Industrial equipment state intelligent monitoring system based on big data

Abstract

The invention relates to the technical field of industrial Internet of Things intelligent manufacturing, and discloses an industrial equipment state intelligent monitoring system based on big data. The system comprises a data acquisition module, a feature abstraction module, a knowledge graph evolution module and an intelligent early warning decision module. Firstly, cleaning verification and multi-scale feature extraction are carried out on original data collected by a sensor, and an equipment state feature abstract is generated; furthermore, an equipment state monitoring knowledge graph is autonomously constructed and dynamically updated by analyzing an association rule between the features; and performing projection and similarity matching on the real-time operation data in the topological structure of the knowledge graph, and triggering hierarchical early warning according to matching deviation. According to the method, the defects that a traditional static monitoring model is poor in adaptability and early warning lacks interpretability are overcome, accurate perception and self-adaptive judgment of dynamic evolution of the equipment state are achieved, and the accuracy and reliability of monitoring and early warning are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) intelligent manufacturing technology, specifically to an intelligent monitoring system for the status of industrial equipment based on big data. Background Technology

[0002] Current industrial equipment condition monitoring mainly relies on alarms or isolated data analysis models based on fixed thresholds. Fixed threshold methods are insensitive to early performance degradation and gradual failures caused by multiple factors. Traditional predictive models often treat monitoring parameters as independent or simply linearly correlated variables, making it difficult to capture the complex nonlinear dynamic coupling relationships between multidimensional parameters such as vibration, temperature, and current, leading to frequent false alarms and missed alarms.

[0003] The knowledge representation and decision-making mechanisms of existing technologies are usually static. After system deployment, their internal rules or model parameters are fixed and lack adaptive evolution capabilities. During long-term operation, wear, aging, and changes in operating conditions can cause the baseline of normal operating status to drift. Static models cannot adapt to this dynamic process, inevitably leading to a decrease in early warning accuracy, or even misjudging new normal operating conditions as abnormalities.

[0004] Traditional early warning decision-making relies on threshold judgments of individual data points or anomaly scoring of isolated models, lacking a macro-level insight into the overall operational status of equipment. This method cannot place real-time data within a context-rich knowledge network for semantic comparison; the decision-making process is isolated, making it difficult to bridge the gap between anomaly detection and root cause analysis, resulting in poor interpretability of early warning results. The purpose of this invention is to address the core shortcomings of insufficient adaptability of static models and the lack of contextual understanding in decision-making mechanisms. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent monitoring system for the status of industrial equipment based on big data, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an intelligent monitoring system for the status of industrial equipment based on big data, the system comprising: An industrial equipment operation data acquisition module is used to continuously capture raw monitoring data streams from a sensor network deployed on industrial equipment, and to perform data cleaning and quality verification on the raw monitoring data streams to form a standardized set of equipment operation data. The equipment status feature abstraction module is used to perform multi-scale feature transformation on the standardized equipment operation data set, generate a time-series correlated equipment status feature image, and perform feature reduction on the equipment status feature image to obtain an equipment status feature summary. The monitoring knowledge graph evolution module is used to parse the feature association rules hidden in the equipment status feature summary, construct the equipment status feature association network based on the feature association rules, and evolve the equipment status feature association network into an industrial equipment status monitoring knowledge graph through a dynamic update mechanism. The intelligent early warning decision module is used to project real-time acquired industrial equipment operation data onto the industrial equipment status monitoring knowledge graph for similarity matching and deviation calculation, and trigger hierarchical early warning decisions based on the matching and calculation results.

[0007] Preferably, the raw monitoring data stream includes vibration signals, temperature readings, pressure values, current and voltage waveforms, equipment control parameters, and ambient temperature and humidity data.

[0008] Preferably, the step of cleaning and quality verification of the original monitoring data stream to form a standardized set of equipment operation data specifically includes: The original monitoring data stream is subjected to outlier removal and missing value imputation to obtain pre-processed monitoring data; The pre-processed monitoring data is subjected to dimensional unification and numerical normalization operations to obtain standardized monitoring data. Standardized monitoring data are aligned and integrated according to equipment number and timestamp to form a standardized set of equipment operation data.

