Communication equipment production intelligent management system based on machine learning

Abstract

The invention relates to the technical field of communication production management, and discloses a communication equipment production intelligent management system based on machine learning. The system comprises a production data acquisition module, a feature engineering construction module, a dynamic clustering analysis module, an anomaly detection engine module and a production decision optimization module. The production data acquisition module acquires multi-source sensor data in real time and converts the multi-source sensor data into a standardized sequence with a unified timestamp; the feature engineering module extracts a time domain statistical feature, a frequency domain energy feature and an equipment state association feature to generate a high-dimensional feature vector set; the dynamic clustering module adopts an incremental algorithm to divide clusters online; the anomaly detection module establishes a multi-level Gaussian mixture model based on a clustering label, and quantifies an anomaly probability through a mahalanobis distance; and the production decision module integrates the results to generate an equipment maintenance priority sequence and a production takt adjustment instruction. According to the system, intelligent monitoring and dynamic optimization of the whole production process of the communication equipment are realized, and the real-time change requirement of a complex production environment is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication production management, in particular to an intelligent management system for communication equipment production based on machine learning. BACKGROUND

[0002] As the core hardware carrier of information networks, the stability and refinement of the production process of communication equipment are directly related to the quality of downstream communication services. With the comprehensive promotion of 5G commercialization and the rapid popularization of Internet of Things technology, the production scale of communication equipment continues to expand, and the automation level and equipment complexity of the production line are also improving. The limitations of traditional production management modes are increasingly highlighted.

[0003] At the data processing level, existing production lines usually deploy multiple types of sensors covering temperature, vibration, pressure and other monitoring dimensions. The data streams generated by these sensors have problems such as structural heterogeneity, inconsistent sampling frequency, and chaotic timestamps. Traditional data collection methods mostly use single format conversion or offline integration methods, which are difficult to achieve real-time standardized processing of multi-source data, resulting in scattered data resources and inability to form a coherent analysis basis, which in turn affects the overall judgment of the production state.

[0004] There are obvious shortcomings in feature extraction. Existing technologies often focus only on single-dimensional feature analysis, or focus on time-domain statistical features such as mean and variance, or extract frequency energy features such as spectral peak values, ignoring the correlation between device operating states. This one-sided feature extraction method cannot fully reflect the actual operating conditions of the device, making the subsequent state recognition and abnormality judgment lack sufficient feature support, and prone to misjudgment or missed judgment.

[0005] In terms of clustering analysis, traditional algorithms mostly use offline batch processing mode. When the production line equipment operating state changes dynamically, the clustering model cannot be updated in time, resulting in a disconnection between the clustering cluster label and the actual device state, and the inability to provide accurate classification basis for anomaly detection. The anomaly detection technology generally relies on a single model architecture, which is not adaptable to the diverse types of anomalies in complex production environments, and lacks quantitative evaluation of the degree of abnormality, and can only achieve simple anomaly alarm, making it difficult to meet the needs of refined management.

[0006] The production decision-making link still relies more on human experience, and the formulation of equipment maintenance plans lacks scientific data-driven basis, often resulting in over-maintenance or inadequate maintenance; the adjustment of production rhythm is mostly post-repair, and cannot be dynamically optimized according to the real-time state of the equipment, not only causing waste of production resources, but also possibly causing the production line to stop due to equipment failure, seriously affecting production efficiency and product quality. These problems jointly restrict the intelligent upgrading of communication equipment production management, and a full-process intelligent management system needs to be built. SUMMARY

[0007] The present application aims to provide a machine learning-based intelligent management system for communication equipment production to solve the problems raised in the background.

[0008] To achieve the above-mentioned purpose, the present application provides a machine learning-based intelligent management system for communication equipment production, which comprises: A production data acquisition module for acquiring real-time raw data streams of various sensors on a communication equipment production line and converting multi-source heterogeneous data into standardized data sequences with unified timestamps; A feature engineering construction module for extracting time domain statistical features, frequency domain energy features and equipment state correlation features from the standardized data sequences to generate a high-dimensional feature vector set; A dynamic clustering analysis module for dividing the high-dimensional feature vector set online using an incremental clustering algorithm and generating dynamic clustering cluster labels according to the similarity of equipment operating states; An anomaly detection engine module for establishing a multi-level Gaussian mixture model based on the dynamic clustering cluster labels and realizing anomaly probability quantification by calculating Mahalanobis distances between real-time data and each level of the model; A production decision optimization module for integrating the probability distribution results output by the anomaly detection engine module to generate equipment maintenance priority ranking and production rhythm adjustment instructions.

[0009] Preferably, the production data acquisition module comprises: A data synchronization unit for resampling and aligning the sampling frequencies of vibration sensors, temperature sensors and current sensors; A data verification unit for detecting missing data segments and using adjacent equipment data correlation for interpolation compensation; A data buffering unit for maintaining a ring buffer to store complete data snapshots of the last N sampling periods.

[0010] Preferably, the feature engineering construction module performs operations including: A time domain window sliding unit for intercepting the standardized data sequences with a fixed time window and calculating the mean, variance and kurtosis index of the data within the window; A frequency domain transformation unit for extracting the energy proportion of the first K harmonic components after performing fast Fourier transform on the time domain window data; A correlation analysis unit for calculating the mutual information entropy value of different sensor data under time shift state.

[0011] Preferably, the incremental clustering algorithm used by the dynamic clustering analysis module includes: A cluster center drift detection unit for triggering cluster splitting when the Euclidean distance between a newly arrived high-dimensional feature vector and the nearest cluster center exceeds an adaptive threshold; The noise filtering unit removes temporary clusters whose duration is less than a set threshold based on historical clustering stability. The dimension compression unit performs local linear embedding to reduce the dimensionality of high-dimensional feature vectors and then visualizes the cluster boundaries.

[0012] Preferably, the multi-level Gaussian mixture model of the anomaly detection engine module includes: Device-level model: Constructs a probability density function for a single communication device that includes all clusters under normal operating conditions; Production line-level model: integrate the common features of equipment at the same workstation across the entire production line to establish a joint probability distribution; A global-level model is used to establish a cross-production line anomaly pattern feature library by integrating historical failure case data.

[0013] Preferably, the instruction generation logic of the production decision optimization module includes: The health rating unit calculates the percentage of the current equipment status that deviates from the standard operating condition based on the Mahalanobis distance. The fault prediction unit assesses the probability distribution curve of the remaining service life by combining the cumulative operating time of the equipment. The scheduling strategy unit dynamically calculates the optimal equipment maintenance time window based on the score values ​​and probability distribution curves.

