based on preliminary predictions and resource availability. Simultaneously, a complete degradation trajectory dataset of the same model as the current equipment and historically experienced the same dominant degradation mode is retrieved from the long-term stored data warehouse. Each historical trajectory is a time-series record of degradation indicators from the moment the equipment's state is first determined to have entered the dominant degradation mode until final repair or failure.

[0055] For effective comparison and fusion, the real-time degradation trajectory of the current device since entering this mode needs to be aligned with each historical trajectory on the timeline. Alignment employs dynamic time warping, an algorithm capable of handling non-linear scaling of two trajectories over time, finding the optimal matching path between them. This maps each data point of the current trajectory to the corresponding time point of each historical trajectory, forming a time-aligned multidimensional fusion dataset. This ensures that subsequent comparisons are performed at the same "degradation stage" rather than at absolute "calendar time."

[0056] Subsequently, based on the fused dataset, the trajectory deviation degree between the current degradation trajectory and each historical trajectory of the same type is calculated. The deviation degree quantifies the extent to which the current device degradation process deviates from the historical typical pattern. An effective calculation method is to use the Fraser distance. This distance metric aims to assess the overall similarity between two curves, considering not only the distance between corresponding points but also the differences in the shape and continuity of the two paths. This distance value can be calculated using methods such as dynamic programming, and used as the trajectory deviation degree. After obtaining the deviation degree from all historical trajectories, its statistical characteristics, such as the mean and standard deviation, are calculated. The risk evolution rate is defined as the rate of change of the deviation degree over time. Since the deviation degree is calculated point-by-point as the current trajectory extends, a sequence of deviation degree changes over time can be constructed, and the average of the first differences can be calculated to obtain the risk evolution rate. If the risk evolution rate is greater than zero, it indicates that the current degradation process is evolving in a way that increasingly deviates from the historical typical path, meaning that the risk is accumulating at an accelerated pace.

[0057] Next, the calculated risk evolution rate is used. Dynamically update the remaining effective running time initially estimated in step S16. The updated model is based on a feedback correction mechanism. It assumes the original remaining lifetime prediction... Based on standard correction trajectory The extrapolated trajectory does not account for new risks arising from deviations from historical typical patterns. Therefore, a correction factor driven by the rate of risk evolution is introduced. This factor is a monotonically increasing function, and its specific form is determined by analyzing the relationship between the rate of change of trajectory deviation and the reduction in actual remaining life in historical data. For example, an exponential form can be used. ,in The sensitivity parameter is obtained by fitting historical cases. It is estimated by the proportion of the final actual lifespan to the standard predicted lifespan. The updated remaining lifespan prediction is: This formula implies that a positive rate of risk evolution will lead to a shorter predicted remaining lifespan, thus providing a more conservative and responsive forecast.

[0058] Then, the updated remaining lifetime forecast is matched with the resource window in the production plan to determine the optimal preventative maintenance intervention point. A resource window refers to a set of time windows during which equipment can be shut down for maintenance without affecting critical production tasks. Determining the intervention point is a constrained optimization problem. The goal is to find a window of opportunity where planned maintenance activities can be completed within that window, and the maintenance start time is as late as possible to utilize the remaining value of the equipment, but must be earlier than the predicted failure time to mitigate risk. Specifically, a maintenance start time point needs to be found that satisfies the following conditions: the maintenance start time plus the maintenance time does not exceed the end time of the selected window; the maintenance start time is not earlier than the start time of the window; and the maintenance start time is not later than the current time plus the updated remaining lifetime forecast minus the maintenance time. The system iterates through all feasible window periods and selects the window that meets the conditions and has the latest maintenance start time. The start time of the window serves as the optimal intervention point. Based on this intervention point, the instruction information determined in step S17 is executed. This "latest feasible time" principle aims to maximize the effective production time of the equipment and utilize its remaining value while ensuring maintenance safety. If no such window exists, an alarm is triggered, requiring resource coordination or adjustments to the production plan to create a maintenance window.