[0009] Preferably, the step of performing multi-scale feature transformation on the standardized equipment operation data set to generate a equipment state feature image with temporal correlation specifically includes: The standardized equipment operation data set is subjected to time-domain statistical feature extraction, frequency-domain energy feature decomposition, and time-frequency domain joint feature analysis to obtain a multi-scale feature set. By sliding and stitching together various features from the multi-scale feature set according to time windows, a device status feature image is constructed.

[0010] Preferably, the step of reducing the device state feature image to obtain the device state feature summary specifically includes: Calculate the correlation strength between each feature dimension in the device status feature image and the device health status; The feature dimensions are sorted according to the correlation strength, and the feature dimensions with a correlation strength exceeding a preset threshold are selected. The selected feature dimensions are recombined to generate a device status feature summary with reduced dimensionality.

[0011] Preferably, the step of parsing the feature association rules implicit in the device status feature summary and constructing a device status feature association network based on the feature association rules specifically includes: The device status feature summary was analyzed using association rule mining methods to discover frequently co-occurring feature combination patterns. An initial device state feature association network is constructed using features as nodes and the co-occurrence strength between features as edges. Redundant edge pruning and core community discovery are performed on the initial device state feature association network to obtain a simplified device state feature association network.

[0012] Preferably, the step of evolving the equipment state feature association network into an industrial equipment condition monitoring knowledge graph through a dynamic update mechanism specifically includes: New device status feature summaries are continuously injected into the device status feature association network; Adjust the weights of network nodes and edges based on the newly injected data, and identify newly added feature relationships; After iterative updates, an industrial equipment condition monitoring knowledge graph containing the evolutionary patterns of equipment states is formed.

[0013] Preferably, the step of projecting the real-time acquired industrial equipment operation data onto the industrial equipment condition monitoring knowledge graph for similarity matching and deviation calculation specifically includes: The real-time industrial equipment operation data is subjected to the same feature abstraction process as historical data to obtain a summary of real-time equipment status features. Calculate the similarity between the real-time equipment status feature summary and the feature vectors of each typical status pattern in the industrial equipment status monitoring knowledge graph; Calculate the degree of deviation of the real-time device state feature summary from the normal state baseline in the knowledge graph.

[0014] Preferably, the step of triggering a tiered early warning decision based on the matching and calculation results specifically includes: The matching calculation results between the real-time equipment status feature summary and the industrial equipment status monitoring knowledge graph are obtained, and the matching calculation results include feature vector similarity and feature deviation. A weighted fusion method is used to synthesize feature vector similarity and feature deviation into a comprehensive anomaly index; The comprehensive anomaly index is compared with multiple preset level thresholds step by step to classify the equipment status level; Query the warning rule library based on the device status level to obtain the corresponding warning action configuration; Configure early warning actions, generate early warning signals, and distribute them to monitoring terminals.

[0015] Preferably, calculating the correlation strength between each feature dimension in the device state feature image and the device health status specifically includes: Collect historical equipment operation data and label the equipment health status to form a training dataset; For each feature dimension in the device state feature image, extract the numerical sequence of the feature dimension in the training dataset; Calculate the statistical correlation measure between the numerical sequence of each feature dimension and the device health status label sequence; The correlation strength value for each feature dimension is determined based on the absolute value of the statistical correlation measure. The association strength values ​​are normalized to obtain standardized association strength weights.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By automatically parsing feature association rules from equipment status feature summaries and constructing a monitoring knowledge graph with a dynamic update mechanism based on these rules, the adaptability limitations of traditional static models due to equipment aging or changes in operating conditions are overcome. This technology enables the knowledge graph to continuously learn from new data generated during equipment operation, adjusting the relationships and weights between its internal nodes in real time, forming a constantly evolving diagnostic knowledge base. This self-evolution mechanism ensures the synchronization between the status assessment benchmark and the actual operating status of the equipment, improving the reliability and accuracy of the system in long-term operation.

[0017] By projecting real-time device operation data onto the topology of a dynamic knowledge graph for similarity matching and deviation calculation, this method changes the traditional early warning logic that relies on fixed thresholds or single model scores. It achieves context-aware assessment of device operation status by analyzing the location of real-time data points in the knowledge network and their topological distances to nodes of various state patterns. This graph-based comparison method can keenly identify subtle anomaly patterns with coordinated changes in multiple parameters, not only advancing the fault warning time window but also effectively distinguishing anomaly types and severity levels through the semantic information contained in the matching results.