[0014] Preferably, the system further includes: The digital twin mapping module is used to establish a 3D visualization interface to display the spatial distribution evolution of dynamic clusters in real time. The model iteration module is used to inject newly generated anomaly case features into the global model weekly for parameter fine-tuning. The audit trail module records all cluster splitting and merging events and their corresponding production batch information.

[0015] Preferably, the digital twin mapping module is implemented in the following ways: The heatmap rendering unit uses color gradients to represent the probability density of anomalies in different areas of the device. The topology analysis unit displays the potential impact range of fault propagation paths between devices; Virtual calibration units allow users to manually annotate abnormal areas in a 3D interface to assist model training.

[0016] Preferably, the parameter fine-tuning process performed by the model iteration module includes: Feature distillation unit extracts unique combinations of features from new cases that distinguish them from existing anomaly patterns; The weight decay unit gradually reduces the model's focus on historical fault features that have exceeded their validity period. The migration test unit verifies the generalization ability of the updated model on similar production lines.

[0017] Preferably, the data association method of the audit trail module includes: The event replay unit displays the entire lifecycle of a specific cluster from its creation to its demise along a timeline. The root cause tracing unit establishes a causal chain between the final equipment failure and the early-appearing abnormal clusters; The knowledge graph unit constructs a network topology that links fault modes with production process parameters.

[0018] Compared with the prior art, the beneficial effects of the present invention are: The production data acquisition module captures raw data streams from various sensors on the production line in real time and converts multi-source heterogeneous data into standardized data sequences with a unified timestamp. This effectively solves the problems of inconsistent formats and time synchronization in traditional data acquisition processes, enabling scattered monitoring data to form a consistent analytical foundation. Subsequent feature extraction and status analysis can then be conducted based on a unified data standard. This standardized processing mode is well-suited to the characteristics of multi-sensor collaborative work in modern production lines, achieving efficient integration and reuse of data resources.

[0019] The feature engineering construction module simultaneously extracts time-domain statistical features, frequency-domain energy features, and equipment state correlation features, generating a high-dimensional feature vector set, breaking through the limitations of traditional single-dimensional feature extraction. Time-domain features can reflect the real-time dynamic changes in equipment operation, frequency-domain features can capture the periodic patterns of equipment vibration, energy consumption, etc., while equipment state correlation features can reflect the mutual influence relationships between different devices. The combination of the three enables the feature vectors to more comprehensively and accurately map the actual operating status of the equipment, providing richer information support for subsequent clustering analysis and anomaly detection.

[0020] The dynamic clustering analysis module employs an incremental clustering algorithm to partition high-dimensional feature vector sets online, enabling real-time responses to changes in equipment operating status and timely updates to cluster labels. Compared to traditional offline clustering algorithms, this dynamic adjustment mode ensures a high degree of matching between clustering results and current production conditions, allowing equipment in different operating states to be accurately classified. This provides a realistic classification benchmark for anomaly detection and avoids analytical biases caused by lagging clustering results.

[0021] The anomaly detection engine module establishes a multi-level Gaussian mixture model based on dynamic clustering labels. It quantifies anomaly probabilities by calculating the Mahalanobis distance between real-time data and each level of the model, significantly improving the accuracy and detail of anomaly identification. The multi-level model architecture can adapt to different types and degrees of anomalies, while the application of Mahalanobis distance fully considers the correlation between features, effectively reducing the risk of false positives. The quantified anomaly probability clearly reflects the severity of the anomaly, providing specific and quantifiable reference information for production decisions and facilitating the rapid identification of key issues.

[0022] The production decision optimization module integrates the probability distribution results of anomaly detection output to generate equipment maintenance priority ranking and production cycle adjustment instructions, breaking away from the traditional decision-making model that relies on human experience. Maintenance priority ranking enables limited maintenance resources to be allocated to high-risk equipment, avoiding resource waste and proactively mitigating potential failures; dynamic adjustment of production cycle ensures that the production process is adapted to the equipment operating status, reducing waiting time or equipment overload during production, and achieving dynamic optimization of production efficiency. Attached Figure Description

[0023] Figure 1 A diagram showing the relationship between equipment health and maintenance priority; Figure 2 A flowchart for building modules for feature engineering; Figure 3 This is a flowchart of the dynamic clustering analysis module; Figure 4 This is a graph showing the relationship between cluster stability and radius variation. Detailed Implementation

[0024] 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.

[0025] Please see Figure 1This invention provides a machine learning-based intelligent management system for communication equipment production. The system includes a production data acquisition module responsible for real-time acquisition of raw data streams from devices such as vibration sensors, temperature sensors, and current sensors deployed on the communication equipment production line. This data has multi-source heterogeneous characteristics. The production data acquisition module converts the raw data into a standardized data sequence with a unified timestamp through data synchronization, data verification, and data buffering. A feature engineering construction module receives the standardized data sequence and extracts time-domain statistical features, frequency-domain energy features, and equipment status correlation features through time-domain window sliding, frequency-domain transformation, and correlation analysis operations, thereby generating a high-dimensional feature vector set. A dynamic clustering analysis module uses an incremental clustering algorithm to partition the high-dimensional feature vector set online. The core of the algorithm includes cluster center drift detection, noise filtering, and dimensionality compression mechanisms, generating dynamic cluster labels based on the similarity of equipment operating states. An anomaly detection engine module establishes a multi-level Gaussian mixture model based on the dynamic cluster labels. The model hierarchy covers equipment-level models, production line-level models, and global-level models. Anomaly probability quantification is achieved by calculating the Mahalanobis distance between real-time data and each level of the model. The production decision optimization module integrates the probability distribution results output by the anomaly detection engine module, and uses health scores, fault prediction, and scheduling strategy logic to generate equipment maintenance priority ranking and production cycle adjustment instructions. The entire system achieves a closed-loop process from data acquisition to decision execution through modular design, ensuring intelligent monitoring and optimization of the communication equipment production process.