[0059] In summary, this invention discloses a fault detection method for color masterbatch production equipment, comprising: collecting multi-source sensor data from the equipment and performing time-series alignment and noise filtering to construct a multi-dimensional data sequence; performing principal component analysis to reduce the dimensionality of the sequence and extracting a low-dimensional degradation vector characterizing the multi-factor coupled degradation trajectory; calculating time-frequency features based on the vector to generate a degradation index sequence that quantifies early wear; determining whether degradation has entered an accelerated phase and extracting potential fault signals by analyzing the growth rate and nonlinear dynamic trajectory of the sequence; classifying the signals using a support vector machine model and outputting a health state probability distribution vector; determining the dominant degradation mode and confidence level based on the vector and generating a scheduling instruction that integrates early warning levels and specific maintenance resources; and finally, fusing the instruction with historical data to dynamically determine the optimal preventive maintenance intervention point and update the equipment reliability assessment model. This invention overcomes the limitation of single signals being insensitive to early, minor wear by deeply integrating multi-source sensor data such as vibration, temperature, pressure, and shear force, and using principal component analysis to extract the essential degradation trajectory from high-dimensional coupled information. By introducing phase space reconstruction and nonlinear dynamic analysis, it achieves accurate identification and early warning of equipment degradation from linear gradual change to nonlinear acceleration. By constructing a health status probabilistic assessment model based on support vector machines and a personalized remaining life prediction model based on confidence, qualitative alarms are upgraded to quantitative and interpretable risk probability outputs. Finally, by linking and optimizing intelligent diagnostic results with maintenance resources and production plans, executable maintenance scheduling instructions are automatically generated and the reliability model is continuously corrected, realizing a paradigm shift from passive alarms and post-event maintenance to early warning and predictive maintenance, significantly improving the operational reliability, maintenance efficiency, and production continuity of masterbatch production equipment.

[0060] Reference Figure 2 The first embodiment of the present invention provides a fault detection system for color masterbatch production equipment, comprising: The data preprocessing module is used to collect multi-source sensor data during the operation of the color masterbatch production equipment, perform time-series alignment and noise filtering on the multi-source sensor data, and construct a multi-dimensional data sequence. The feature dimensionality reduction module is used to perform principal component analysis on the multidimensional data sequence to reduce dimensionality and extract a low-dimensional degradation vector that characterizes the multi-factor coupled degradation trajectory of the device. The index generation module is used to calculate the time-domain statistical features and frequency-domain energy distribution features within the sliding time window based on the low-dimensional degradation vector, and splice them to form a time-frequency joint feature vector. It also calculates the deviation of the time-frequency joint feature vector from the equipment health benchmark to generate a degradation index sequence that quantifies the early wear degree of the equipment. The trend recognition module is used to perform a first-order difference operation on the degradation index sequence to obtain a growth rate sequence. If the growth rate sequence exceeds a preset acceleration threshold, it is determined that the equipment degradation has entered a nonlinear acceleration stage, and potential fault signals of this stage are extracted. The state classification module is used to analyze the probability distribution vector of the device belonging to different health states based on the potential fault signals and in combination with a preset support vector machine model. The scheduling generation module is used to determine the dominant degradation mode and mode confidence based on the probability distribution vector, retrieve the corresponding degradation evolution standard curve based on the dominant degradation mode, and use the mode confidence to correct the standard curve to calculate the remaining effective operating time of the equipment. If the remaining effective operating time is less than a preset time threshold, instruction information containing the corresponding warning level and maintenance scheduling is generated. The decision optimization module is used to determine preventive maintenance intervention points and update the pre-built equipment reliability assessment model based on the instruction information and degradation trajectory data from the historical operation of equipment of the same model.

[0061] It should be noted that the color masterbatch production equipment fault detection system provided in this embodiment of the invention is used to execute all the process steps of the color masterbatch production equipment fault detection method in the above embodiment. The working principle and beneficial effects of the two are one-to-one, so they will not be described again.

[0062] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0063] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A colorant masterbatch production apparatus failure detection method characterized by, include: Collect multi-source sensor data during the operation of the color masterbatch production equipment, preprocess the multi-source sensor data, and construct a multi-dimensional data sequence; Principal component analysis is performed on the multidimensional data sequence to reduce its dimensionality and extract a low-dimensional degradation vector that represents the multi-factor coupled degradation trajectory of the device. Based on the low-dimensional degradation vector, the time-domain statistical features and frequency-domain energy distribution features within the sliding time window are calculated and concatenated to form a time-frequency joint feature vector. The deviation of the time-frequency joint feature vector from the equipment health benchmark is calculated to generate a degradation index sequence that quantifies the early wear degree of the equipment. A first-order difference operation is performed on the degradation index sequence to obtain a growth rate sequence. If the growth rate sequence exceeds a preset acceleration threshold, it is determined that the equipment degradation has entered a nonlinear acceleration stage, and potential fault signals are extracted from it. Based on the potential fault signals, and combined with a preset support vector machine model, the probability distribution vectors of the equipment belonging to different health states are analyzed. The dominant degradation mode and mode confidence are determined based on the probability distribution vector. The corresponding standard curve of degradation evolution is retrieved based on the dominant degradation mode, and the remaining effective operating time of the equipment is estimated by correcting the standard curve using the mode confidence. If the remaining effective running time is less than a preset time threshold, then instruction information containing the corresponding warning level and maintenance scheduling is generated; Based on the instruction information and degradation trajectory data from the historical operation of equipment of the same model, preventive maintenance intervention points are determined, and the instruction information is executed at the preventive maintenance intervention points.