[0018] The results of knowledge graph topology matching naturally form the basis for hierarchical early warning decisions, making the early warning output both accurate and interpretable. The matching degree between real-time data and different state patterns in the knowledge graph directly reflects the degree of deviation of the current device state, providing a quantitative basis for early warning level classification. At the same time, the matching process intuitively presents the association path between abnormal features and existing fault modes in the graph, providing clear logical clues for operation and maintenance personnel to trace the root cause of anomalies, realizing a leap from simple alarms to assisted diagnosis. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent monitoring system for industrial equipment status based on big data as described in this invention. Figure 2 A flowchart for data cleaning and standardization; Figure 3 A flowchart for feature reduction; Figure 4 This is a feature correlation strength analysis diagram; Figure 5 This is a heatmap of association rules. Detailed Implementation

[0020] 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.

[0021] Please see Figure 1 This invention provides an intelligent monitoring system for industrial equipment status based on big data. The system includes: an industrial equipment operation data acquisition module that continuously captures raw monitoring data streams from a sensor network deployed on the industrial equipment, and performs data cleaning and quality verification on the raw monitoring data streams to form a standardized set of equipment operation data; an equipment status feature abstraction module that performs multi-scale feature transformation on the standardized set of equipment operation data to generate a time-series correlated equipment status feature image, and performs feature reduction on the equipment status feature image to obtain an equipment status feature summary; a monitoring knowledge graph evolution module that parses the feature association rules implicit in the equipment status feature summary, constructs an equipment status feature association network based on the feature association rules, and evolves the equipment status feature association network into an industrial equipment status monitoring knowledge graph through a dynamic update mechanism; and an intelligent early warning decision module that projects the real-time acquired industrial equipment operation data onto the industrial equipment status monitoring knowledge graph for similarity matching and deviation calculation, and triggers hierarchical early warning decisions based on the matching and calculation results.

[0022] Example 1: See Figure 2In practical implementation, the industrial equipment operation data acquisition module is responsible for continuously capturing raw monitoring data streams from the sensor network deployed on the industrial equipment. These raw monitoring data streams contain various physical parameters, specifically vibration signals, temperature readings, pressure values, current and voltage waveforms, equipment control parameters, and environmental temperature and humidity data. Vibration signals are acquired by accelerometers or vibration sensors, reflecting the operating status of the equipment's mechanical components; temperature readings are obtained through thermocouples or infrared sensors, monitoring the thermal state of key components; pressure values ​​are measured by pressure sensors, indicating the pressure status of fluid systems or sealed containers; current and voltage waveforms are recorded using current transformers and voltage sensors, characterizing the load and energy consumption of the electrical system; equipment control parameters originate from programmable logic controllers or distributed control systems, including operating commands such as equipment start / stop status and set speeds; and environmental temperature and humidity data are acquired through temperature and humidity sensors, describing the external environmental conditions of the equipment. The sensor network is built based on Industrial Internet of Things (IIoT) protocols, such as MQTT or OPCUA. Sensor nodes continuously transmit data to an edge gateway or directly to a cloud platform at a specific sampling frequency, forming a continuous raw monitoring data stream.

[0023] In practice, data cleaning and quality verification of the raw monitoring data stream are crucial steps in forming a standardized set of equipment operation data. The data cleaning and quality verification process includes outlier removal, missing value imputation, dimensional unification, and numerical normalization. Outlier removal identifies and removes erroneous data that significantly deviates from the normal range in the raw monitoring data stream. Statistical methods such as Z-score-based standard deviation analysis or box plot methods based on interquartile ranges are used. An absolute value threshold for the Z-score or upper and lower limits for the box plot are set, and data points exceeding the threshold are identified as outliers and removed. Missing value imputation fills in missing data in the data stream caused by sensor communication interruptions or acquisition failures. Linear interpolation methods or regression models trained on historical data are used. Linear interpolation fills in gaps between adjacent valid data points, while regression models use data from other related sensors to predict missing values. After outlier removal and missing value imputation, a set of complete, error-free, pre-processed monitoring data is obtained.