[0026] Example 1: See Figure 2The data synchronization unit of the production data acquisition module is responsible for processing the multi-source heterogeneous data streams generated by vibration sensors, temperature sensors, and current sensors. The sampling frequency of the vibration sensor is set to 10kHz, the data update frequency of the temperature sensor is set to 100Hz, and the current sensor outputs readings at a fixed period of 1kHz. The data synchronization unit uses a resampling algorithm to unify the data streams of all sensors onto a 10kHz timestamp sequence. During the resampling process, linear interpolation technology is used to fill in the intermediate sampling points for the data from the temperature sensor and the current sensor, so that data from different physical sources can be aligned and correlated at the same point in time. After data synchronization, the data verification unit performs an integrity check. It scans the standardized data sequence to identify missing data segments caused by transmission errors or sensor malfunctions. The identification of missing data segments is based on the continuity of timestamps and the reasonableness of data values. Timestamp continuity requires the time interval between adjacent data points to be strictly equal to 0.1 milliseconds. The reasonableness check of data values ​​excludes abnormal readings exceeding the physical range. Once data loss is detected, the data verification unit initiates an interpolation compensation program. This program uses the correlation between sensor data from adjacent devices to establish a linear regression model, calculating an approximate value for the missing location using known data. This spatial correlation-based compensation method effectively restores the integrity of the data stream. The data buffer unit intervenes after data verification. It maintains a fixed-size circular buffer that stores complete data snapshots of the most recent 30,000 sampling periods in a first-in, first-out (FIFO) manner. The size of the buffer is dynamically set based on the maximum data generation rate of the production line and the throughput capacity of subsequent processing modules. The circular buffer design allows the system to remain stable under high-throughput data inflow while providing a continuous data window for the feature engineering module.

[0027] The time-domain window sliding unit of the feature engineering construction module obtains standardized data sequences from the data buffer unit. It extracts the data stream with a fixed-length time window, set to 1000 sampling points corresponding to 100 milliseconds. The window advances in a sliding step of 500 sampling points to achieve continuous data segmentation. For each window, the time-domain window sliding unit calculates the mean, variance, and kurtosis. The mean reflects the average level of the data within the window, the variance characterizes the dispersion of the data, and the kurtosis describes the difference between the data distribution and the normal distribution. These three time-domain statistics together constitute the basic feature set describing the short-term operating state of the equipment. The frequency-domain transformation unit receives the windowed data output from the time-domain window sliding unit. It applies a Fast Fourier Transform (FFT) algorithm to each data window, converting the time-domain signal into a frequency-domain representation, revealing the frequency composition characteristics of the signal. The frequency domain transformation unit extracts the energy proportion of the top 20 harmonic components from the transformation result. The energy proportion is calculated as the ratio of the energy of each harmonic component to the total energy. These frequency domain features can effectively capture the periodic vibration modes and energy distribution changes during equipment operation. The correlation analysis unit processes data from multiple sensors simultaneously. The correlation analysis unit calculates the mutual information entropy values ​​of different sensor signals under time-shifted conditions. The time-shift range covers the possible delay in equipment action from -50 milliseconds to +50 milliseconds. The mutual information entropy value quantifies the statistical dependence between two sensor signals. The correlation analysis unit determines the optimal time-shift relationship between sensors by finding the maximum mutual information entropy value. This cross-sensor correlation feature can discover the interactions and influences between internal components of the equipment.

[0028] In the data synchronization unit implementation of the production data acquisition module, the resampling algorithm uses linear interpolation to unify sensor data of different frequencies to the highest sampling frequency. Linear interpolation calculates the interpolated value between adjacent sampling points according to the time ratio, maintaining the basic shape of the data waveform. The data synchronization unit adds a unified timestamp to each sensor data stream, with timestamp accuracy reaching the millisecond level, meeting the time sequence analysis requirements of high-speed production lines. The interpolation compensation program of the data verification unit adopts a regression model based on the K-nearest neighbor algorithm. The K-nearest neighbor algorithm finds the 10 most similar patterns to the adjacent data of the missing segment in historical data, and uses the statistical characteristics of these patterns to generate compensation values. The interpolation compensation program specifically handles the case of continuous missing points, and improves the compensation accuracy of long missing segments by establishing an autoregressive model. The circular buffer of the data buffer unit is implemented using a circular queue data structure. The read and write pointers of the circular queue mark the data storage and retrieval positions respectively. When the buffer is full, it automatically overwrites the oldest data. The data buffer unit also provides a random access interface for data snapshots, supporting retrospective analysis of the feature engineering construction module.

[0029] The time-domain window sliding unit of the feature engineering construction module adopts an overlapping window sliding strategy, with 50% overlap between adjacent windows. This overlapping sliding ensures continuous capture of device operating status and avoids missing important features due to window segmentation. The variance value calculated by the time-domain window sliding unit is unbiased corrected to adapt to the statistical characteristics of small sample windows. The kurtosis index is calculated using the ratio of the fourth-order central moment to the square of the variance, accurately reflecting the peak degree of the data distribution. The fast Fourier transform of the frequency domain transform unit is implemented using the Cooley-Tukey algorithm. Before the transform, a Hanning window function is applied to the data window to reduce spectral leakage. The Hanning window function suppresses discontinuities at window edges and improves frequency resolution accuracy. The harmonic component energy proportions extracted by the frequency domain transform unit are arranged in descending order, focusing on the main frequency components of the signal. The energy proportion calculation uses the sum of squares of the amplitudes as the energy measure, which conforms to the physical definition of energy. The mutual information entropy value of the correlation analysis unit is calculated based on histogram estimation of the joint probability distribution. The number of histogram intervals is automatically adjusted according to the amount of data. The mutual information entropy value is applicable to both discrete and continuous variables. The correlation analysis unit finds the peak value of mutual information entropy by scanning different time shifts. The time shift corresponding to the peak value reflects the actual propagation delay between sensor signals.