2. The color masterbatch production apparatus failure detection method according to claim 1, characterized by, The process involves collecting multi-source sensor data during the operation of the color masterbatch production equipment, preprocessing the multi-source sensor data, and constructing a multi-dimensional data sequence, including: Acquire multi-source sensing data output from multiple sensors, wherein the multi-source sensing data includes vibration, temperature, pressure, and shear force signals; The multi-source sensor data are subjected to analog-to-digital conversion and digital filtering respectively to obtain denoised digital signals; Perform a high-precision timestamp alignment operation on the denoised digital signal to generate a synchronized multidimensional data stream; The multidimensional data stream is subjected to data integrity verification and missing value imputation to construct the multidimensional data sequence.

3. The color masterbatch production apparatus failure detection method according to claim 1, characterized by, The step of performing principal component analysis to reduce the dimensionality of the multidimensional data sequence and extracting a low-dimensional degradation vector characterizing the multi-factor coupled degradation trajectory of the device includes: The multidimensional data sequence is subjected to zero-mean standardization to obtain a standardized data matrix, and the covariance matrix between each dimension is calculated based on the standardized data matrix. The covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvectors and eigenvalues. The principal component directions are selected according to the eigenvalues ​​in descending order to form the projection transformation matrix. Multiply the standardized data matrix by the projection transformation matrix to obtain the principal component score matrix; The principal component score matrix is ​​reconstructed into a low-dimensional degenerate vector in chronological order.

4. The color masterbatch production apparatus failure detection method according to claim 1, characterized by, Based on the low-dimensional degradation vector, the time-domain statistical characteristics and frequency-domain energy distribution characteristics within the sliding time window are calculated and concatenated to form a time-frequency joint feature vector. The deviation of the time-frequency joint feature vector from the equipment health baseline is then calculated to generate a degradation index sequence that quantifies the early wear degree of the equipment, including: A series of time window slices are obtained by truncating the low-dimensional degenerate vector using a sliding window of fixed length. For each time window slice, calculate the statistical mean vector and the statistical variance vector; Perform a Fast Fourier Transform on the time window slices to extract the energy percentage within the preset frequency band and obtain the spectral energy feature vector; The statistical mean vector, the statistical variance vector, and the spectral energy feature vector are concatenated to form a time-frequency joint feature vector; Based on the data collected during the initial health operation phase of the equipment, the average value of the time-frequency joint feature vector calculated through the same process is used to predetermine the equipment health baseline vector; Calculate the Euclidean distance between the time-frequency joint feature vector and a predetermined device health baseline vector to generate a degradation index sequence.

5. The color masterbatch production apparatus failure detection method according to claim 1, characterized by, The degradation index sequence is subjected to a first-order difference operation to obtain a growth rate sequence. If the growth rate sequence exceeds a preset acceleration threshold, the equipment degradation is determined to have entered a nonlinear acceleration stage, and potential fault signals are extracted from it, including: The degradation index sequence is subjected to a first-order difference operation to obtain an instantaneous growth rate sequence, and the instantaneous growth rate sequence is mapped to the reconstructed phase space to form a dynamic evolution trajectory; Calculate the local tangent slope at each point on the dynamic evolution trajectory. If the local tangent slope remains constant... If the speed continues to exceed the preset acceleration threshold, the device is determined to have entered the nonlinear acceleration stage of degradation. Local maxima of the degradation index sequence during the nonlinear acceleration phase are extracted, an evolution envelope is constructed based on the local maxima, and the divergence of the evolution envelope is calculated. The signal segments whose divergence exceeds a preset fluctuation threshold are marked as potential fault signals.