[0024] In practice, the pre-processed monitoring data undergoes dimensional unification and numerical normalization. Dimensional unification converts data from different sensors with varying physical units into a consistent measurement standard. For example, all temperature readings are converted to degrees Celsius, all pressure values ​​to Pascals, and all vibration signal amplitudes to meters per second squared. Numerical normalization scales the dimensionally unified data values ​​to a uniform range, eliminating biases caused by differences in dimensions and orders of magnitude. Common methods include min-max scaling and Z-score normalization. Min-max scaling linearly transforms the data to the [0,1] interval, while Z-score normalization converts the data into a standard normal distribution with a mean of 0 and a variance of 1. The standardized monitoring data is then aligned and integrated according to equipment number and timestamp to form the final standardized equipment operation data set. The alignment operation strictly aligns data generated by the same equipment on different sensors on the timeline based on the unique equipment number and precise timestamp information attached to each data point. The timestamp synchronization accuracy needs to reach millisecond level or higher to reflect the overall operating status of the equipment at the same moment. The integration process organizes the aligned multi-source data into a structured dataset based on time series, which can be stored using a time-series database such as InfluxDB or TimescaleDB. Each record contains the device number, timestamp, and standardized values ​​from multiple sensor measurement points. This standardized device operation dataset serves as input for subsequent modules, and its data structure ensures data consistency and traceability.

[0025] In some embodiments, the data cleaning process of the raw monitoring data stream can introduce more complex quality verification rules, such as data rationality checks based on the equipment's physical model. Rules could stipulate that temperature readings should not exceed the material's melting point, pressure values ​​should be within the equipment's design pressure range, and current values ​​should not exceed a specific multiple of the motor's rated current. In specific implementations, the data cleaning and quality verification module can be deployed on edge computing devices to achieve preliminary data processing at the acquisition end, reducing uplink data transmission bandwidth pressure. It is understood that the missing value imputation method in data cleaning is not limited to linear interpolation; for sequences with significant nonlinear characteristics, spline interpolation or imputation methods based on the K-nearest neighbor algorithm can be used. In specific implementations, the method selected for numerical normalization needs to be adjusted according to the requirements of subsequent feature abstraction algorithms. If distance-based machine learning algorithms are subsequently used, Z-score normalization usually yields better results.

[0026] In some embodiments, the standardized device operation data set can be stored in columnar format as Parquet or ORC files to improve read / write efficiency under large data volumes. It is understood that the data alignment process needs to handle inconsistencies in sampling rates between different sensors. For sensor data with lower sampling rates, forward padding or nearest-neighbor interpolation methods can be used to upscale it to the frequency of the sensor with the highest sampling rate. Optionally, for large data measurement points such as high-frequency vibration signals, data downsampling or compression steps can be introduced before data cleaning to reduce data volume while retaining key information. Optionally, the quality verification process can include a data continuity check to identify and mark or correct any unwanted step jumps in the data stream.

[0027] Example 2: In specific implementation, the equipment state feature abstraction module performs multi-scale feature transformation on the standardized equipment operation data set to generate a time-series correlated equipment state feature image. The multi-scale feature transformation process first performs time-domain statistical feature extraction, frequency-domain energy feature decomposition, and time-frequency joint feature analysis on the standardized equipment operation data set to obtain a multi-scale feature set. Time-domain statistical feature extraction calculates a series of statistics for each sensor data sequence in the standardized equipment operation data set within a fixed-length time window. These statistics include mean, variance, peak factor, kurtosis, waveform factor, and impulse factor. The mean reflects the average level of the data sequence, the variance characterizes the dispersion of the data, the peak factor is the ratio of the peak value to the root mean square (RMS) used to detect impulse components, the kurtosis describes the difference between the data distribution and the normal distribution, the waveform factor is the ratio of the RMS to the mean absolute value, and the impulse factor is the ratio of the peak value to the mean absolute value. These statistics together constitute the time-domain feature vector. Frequency domain energy feature decomposition transforms time-domain signals into the frequency domain using a Fast Fourier Transform (FFT) algorithm. The FFT converts discrete-time sequences into frequency components, then calculates the energy proportion, centroid frequency, frequency variance, and spectral kurtosis for each frequency band. The energy proportion is the ratio of energy in a specific frequency band to the total energy; the centroid frequency is an indicator of the frequency-weighted average; the frequency variance describes the dispersion of the frequency distribution; and the spectral kurtosis characterizes the peak characteristics of the frequency domain. Joint time-frequency domain feature analysis employs wavelet transform. The wavelet transform obtains the signal's energy distribution in the time-frequency plane through translation and scaling of the mother wavelet function in time and scale. It extracts the energy entropy of wavelet coefficients, the energy features of wavelet packet nodes, and time-frequency moment features. Energy entropy measures the uncertainty of time-frequency energy, wavelet packet node energy provides frequency band division for multi-resolution analysis, and time-frequency moment features, including first and second moments, describe the center and dispersion of the time-frequency distribution.