[0030] The collaborative operation of the production data acquisition module and the feature engineering construction module is built on a strict data pipeline, maintaining temporal consistency from data acquisition to feature extraction. The data synchronization unit of the production data acquisition module outputs a multi-sensor data stream with a unified timestamp. This data stream enters the data verification unit for quality checks, and qualified data is temporarily stored in the data buffer unit. The time-domain window sliding unit of the feature engineering construction module retrieves the latest data window from the buffer, and time-domain feature calculation and frequency-domain transformation are performed in parallel. The correlation analysis unit simultaneously acquires data from multiple sensors and calculates mutual information entropy. The entire process adopts a pipelined processing mode, with data flowing through different processing stages. Each stage adds corresponding feature labels, and the final output high-dimensional feature vector set includes time-domain statistical features, frequency-domain energy features, and equipment status correlation features, providing standardized input for the dynamic clustering analysis module. The dimension of the feature vector is determined by the number of time-domain features, frequency-domain features, and correlation features. Each feature dimension undergoes normalization to eliminate the impact of dimensional differences on subsequent clustering algorithms. The resampling accuracy of the data synchronization unit directly affects the quality of subsequent feature extraction; the interpolation compensation algorithm of the data verification unit depends on the spatial correlation model between devices; and the capacity configuration of the circular buffer needs to balance memory usage with the requirement to retain historical data. Example 2: See Figure 3The dynamic clustering analysis module employs an incremental clustering algorithm, which includes a cluster center drift detection unit, a noise filtering unit, and a dimensionality compression unit. The cluster center drift detection unit monitors the Euclidean distance between newly arriving high-dimensional feature vectors and existing cluster centers. The calculation process uses vectorized operations; the distance matrix is ​​calculated simultaneously between each new feature vector and all existing cluster centers. When the minimum Euclidean distance exceeds 1.5 times the current cluster radius, a cluster splitting mechanism is triggered. The cluster splitting mechanism creates new cluster centers and reallocates adjacent feature vectors. The initial position of the new cluster center is set as the coordinates of the feature vector that triggered the split. The reallocation of adjacent feature vectors is based on the nearest neighbor principle, and the cluster radius is dynamically updated according to the distribution density of feature vectors within the cluster. The noise filtering unit analyzes the stability index of the clustering results. The stability index is obtained by calculating the rate of change of the cluster structure within a continuous time window, using ten consecutive sampling periods as the time window. When the lifespan of a cluster is detected to be lower than a preset threshold, the noise filtering unit marks the detected cluster as a temporary cluster. The identification of temporary clusters relies on the duration statistics and the trend of changes in the number of members. The noise filtering unit maintains a list of temporary clusters. If a cluster in the list has not been reallocated after three complete production cycles, it is removed. The removal operation releases the storage resources occupied by the temporary clusters and updates the clustering topology.

[0031] The dimensionality compression unit performs local linear embedding (LSE) to reduce the dimensionality of high-dimensional feature vectors. LSE achieves dimensionality compression by preserving the geometric relationships of feature vectors within their local neighborhoods. The LSE dimensionality reduction process consists of three steps: neighborhood construction, weight calculation, and low-dimensional embedding. Neighborhood construction uses the k-nearest neighbor algorithm to find the nearest neighbor of each feature vector. Weight calculation is obtained by minimizing the local linear reconstruction error. Low-dimensional embedding maps the high-dimensional data to a two- or three-dimensional space. The dimensionality-reduced feature vectors retain the original data's topological structure. The dimensionality compression unit passes the low-dimensional coordinates to the visualization interface, which generates a dynamic clustering boundary map. The cluster center drift detection unit's adaptive threshold adjustment mechanism is based on historical clustering quality assessment results. The assessment metrics include intra-cluster compactness and inter-cluster separation. Intra-cluster compactness is measured by calculating the average distance from the intra-cluster feature vector to the cluster center, while inter-cluster separation is characterized by the minimum distance between different cluster centers. The threshold adjustment process employs a smooth transition strategy to avoid drastic fluctuations, with each adjustment not exceeding 10% of the current value. Cluster split events are recorded with timestamps, trigger vector identifiers, and cluster center coordinates before and after the split. This information is sent to the audit trail module for subsequent analysis. The temporary cluster determination algorithm of the noise filtering unit considers both duration and stability factors. The duration threshold is set at ten minutes, corresponding to sixty sampling periods. The stability factor is evaluated by checking the coefficient of variation of the number of cluster members. Clusters with a coefficient of variation exceeding 0.5 are determined to be unstable clusters. Unstable clusters are retained for observation even if their survival time exceeds the threshold. The noise filtering unit periodically generates cluster stability reports, which include the survival time and stability indicators of all currently active clusters.

[0032] The local linear embedding algorithm of the dimensionality compression unit employs an improved neighborhood selection strategy. The neighborhood size is adaptively determined based on the local density of the feature vectors, with smaller neighborhoods in high-density regions and larger neighborhoods in low-density regions. Low-dimensional coordinates are updated every 30 seconds, synchronized with the clustering analysis cycle. The visualization interface uses different colors to render each cluster region, with the color mapping dynamically adjusted based on cluster creation time and stability indicators. The data processing pipeline of the dynamic clustering analysis module begins with feature vector standardization, scaling the feature values ​​to zero mean and unit variance. The standardized feature vectors then enter the cluster center matching stage. The matching algorithm uses approximate nearest neighbor search to improve computational efficiency, accelerating distance calculation by constructing a ball tree data structure. Successfully matched feature vectors update the statistical information of the corresponding clusters, including the moving average of the cluster center coordinates and recalculation of the cluster radius.

[0033] A data interaction mechanism exists between the cluster center drift detection unit and the noise filtering unit. The cluster center drift detection unit synchronizes newly created cluster information to the noise filtering unit in real time. The noise filtering unit decides whether to include a new cluster in the stability monitoring scope based on its survival time. The dimensionality compression unit obtains the latest cluster center coordinates from the cluster center drift detection unit and a list of valid clusters from the noise filtering unit. The dimensionality compression unit only performs dimensionality reduction visualization on valid clusters. The memory management of the dynamic clustering analysis module adopts a block allocation strategy. Each cluster is allocated an independent memory block to store member feature vectors, and the memory block size is dynamically adjusted according to the cluster size. When the number of members in a cluster exceeds a threshold, a memory expansion operation is automatically triggered, with the expansion amount being twice the current size. The allocation of computing resources is dynamically adjusted based on the clustering complexity, which is quantified by the product of the current number of active clusters and the feature vector dimension. The output of the incremental clustering algorithm is a dynamic cluster label. The label generation process uses a hierarchical encoding scheme, and the encoding includes cluster type identifier, creation timestamp, and status level information. Cluster type identifiers distinguish between normal operating, transitional, and abnormal states of the equipment. A timestamp records the cluster's creation time, and the state level is divided into three levels based on the cluster's stability and size. Dynamic clustering labels are encapsulated into a standard data structure, which includes label encoding, cluster center coordinates, and statistical characteristics.

[0034] The performance optimization of the dynamic clustering analysis module includes computation caching and incremental update strategies. The computation cache stores the most recently used distance matrix results, and the incremental update strategy avoids recalculating unchanged cluster relationships. The cache size is dynamically adjusted based on available memory, and the least recently used algorithm is used for cache replacement. Incremental updates use a version number mechanism to identify data changes. Message transmission employs an acknowledgment mechanism to ensure data reliability. The anomaly detection engine module returns an acknowledgment signal upon receiving a message, and unacknowledged data is automatically retransmitted after a timeout. The collaborative work of these three units forms a complete incremental clustering analysis loop. Each newly arrived feature vector triggers the execution of this loop, and the loop cycle is consistent with the output frequency of the feature engineering construction module. The cluster center drift detection unit processes the matching relationship between feature vectors and existing clusters, the noise filtering unit maintains the stability of the clustering results, and the dimensionality compression unit provides visualization support.