6. The color masterbatch production apparatus failure detection method according to claim 1, characterized by, The step of analyzing the probability distribution vectors of the equipment belonging to different health states based on the potential fault signals and in conjunction with a preset support vector machine model includes: Extract time-domain statistical amplitude features from the potential fault signals to construct a multi-dimensional feature input vector; The multidimensional feature input vector is input into the support vector machine model to calculate the signed geometric distance from the multidimensional feature to the optimal classification hyperplane; The signed geometric distance is input into a preset probability calibration function to obtain the posterior probability that the potential fault signal belongs to various health states. The posterior probabilities output by multiple binary classifiers in the support vector machine model are aggregated to form a probability distribution vector.

7. The color masterbatch production apparatus failure detection method according to claim 1, characterized by, The process of determining the dominant degradation mode and mode confidence based on the probability distribution vector, retrieving the corresponding standard curve of degradation evolution based on the dominant degradation mode, and using the mode confidence to correct the standard curve and estimate the remaining effective operating time of the equipment includes: Identify the component with the largest value in the probability distribution vector, determine the health state corresponding to the component with the largest value as the dominant degradation mode, and use the value of the component with the largest value as the mode confidence. According to the dominant degradation mode, the corresponding standard curve of degradation evolution is retrieved from the pre-established standard curve library of degradation evolution generated based on the statistical analysis of historical failure data of similar equipment. The standard curve is scaled and corrected using the pattern confidence level to obtain a corrected degradation trajectory, and the remaining effective operating time of the device is estimated along the corrected degradation trajectory.

8. The color masterbatch production apparatus failure detection method according to claim 1, characterized by, If the remaining effective running time is less than a preset time threshold, then instruction information containing the corresponding warning level and maintenance scheduling is generated, including: When the remaining effective running time is less than the preset emergency time threshold, a high-risk warning level is determined; Based on the high-risk warning level and the current maintenance resource status, instruction information including specific maintenance time, spare parts list and personnel arrangement is generated.

9. The method for fault detection of color masterbatch production equipment according to claim 1, characterized in that, The step of determining preventative maintenance intervention points based on the instruction information and degradation trajectory data from the historical operation of equipment of the same model includes: Taking the planned maintenance time points and historical degradation trajectory data in the instruction information as input, perform time axis alignment and data fusion operations to output a multi-dimensional fused dataset. Calculate the deviation between the current degraded trajectory and historical similar trajectories in the multidimensional fusion dataset, and calculate the risk evolution rate based on the deviation. The remaining life prediction value of the equipment is dynamically updated based on the risk evolution rate. The updated remaining life forecast is matched with the resource gaps in the production plan to determine the optimal preventative maintenance intervention point.

10. A fault detection system for color masterbatch production equipment, characterized in that, include: The data processing module is used to collect multi-source sensor data during the operation of the color masterbatch production equipment, preprocess the multi-source sensor data, and construct a multi-dimensional data sequence. The feature dimensionality reduction module is used to perform principal component analysis on the multidimensional data sequence to reduce dimensionality and extract a low-dimensional degradation vector that characterizes the multi-factor coupled degradation trajectory of the device. The index generation module is used to calculate the time-domain statistical features and frequency-domain energy distribution features within the sliding time window based on the low-dimensional degradation vector, and splice them to form a time-frequency joint feature vector. It also calculates the deviation of the time-frequency joint feature vector from the equipment health benchmark to generate a degradation index sequence that quantifies the early wear degree of the equipment. The trend recognition module is used to perform a first-order difference operation on the degradation index sequence to obtain a growth rate sequence. If the growth rate sequence exceeds a preset acceleration threshold, it is determined that the equipment degradation has entered a nonlinear acceleration stage, and potential fault signals are extracted from it. The state classification module is used to analyze the probability distribution vector of the device belonging to different health states based on the potential fault signals and in combination with a preset support vector machine model. The pattern prediction module is used to determine the dominant degradation pattern and the pattern confidence based on the probability distribution vector, retrieve the corresponding degradation evolution standard curve based on the dominant degradation pattern, and use the pattern confidence to correct the standard curve to estimate the remaining effective operating time of the equipment. The instruction generation module is used to generate instruction information containing the corresponding warning level and maintenance schedule if the remaining effective running time is less than a preset time threshold. The intervention positioning module is used to determine preventive maintenance intervention points based on the instruction information and degradation trajectory data from the historical operation of equipment of the same model, and to execute the instruction information at the preventive maintenance intervention points.

Images