[0028] In practice, various features in the multi-scale feature set are slidably stitched together according to time windows to construct a device state feature image. The size of the time window is determined based on the device's operating cycle and sampling frequency. For example, for rotating machinery, the time window can be set as an integer multiple of the rotation cycle. The sliding step size is usually smaller than the window length to maintain temporal continuity. The sliding stitching operation first organizes the time-domain statistical features, frequency-domain energy features, and time-frequency joint features within each time window into feature vectors. The time-domain statistical feature vector contains multiple statistics, the frequency-domain energy feature vector contains multiple frequency band indices, and the time-frequency joint feature vector contains wavelet transform-derived parameters. Then, these feature vectors are stitched together in a predetermined order into a high-dimensional composite feature vector. The dimension of the composite feature vector is the sum of the dimensions of each sub-feature. The stitching order can be arranged in the order of time domain, frequency domain, and time-frequency domain, or adjusted according to the importance of the features. Finally, the window is moved along the time axis by a sliding step size, and the above process is repeated to generate a series of time-continuous high-dimensional feature vector sequences. This sequence is the device state feature image. The device state feature image is stored in matrix form, with rows representing time points and columns representing feature dimensions, thus preserving temporal correlation.

[0029] In some embodiments, time-domain statistical feature extraction may include the calculation of higher-order statistics, such as third-order moment skewness and fourth-order moment kurtosis. Skewness measures the asymmetry of the data distribution, and kurtosis reflects the sharpness of the distribution. Frequency-domain energy feature decomposition can use short-time Fourier transform (SFT) instead of fast Fourier transform (FFT). SFT obtains the time-varying spectrum through a windowed sliding method, suitable for non-stationary signal analysis. Joint time-frequency domain feature analysis can use empirical mode decomposition (EMD), which adaptively decomposes the signal into intrinsic mode functions (EMFs) and then extracts the energy features and instantaneous frequencies of the EMFs. The generation of multi-scale feature sets can introduce a feature selection step, applying filtering or wrapping methods immediately after feature extraction to reduce redundancy, such as removing highly correlated features using correlation coefficients. The sliding stitching process can optimize the window overlap rate, adjustable from 50% to 90%, to balance temporal resolution and computational efficiency. The storage format of the device state feature image can use sparse matrix representation; sparse storage saves space when there are many zero values ​​in the feature vector.

[0030] It is understandable that the window length for time-domain statistical feature extraction needs to match the dynamic characteristics of the device; an excessively long window may result in smooth and rapid changes, while an excessively short window may introduce noise. The frequency resolution of frequency-domain energy feature decomposition depends on the sampling frequency and window length; high frequency resolution requires a long window, but this reduces temporal resolution. It is also understandable that the feature map constructed by sliding stitching may face the curse of dimensionality, thus subsequent feature reduction steps are essential. The temporal correlation of device state feature maps makes them suitable for deep learning models such as recurrent neural networks or long short-term memory networks to capture long-term dependencies. Optionally, time-domain statistical feature extraction can integrate automated outlier robust estimation methods, such as using the median and absolute median difference instead of the mean and variance, to reduce the impact of outliers. Frequency-domain energy feature decomposition can be combined with envelope analysis; first, a Hilbert transform is performed on the signal to extract the envelope, and then spectral analysis is performed on the envelope to enhance fault characteristics.

[0031] In practical implementation, the mathematical expression of frequency domain energy eigenvalue decomposition can be expressed using the discrete Fourier transform formula: in: Describes the complex spectrum of the k-th frequency component. It is the nth sampling point of the time-domain signal. This is the total number of sampling points. It is a frequency index. It is a time index. It is the imaginary unit. This formula converts the time-domain sequence into a frequency-domain representation, and then the energy spectrum is obtained by taking the modulus and squaring, which is used to calculate the frequency-domain energy characteristics.