[0035] See Figure 4This study presents the changing trends of core performance indicators of the incremental clustering algorithm in real-time data processing. Specifically, the average stability index is obtained by continuously calculating the rate of change of the cluster structure within a sliding time window using the noise filtering unit. This rate of change is based on the coefficient of variation of the number of cluster members over a window of ten sampling periods. The average radius is dynamically updated by the cluster center drift detection unit, calculated using the moving average of the Euclidean distance matrix based on the distribution density of feature vectors within the cluster. During parameter configuration, the stability threshold is set to a coefficient of variation of 0.5, the adaptive radius adjustment is limited to 10% of the current value, and the noise filtering lifespan threshold is 60 sampling periods. The blue line in the figure represents the initial high stability, corresponding to the stable cluster structure during the initialization stage of clustering. The subsequent decrease in fluctuation reflects the generation process of new clusters after the cluster splitting mechanism is triggered. The red line shows the rapid increase in radius, which is consistent with the triggering condition of the distance threshold (1.5 times the cluster radius) of the cluster center drift detection unit. The later convergence of the two indices verifies the cluster boundary optimization achieved by the dimensionality compression unit through local linear embedding. The entire evolution process reflects the noise filtering unit's efficiency in eliminating temporary clusters and the adaptive threshold adjustment mechanism of the cluster center drift detection unit in the incremental clustering algorithm.

[0036] Example 3: The multi-level Gaussian mixture model of the anomaly detection engine module includes a device-level model, a production line-level model, and a global-level model. The device-level model constructs a probability density function for a single communication device, encompassing all clusters under normal operating conditions. This probability density function characterizes the data distribution characteristics of the device under different operating modes through a weighted combination of multiple Gaussian distributions. The parameters of each Gaussian distribution include a mean vector and a covariance matrix. The production line-level model integrates the common features of devices at the same workstation across the entire production line to establish a joint probability distribution. This joint probability distribution uses a multivariate Gaussian mixture model to describe the correlation between multiple devices, and the model parameters are learned from historical normal data using an expectation-maximization algorithm. The global-level model integrates historical fault case data to establish a cross-production line anomaly pattern feature library. This library stores fault feature vectors selected through feature selection and their corresponding weight coefficients.

[0037] The instruction generation logic of the production decision optimization module includes a health rating unit, a fault prediction unit, and a scheduling strategy unit. The health rating unit calculates the percentage deviation of the current equipment state from the standard operating condition based on Mahalanobis distance. The formula for calculating Mahalanobis distance is: ; in: Represents Mahalanobis distance, Represents the current device feature vector. This represents the feature mean vector under normal conditions. To represent the transpose of a matrix, It is the covariance matrix under normal conditions. This represents the inverse matrix. The health rating unit maps Mahalanobis distance to a percentage range of 0 to 100%, using a piecewise linear transformation function. The fault prediction unit assesses the probability distribution curve of remaining service life based on the equipment's cumulative operating time. This probability distribution curve is constructed based on the Wiener process degradation model, and the model parameters are learned from historical fault data using the maximum likelihood estimation method. The scheduling strategy unit dynamically calculates the optimal equipment maintenance time window based on the health rating values ​​and the probability distribution curve. The calculation of the maintenance time window considers the balance between production plan constraints and equipment failure risks.

[0038] The equipment-level model training process uses normal-state cluster labels output by the dynamic clustering analysis module, with each cluster corresponding to a Gaussian distribution component. The Gaussian mixture model parameter learning employs an iterative optimization algorithm, alternately executing the expectation step and the maximization step until the model converges. The expectation step calculates the posterior probability of each feature vector belonging to each Gaussian distribution, and the maximization step re-estimates the model parameters based on the posterior probability. The equipment-level model is updated weekly, incorporating the most recent seven days of normal operation data to ensure the model can track the slow degradation of equipment performance. Regularization is implemented during model training, with regularization parameters determined through cross-validation to avoid overfitting. The production line-level model requires collecting synchronous operation data from multiple devices at the same workstation, with the data collection period covering the complete work cycle of the devices. The covariance matrix of the joint probability distribution model contains information on the correlation between devices, with off-diagonal elements quantifying the interdependence of different device operating states. The production line-level model training uses a distributed computing framework, distributing data across multiple computing nodes for parallel processing, and the model update frequency is synchronized with the equipment-level model. An outlier detection mechanism is implemented during model training. Outlier detection is based on the three-standard-deviation principle of Mahalanobis distance; outliers exceeding this range are automatically excluded from the training data. A version management mechanism is implemented for the global model's anomaly pattern feature library. Each anomaly pattern feature record includes a feature vector, fault type, occurrence time, and severity label. Updates to the feature library require a quality verification process; newly added anomaly pattern features must originate from confirmed fault events, and duplicate records are avoided through similarity checks. Feature library retrieval employs a nearest neighbor search algorithm based on cosine similarity, and an inverted index is built during the search process to accelerate query speed.

[0039] The Mahalanobis distance calculation in the health rating unit is regularized by adding small positive numbers to the main diagonal of the covariance matrix to avoid matrix singularity. A threshold mechanism is used for percentage value conversion; when the Mahalanobis distance exceeds three standard deviations of the normal distribution, the health rating is set to zero. The health rating unit outputs a rating result every five minutes, accompanied by a confidence interval estimate, reflecting the reliability of the rating calculation. The rating results are stored in a time-series database with multi-level indexes to support fast historical data retrieval. The Wiener process degradation model in the fault prediction unit includes two key parameters: the drift coefficient and the diffusion coefficient. The drift coefficient characterizes the average rate of equipment performance degradation, while the diffusion coefficient describes the random fluctuation amplitude of the degradation process. The probability distribution curve is updated based on the current equipment health rating, showing the probability that the equipment will still function normally at different future time points. The probability distribution curve is generated using Monte Carlo simulation, with 10,000 simulations performed to ensure statistical significance.

[0040] The scheduling strategy unit employs a stochastic optimization algorithm to solve for the maintenance time window. The algorithm's objective function minimizes the sum of expected production loss and maintenance cost. The output of the maintenance time window includes the specific maintenance time point, estimated maintenance duration, and maintenance priority score. The priority score is calculated based on a comprehensive assessment of the equipment health decline rate and the severity of the failure consequences. The scheduling strategy unit establishes a data interface with the production planning system, transmitting equipment availability predictions and maintenance demand information, which the production planning system uses to adjust production schedules. The scheduling algorithm employs a genetic algorithm with a population size of 100, 500 iterations, a crossover probability of 0.8, and a mutation probability of 0.1. The anomaly detection engine module implements a collaborative reasoning mechanism across its three-tiered models: the equipment-level model provides fine-grained anomaly detection capabilities, the production line-level model identifies systemic anomaly patterns, and the global-level model detects rare but severe fault types. When the detection results of the three models diverge, the system uses a weighted voting mechanism to make a final judgment. The weight coefficients are dynamically adjusted based on the model's historical accuracy. The weights of the weighted voting mechanism are updated every 24 hours, taking into account the accuracy performance of the model's most recent 100 predictions.

[0041] The production decision optimization module incorporates multi-objective optimization considerations in its decision-making process, simultaneously optimizing equipment reliability, production efficiency, and maintenance costs. Data exchange between the health scoring unit and the fault prediction unit utilizes streaming processing. The fault prediction unit receives the health scoring sequence in real time, and this sequence data is used to update the degradation model parameters. Data stream processing employs windowed management, with a window size of 100 data points and a window sliding step of 1 data point. Integration between the anomaly detection engine module and the production decision optimization module is achieved through a message middleware. The message format uses the JSON standard to encapsulate detection results and decision instructions. Message transmission guarantees at least-once delivery semantics, ensuring no loss of critical data. Version control strategies are defined for inter-module interfaces, and bidirectional compatibility checks are implemented during version changes to ensure a smooth transition during system upgrades. The message middleware's topic partitions are hash-assigned based on device IDs, ensuring that data from the same device is always processed by the same consumer.

[0042] The inference process of the equipment-level model is optimized computationally, including Cholesky decomposition of the covariance matrix and pre-calculation of the determinant. The production line-level model employs dimensionality reduction techniques to decrease computational complexity; principal component analysis is used to preserve 95% of the data variance. Global-level model queries are accelerated using indexes; the index structure is built based on hash values ​​of feature vectors, and similarity search uses an approximate nearest neighbor algorithm to improve response speed. The health scoring unit considers real-time requirements; scoring calculations utilize the parallelization of matrix operations, achieving millisecond-level response through GPU acceleration. The fault prediction unit uses the fourth-order Runge-Kutta method to solve stochastic differential equations, ensuring the numerical stability of the probability distribution curve. Step size control employs an adaptive strategy, with an error tolerance set to 10^-6. The scheduling strategy unit's optimization algorithm implements a heuristic pruning strategy, eliminating obviously inferior solution space regions to improve search efficiency. The pruning threshold is dynamically adjusted based on the quality of historical best solutions, with the adjustment not exceeding 5% of the current threshold. The anomaly detection engine module employs a progressive learning strategy for model updates, incrementally incorporating new data into the existing model to avoid the computational overhead of complete retraining. Progressive learning utilizes a sliding window mechanism with a window size of 1000 samples; as a new sample enters the window, the oldest sample is removed. The production decision optimization module implements a flexible adjustment mechanism, automatically adapting decision parameters to new operating conditions when the production environment changes. Parameter adjustment uses gradient descent with a learning rate of 0.01 and 1000 iterations.

[0043] Example 4: The digital twin mapping module establishes a 3D visualization interface to display the spatial distribution evolution of dynamic clusters in real time. The 3D visualization interface is built based on the actual layout of the physical production line, including equipment geometric models, conveyor belt movement trajectories, and material flow paths. The heatmap rendering unit uses color gradients to represent the anomaly probability density of equipment in different areas. The color mapping adopts a continuous spectrum from dark green to dark red, with dark green indicating a safe state with an anomaly probability below 5%, and dark red marking high-risk areas with an anomaly probability exceeding 80%. The topology graph analysis unit displays the potential impact range of fault propagation paths between equipment. The fault propagation model is built based on equipment connection relationships and historical fault data, and the analysis results are superimposed on the 3D scene in the form of a directed graph. The virtual calibration unit provides interactive annotation tools, allowing users to manually delineate abnormal areas and add text annotations on the 3D interface. The annotation data is automatically synchronized to the training dataset of the model iteration module.

[0044] The model iteration module injects newly generated anomaly case features into the global model weekly for parameter fine-tuning. This fine-tuning employs a mini-batch gradient descent algorithm with a learning rate of 0.001 and a fixed batch size of 32 samples. The feature distillation unit extracts unique feature combinations from new cases that distinguish them from existing anomaly patterns. Feature distillation uses an attention mechanism to calculate feature importance weights; new features with weights exceeding a threshold are retained for model updates. The weight decay unit gradually reduces the model's attention to historical fault features that have exceeded their expiration date. The decay coefficient decreases exponentially based on the feature's age, with a half-life set to 90 days. The transfer testing unit verifies the updated model's generalization ability on similar production lines. The test dataset contains the most recent 30 days' runtime data from three similar production lines, and performance metrics include precision, recall, and F1 score. The audit trail module records all cluster splitting and merging events and corresponding production batch information. Event records are stored in a structured format, including timestamps, device numbers, cluster identifiers, event types, and production batch numbers. The event replay unit displays the entire lifecycle of a specific cluster from its creation to its demise along a timeline. The timeline interface supports zooming and dragging, with a minimum time precision of milliseconds. The root cause tracing unit establishes a causal chain between the final equipment failure and early-appearing anomalous clusters. Causal analysis employs the Granger causality test, with a significance level set at 0.05. The knowledge graph unit constructs a network linking failure modes and production process parameters. Network nodes represent equipment components, failure types, and process parameters, while edges represent the strength of the association between nodes.

[0045] The heatmap rendering unit of the digital twin mapping module implements a multi-level detail display mechanism, automatically adjusting the rendering granularity of the heatmap as the user zooms in. A long-distance view displays an overview of the anomaly probability distribution across the entire production line, while a close-up view displays detailed status information for individual devices. The heatmap rendering unit uses WebGL technology for GPU-accelerated rendering, with vertex shaders handling position coordinates and fragment shaders calculating color interpolation. Color interpolation is performed in the HSL color space to ensure smooth and natural color transitions. The topology analysis unit uses a force-directed algorithm to automatically place device nodes, with the repulsion coefficient set to -100, the attraction coefficient to 0.1, and the iteration count set to 1000. The virtual calibration unit's annotation tool offers both brush and lasso selection options. The brush radius is adjustable from 5 to 50 pixels, and the lasso tool supports continuous polygon selection. Annotated data uses a differential synchronization mechanism, transmitting only changed portions to reduce network bandwidth consumption. The feature distillation unit of the model iteration module performs feature importance evaluation using a permutation importance algorithm. The importance ranking algorithm observes changes in model prediction accuracy by randomly shuffling the values ​​of individual features; features with a more significant drop in accuracy are considered more important. The weight decay unit uses an exponential decay formula with a daily decay coefficient of 0.992 to ensure a smooth reduction in the influence of historical features. The performance evaluation of the transfer testing unit employs a five-fold cross-validation strategy, evenly dividing the test data into five mutually exclusive subsets. Four subsets are used alternately to train one subset for testing, and the average of the five test results is taken as the final performance metric. Version control for the model iteration module uses semantic version numbers: major version numbers increment when interfaces become incompatible, minor version numbers increment when new features are added, and revision numbers increment when issues are fixed.