[0032] In practical implementation, the overall process of multi-scale feature transformation requires coordinated computing resources. For large-scale data scenarios, distributed computing frameworks such as Apache Spark can be used to process feature extraction from different sensors in parallel. Time-domain statistical feature extraction, frequency-domain energy feature decomposition, and time-frequency joint feature analysis can be pipelined to reduce memory consumption. The construction process of the device status feature map should record feature metadata, including feature name, extraction parameters, and timestamp, for easy traceability and debugging. Verification of time-series correlated device status feature maps can be performed using visualization tools, such as plotting the feature map as a heatmap or time series diagram to visually examine the feature change patterns over time. The data structure of the feature map is usually in tensor form. A three-dimensional tensor can represent time, feature dimensions, and service number, supporting batch processing.

[0033] Example 3: See Figure 3In practical implementation, the equipment status feature abstraction module performs feature reduction on the equipment status feature map to obtain an equipment status feature summary. The core of the feature reduction process is to calculate the correlation strength between each feature dimension in the equipment status feature map and the equipment health status. Calculating the correlation strength requires collecting historical equipment operation data and labeling it with equipment health status tags to form a training dataset. The historical equipment operation data comes from long-term operating equipment sensor records. The equipment health status tags are determined based on equipment maintenance records, fault logs, and expert evaluations. The tags can be divided into multiple levels such as "normal," "slightly deteriorated," "severely deteriorated," and "faulty." The training dataset is a collection of data containing historical feature vector sequences and their corresponding health status tag sequences. For each feature dimension in the equipment status feature map, the numerical sequence of that feature dimension in the training dataset needs to be extracted. A feature dimension refers to each column in the equipment status feature map matrix, representing a feature. The numerical sequence is a vector composed of the values ​​of that feature at all time points. Then, the statistical correlation measure between the numerical sequence of each feature dimension and the equipment health status tag sequence is calculated. The statistical correlation measure is used to quantify the strength of the linear or monotonic relationship between the two sequences. The Pearson correlation coefficient is suitable for measuring linear relationships, while the Spearman rank correlation coefficient is suitable for measuring monotonic relationships. Based on the absolute value of the statistical correlation metric, the correlation strength value for each feature dimension is determined. A larger absolute value indicates a stronger correlation between the feature dimension and changes in device health status. The correlation strength value is a non-negative real number. Finally, the correlation strength values ​​are normalized to obtain standardized correlation strength weights. Normalization ensures that the sum of all weights is 1, facilitating comparison and subsequent selection. Normalization methods can employ min-max scaling or the softmax function.

[0034] In practice, feature dimensions are sorted based on the calculated standardized correlation strength weights, and feature dimensions with correlation strength exceeding a preset threshold are selected. The sorting operation arranges all feature dimensions in descending order of standardized correlation strength weights, generating an ordered list of feature dimensions. The preset threshold is a configurable parameter used to filter out features with weak correlations; it can be set as an absolute value or a relative proportion. The selection process traverses the ordered list, filtering out feature dimensions with standardized correlation strength weights greater than or equal to the preset threshold, forming a subset of candidate feature dimensions. The selected feature dimensions are then recombine to generate a device status feature summary with reduced dimensionality. This recombination operation extracts the original data corresponding to the candidate feature dimension subset from the high-dimensional device status feature image, constructing a new feature matrix with a reduced number of features. This new feature matrix is ​​the device status feature summary, which retains the feature information most relevant to the device's health status while significantly reducing data dimensionality.

[0035] In some embodiments, the labeling of equipment health status can be automated, for example, using equipment operating parameters as proxy labels for status deterioration. The calculation of statistical correlation metrics can be combined with hypothesis testing; for example, when calculating the Pearson correlation coefficient, the p-value is output simultaneously, retaining only associations with p-values ​​less than the significance level to ensure statistical significance of the association strength. The ranking of feature dimensions can comprehensively consider multiple correlation indicators; for example, a weighted fusion of the Pearson and Spearman correlation coefficients can be used to obtain a comprehensive association strength score. The determination of the preset threshold can employ adaptive methods; for example, the elbow rule can be used to observe the decreasing curve of association strength values, selecting the value at the inflection point of the curve as the threshold. The generation of equipment status feature summaries can incorporate feature transformation techniques; for example, principal component analysis can be performed on the selected features to further compress the dimensions while retaining most of the variance.