[0046] The audit trail module uses a time-series database for event storage, which manages event records by time partitioning, with each partition storing one week's worth of data. The event playback unit generates video using keyframe technology, with a keyframe interval of 10 seconds, and non-keyframes generated using linear interpolation. The root cause analysis in the root cause tracing unit considers the time lag effect, with the lag time range set from 0 to 60 minutes, and scanning and detection in 5-minute increments. The knowledge graph unit's graph database uses an attribute graph model, where nodes and edges can be appended with multiple attribute key-value pairs, supporting complex graph traversal queries. The knowledge graph unit's query language uses Cypher statements, supporting pattern matching and pathfinding operations. The digital twin mapping module and the model iteration module establish a bidirectional data flow; the digital twin mapping module pushes user-annotated data to the model iteration module in real time, and the model iteration module feeds back updated model parameters to the digital twin mapping module. The audit trail module provides historical data query services for both modules, with the query interface supporting filtering by multiple conditions such as time range, device number, and event type. The three modules share a unified identity authentication system, and user operation records detail the operation time, content, and user identity information. The digital twin mapping module uses WebGL technology for real-time data rendering, achieving a 3D scene update frequency of 60 frames per second. The model iteration module employs a parameter server architecture for distributed training, with worker nodes responsible for calculating gradients and the parameter server aggregating gradients and updating the model. The audit trail module utilizes a multi-level caching strategy for query optimization: hot data is stored in an in-memory database, warm data in solid-state drives (SSDs), and cold data in hard disk drives (HDDs).

[0047] The system implements a monitoring and alarm mechanism. When the rendering latency of the digital twin mapping module exceeds 100 milliseconds, the training loss of the model iteration module fails to decrease after three consecutive iterations, or the query response time of the audit trail module exceeds 1 second, the system automatically triggers an alarm to notify operations and maintenance personnel. Alarm information is sent via both email and SMS, including the alarm level, alarm content, and suggested handling measures. Version management for the three modules adopts a branching strategy: the development branch is used for new feature development, the testing branch for feature verification, and the production branch for formal deployment. An automated test suite is executed before each version upgrade, with test case coverage required to reach over 90%. Data backup implements a daily incremental backup and weekly full backup strategy, with backup data retained for one year. Backup files are encrypted and stored in a multi-region object storage service. Inter-module communication uses the gRPC framework, and interface definitions use the Protocol Buffers description language. Service discovery is implemented using Consul, supporting health checks and service circuit breakers. Log collection uses the ELK technology stack; log data is cleaned and enriched before being stored in an Elasticsearch cluster, with a visual query interface provided through Kibana. Performance monitoring uses Prometheus to collect metric data, while Grafana provides custom dashboard functionality. The digital twin mapping module supports both orthographic and perspective projection modes for view projection. Orthographic projection maintains the object's size, while perspective projection simulates human visual perception. The model iteration module uses time-series cross-validation for model validation, maintaining the chronological order of the training and test sets to prevent future information leaks. The audit trail module automatically compresses data during archiving using the Zstandard compression algorithm, with a compression level of 3 to balance compression ratio and computational cost. Table 1 illustrates the correspondence between heatmap color coding and anomaly probabilities in the digital twin mapping module.

[0048] 1: Heatmap Color Coding Table

[0049] The digital twin mapping module uses a linear interpolation algorithm for color rendering, ensuring smooth transitions between adjacent color ranges. The model iteration module's version log includes model hash values, training data summaries, and performance metric snapshots. The audit trail module's event index uses an inverted index structure, supporting full-text search and fuzzy queries. System logs for all three modules are structured, with fields including request ID, timestamp, log level, module name, and detailed message. System security implements role-based access control, with different user roles having different data access and operation permissions, following the principle of least privilege. The digital twin mapping module supports smooth transition animations for perspective switching, with an animation duration set to 300 milliseconds to avoid visual jarring. The model iteration module's incremental training implements gradient pruning, with a gradient norm threshold set to 1.0 to prevent gradient explosion. The audit trail module's transaction processing uses an optimistic locking mechanism, automatically retrying three times in case of version conflicts to ensure data consistency.

[0050] Example 5: The parameter fine-tuning process executed by the model iteration module includes a feature distillation unit, a weight decay unit, and a transfer testing unit. The feature distillation unit extracts unique feature combinations from new cases that distinguish them from existing abnormal patterns, and uses a multi-head self-attention mechanism to calculate feature importance weights. The feature encoding layer maps the original feature vectors to a 512-dimensional space, and the attention calculation layer contains eight attention heads, each focusing on a different subspace of the feature space. The feature selection layer sets a weight threshold of 0.8, a stability threshold of 0.7, and a discriminative threshold of 0.6; features that simultaneously meet all three conditions enter the feature fusion stage. The weight decay unit uses a distributed scheduling framework to process feature records, performing exponential decay calculations every morning at midnight, with a fixed decay coefficient of 0.992. When a feature weight is below 0.1, it is moved to the archive database, and the weight change trend is monitored; features that increase for seven consecutive cycles trigger a re-evaluation process. The transfer testing unit includes test case generation, a test execution engine, and a result analysis component. The test case generator uses the K-means clustering algorithm to select representative samples, the test execution engine runs model inference in a Docker container environment, and the result analysis component uses the Bootstrap method to calculate the confidence interval of performance metrics. The audit trail module employs three data association methods: an event playback unit, a root cause analysis unit, and a knowledge graph unit. The event playback unit stores event records using a time-series database, with a keyframe interval of 10 seconds and a rendering engine supporting adjustable playback speeds from 0.5x to 8x. The root cause analysis unit performs ADF stationarity tests and Granger causality analysis, with a significance level set at 0.05 and a maximum lag order of 12. The knowledge graph unit uses the Neo4j graph database to store data, automatically updating the graph structure weekly. Entity recognition combines rule-based methods and a bidirectional LSTM model.