[0036] It is understandable that the quality of the training dataset directly determines the accuracy of the correlation strength calculation; therefore, it is necessary to ensure that historical data covers various operating states and typical failure modes of the equipment. Statistical correlation measures mainly reflect linear or monotonic relationships between variables. For complex nonlinear relationships, their evaluation capability is limited, and it may be necessary to introduce nonlinear measures such as mutual information as a supplement. Optionally, different correlation calculation methods can be used for different types of equipment health status labels when calculating correlation strength. For continuous health indicators, Pearson correlation coefficient is used; for categorical health status, ANOVA F-value or mutual information is used. After generating a summary of equipment status features, it can be visualized, such as by plotting feature importance bar charts or parallel coordinate graphs, to intuitively display the retained key features and their numerical distribution, assisting engineers in understanding the equipment status characterization.

[0037] In practice, the formula for calculating the Pearson correlation coefficient can be expressed as: in: The Pearson correlation coefficient represents the relationship between the feature dimension sequence x and the health status label sequence y. It is the number of samples in the training dataset. It is the value of the feature dimension x at the i-th sample point. It is the value of the health status label y at the i-th sample point. It is the arithmetic mean of all values ​​in the feature dimension x. It is the arithmetic mean of all values ​​for the health status label y. (Calculated...) The absolute value is used as the initial correlation strength value for this feature dimension.

[0038] In practical implementation, the feature reduction process can be accelerated using a distributed computing framework, such as distributing the task of calculating the correlation strength of different feature dimensions to multiple computing nodes for parallel execution. Standardized correlation strength weights can be persistently stored in a database, forming a feature importance knowledge base to guide feature selection for new devices. The setting of preset thresholds can incorporate domain knowledge; for example, a more conservative threshold can be used on critical devices to retain more features, while a more aggressive threshold can be used on less important devices to achieve stronger dimensionality reduction. The data structure of the device status feature summary typically adopts a time-series matrix form similar to the original device status feature map, but with a significantly reduced number of columns, facilitating processing by the subsequent knowledge graph construction module. The entire feature reduction process should be designed as a configurable pipeline, allowing users to adjust parameters such as the correlation strength calculation method and preset thresholds according to specific device types and monitoring objectives to optimize monitoring results.

[0039] See Figure 4 This paper presents the correlation strength analysis results between various feature dimensions and equipment health status in industrial equipment condition monitoring. By calculating the Pearson correlation coefficient between each feature and equipment health status, and then standardizing it to obtain the correlation strength weight, the paper intuitively reflects the importance of different features in equipment condition monitoring. The figure uses a horizontal bar chart to clearly show the top ten key features and their corresponding standardized weight values. The length of each bar represents the relative importance of that feature in the equipment health status assessment; a larger weight value indicates a stronger correlation between the feature and changes in equipment status. The red dashed threshold line in the chart is used for feature selection; features with weights exceeding this threshold are retained for constructing the equipment status feature summary. This visualization method provides engineers with an intuitive ranking of feature importance, helping to optimize the feature selection strategy of the monitoring system and improve the accuracy and efficiency of condition monitoring.

[0040] Example 4: In specific implementation, the monitoring knowledge graph evolution module analyzes the feature association rules implicit in the equipment status feature summary, constructs an equipment status feature association network based on these rules, and uses association rule mining methods to analyze the equipment status feature summary during the analysis process. Association rule mining methods are used to discover frequently co-occurring feature combination patterns in the equipment status feature summary. Frequently co-occurring feature combination patterns refer to the regularity of a set of feature dimensions simultaneously appearing abnormally or within a specific numerical range in multiple time samples of the equipment status feature summary; for example, "increased vibration energy in the high-frequency band" and "increased bearing temperature" often appear simultaneously. Association rule mining can use the Apriori algorithm or the FP-Growth algorithm. The Apriori algorithm iteratively generates frequent itemsets and calculates support and confidence, while the FP-Growth algorithm avoids generating candidate itemsets by constructing a frequent pattern tree. Support represents the frequency of feature combinations appearing in the dataset, and confidence represents the conditional probability that one set of features also appears when another set of features appears. By setting minimum support and minimum confidence thresholds, meaningful strong association rules are selected, revealing the intrinsic connections between equipment status features.