[0051] The model iteration module and the audit trail module establish a bidirectional data exchange channel, synchronizing data every six hours via a Kafka message queue. The feature distillation unit transmits feature importance weights to the knowledge graph unit to optimize network boundary weights, while the audit trail module feeds back root cause analysis results to the feature distillation unit to adjust the initial weights of new features. Data exchange uses Avro serialization format to ensure data transmission efficiency. The weight decay unit's distributed framework uses Apache Spark to implement feature sharding, with feature shards allocated based on feature ID hash values. The migration testing unit uses Kubernetes for container orchestration, with each test case allocated an independent namespace. The event replay unit's time-series database is implemented using InfluxDB, storing data in 24-hour partitions. The root cause tracing unit generates HTML reports containing causal relationship graphs and time-aligned curves, while the knowledge graph unit uses R-tree indexes to support spatial queries.

[0052] The system employs a microservice architecture to encapsulate various functional units, with inter-service communication utilizing the gRPC framework. The feature distillation unit implements the NSGA-II multi-objective optimization algorithm, considering feature importance, sparsity, and diversity. The weight decay unit supports customizable linear decay, exponential decay, and piecewise decay functions. The audit trail module's event storage utilizes a multi-level architecture of Redis and Cassandra to achieve hot and cold data separation. The model iteration module and audit trail module execute a joint optimization process every 24 hours, comprising three stages: data synchronization, model recalculation, and result verification. System logs record the optimization process in structured JSON format and are monitored and analyzed in real-time using the ELK technology stack. The feature distillation unit maintains the history of feature weight versions, the weight decay unit records manual adjustment information, and the migration test unit saves test environment configuration parameters. The event replay unit provides event tagging and time-point annotation functions, supporting multi-user collaborative analysis. The root cause tracing unit introduces the transitive entropy method to detect nonlinear causal relationships, and the knowledge graph unit implements force-directed layout visualization. The collaborative optimization of the model iteration module and audit trail module ensures the system continuously adapts to changes in the production environment, constantly improving anomaly detection accuracy through a closed-loop learning mechanism.

[0053] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based intelligent management system for communication equipment production, characterized in that, The application relates to a production data acquisition module for collecting raw data streams of various sensors on a communication equipment production line in real time and converting multi-source heterogeneous data into standardized data sequences with unified timestamps. The application relates to a feature engineering construction module for extracting time domain statistical features, frequency domain energy features and equipment state correlation features from the standardized data sequences to generate a high-dimensional feature vector set. The application relates to a dynamic clustering analysis module for performing online division on the high-dimensional feature vector set by using an incremental clustering algorithm to generate dynamic clustering cluster labels according to equipment running state similarity. The application relates to an abnormality detection engine module for establishing a multi-level Gaussian mixture model based on the dynamic clustering cluster labels and realizing abnormal probability quantification by calculating Mahalanobis distances of real-time data and each level model. The application relates to a production decision optimization module for integrating probability distribution results output by the abnormality detection engine module to generate equipment maintenance priority ranking and production rhythm adjustment instructions. The production data acquisition module comprises a data synchronization unit for resampling and aligning sampling frequencies of vibration sensors, temperature sensors and current sensors. 2.The machine learning based intelligent management system for communication equipment production according to claim 1, wherein, The feature engineering construction module performs operations including a time domain window sliding unit for intercepting the standardized data sequences with a fixed time window, calculating mean value, variance and kurtosis index of data in the window, a frequency domain transformation unit for extracting energy proportion of the first K harmonic components after fast Fourier transformation of the time domain window data, and a correlation analysis unit for calculating mutual information entropy values of different sensor data in a time shift state. The dynamic clustering analysis module adopts an incremental clustering algorithm implementation mode including a cluster center drift detection unit for triggering cluster splitting when the Euclidean distance between a newly-arrived high-dimensional feature vector and a nearest cluster center exceeds an adaptive threshold, a noise filtering unit for eliminating temporary clusters with a duration lower than a set threshold according to historical clustering stability, and a dimension compression unit for visualizing clustering boundaries after local linear embedding dimension reduction of the high-dimensional feature vector. The abnormality detection engine module comprises a device-level model for constructing a probability density function containing all clustering clusters in a normal working state for a single communication equipment, a production line-level model for establishing a joint probability distribution by integrating common features of same station equipment on the whole production line, and a global-level model for establishing an abnormal mode feature library across production lines by fusing historical fault case data. The production decision optimization module comprises a health degree scoring unit for calculating a percentage value of a current equipment state deviating from a standard working condition according to the Mahalanobis distance, a fault prediction unit for evaluating a probability distribution curve of remaining service life by combining equipment cumulative running length, and a scheduling strategy unit for dynamically calculating an optimal equipment maintenance time window based on the scoring value and the probability distribution curve. 3.The machine learning based intelligent management system for communication device production according to claim 2, wherein, The application further comprises a digital twin mapping module for establishing a three-dimensional visual interface to display spatial distribution evolution of the dynamic clustering clusters in real time, and a model iteration module for injecting newly-generated abnormal case features into the global-level model for parameter fine-tuning every week. ​ ​ ​ 4.The machine learning based intelligent management system for communication device production according to claim 3, wherein, ​ ​ ​ ​ 5.The machine learning based intelligent management system for communication equipment production according to claim 4, characterized in that, ​ ​ ​ ​ 6.The machine learning based intelligent management system for communication device production according to claim 5, wherein, ​ ​ ​ ​ 7.The machine learning based intelligent management system for communication device production according to claim 6, wherein, ​ ​ ​ An audit tracking module records all cluster splitting and merging events and corresponding production batch information. 8.The machine learning based intelligent management system for communication device production according to claim 7, wherein, Implementations of the digital twin mapping module include: A heat map rendering unit represents abnormal probability density of different regional devices using a color gradient; A topology graph analysis unit displays potential impact range of fault propagation paths between devices; A virtual calibration unit allows users to manually label abnormal regions in a three-dimensional interface to assist model training. 9.The machine learning based intelligent management system for communication device production according to claim 8, wherein, A parameter fine-tuning process performed by the model iteration module includes: A feature distillation unit extracts unique feature combinations that distinguish new cases from existing abnormal patterns; A weight decay unit gradually reduces model attention to historical fault features that exceed a valid period; A migration test unit verifies generalization capability indicators of the updated model on similar production lines. 10.The machine learning based intelligent management system for communication equipment production according to claim 9, wherein, Data correlation methods of the audit tracking module include: An event playback unit displays the full life cycle of a specific cluster from its inception to its demise on a timeline; A root cause tracing unit establishes a causal chain between the final device failure and an early abnormal cluster; A knowledge graph unit constructs an associated network topology structure of fault patterns and production process parameters.

Images