[0041] In practical implementation, an initial equipment state feature association network is constructed using features as nodes and the co-occurrence strength between features as edges. Nodes represent each feature dimension in the equipment state feature summary, such as specific features like "mean axial vibration" and "motor current harmonic distortion rate." Edges represent association rules between feature nodes, and the weight of the edge is determined by a metric of the association rule, such as lift or confidence. Lift measures the effectiveness of the rule; a lift greater than 1 indicates a positive correlation. The initial equipment state feature association network is a weighted undirected graph, which can be stored and represented using a graph database. Then, redundant edge pruning and core community discovery are performed on the initial network to obtain a simplified network. Redundant edge pruning removes weakly associated edges with weights below a threshold and isolated nodes with low connectivity, making the network structure clearer. Core community discovery uses graph clustering algorithms to identify closely connected feature node clusters in the network, such as the Louvain community discovery algorithm or label propagation algorithm. Each community represents a functionally related feature group, such as a feature community related to "thermal state" or a feature community related to "mechanical vibration."

[0042] In practical implementation, a dynamic update mechanism evolves the equipment status feature association network into an industrial equipment condition monitoring knowledge graph. This mechanism continuously injects new equipment status feature summaries into the network. These new summaries originate from incremental data generated after feature abstraction of real-time equipment operation data and are injected into the system in a streaming manner. The weights of network nodes and edges are adjusted based on the newly injected data, and new feature associations are identified. Weight adjustment can employ a sliding window model or an exponential decay model to update historical weights. The sliding window model considers only data from the most recent period, while the exponential decay model assigns a decay factor to historical weights. The identification of new feature associations is achieved through incremental association rule mining algorithms, such as the Fast-Update algorithm. When a co-occurrence pattern of features not covered by existing rules appears in the new data, a new edge is created. After iterative updates, an industrial equipment condition monitoring knowledge graph containing the evolutionary patterns of equipment status is formed. This knowledge graph not only includes feature nodes and associated edges but also node attribute information and edge temporal evolution information, dynamically reflecting changes in the relationships between equipment status features.

[0043] In some embodiments, association rule mining can employ fuzzy association rule methods to process continuous numerical features in the device state feature summary. First, the continuous features are discretized into several fuzzy intervals, and then the association relationships between these intervals are mined. Core community discovery can be combined with modularity optimization metrics to automatically determine the optimal number of community divisions in the network, ensuring tight connections within communities and sparse connections between communities.

[0044] In practice, the support of association rules can be calculated using the following formula: in: Representation rules Support and It is a feature set. The dataset contains itemsets. and itemsets The number of transactions, This represents the total number of transactions in the dataset. A higher support value indicates that the feature combination pattern occurs more frequently. See Table 1, which shows a simplified association rule for device status features.

[0045] Table 1: Association Rules for Equipment Status Features In practical implementation, constructing the equipment status feature association network requires defining the data structures for nodes and edges. Nodes include attributes such as feature identifier, feature name, and feature type, while edges include attributes such as source node, target node, association strength, and rule source. Implementing a dynamic update mechanism requires designing an efficient data structure to support incremental updates, such as using an adjacency list to store the network structure. The query interface for the industrial equipment status monitoring knowledge graph needs to support graph traversal queries, such as finding all features directly associated with a given feature, or finding the shortest path between two features.

[0046] See Figure 5 This heatmap displays the results of association rule analysis between equipment status characteristics. The heatmap visually represents the strength of statistical correlation between different feature dimensions using varying color intensity; a higher correlation coefficient indicates a stronger correlation between two features during equipment operation. The color of each cell in the graph represents the Pearson correlation coefficient between the corresponding row and column features, and the numerical labels precisely display the specific value of the correlation. Diagonal cells are hidden because they represent the correlation between a feature and itself, highlighting the interrelationships between different features. This heatmap visualization helps engineers quickly identify strong correlation patterns in equipment status characteristics, discovering feature combinations that frequently exhibit anomalies or consistent trends. Analyzing the correlations between features provides a better understanding of equipment operating mechanisms, offering crucial information for fault diagnosis and preventative maintenance, and also helps optimize feature selection strategies for monitoring systems.

[0047] Example 5: In specific implementation, the intelligent early warning decision module projects real-time acquired industrial equipment operation data onto the industrial equipment condition monitoring knowledge graph for similarity matching and deviation calculation. The real-time industrial equipment operation data originates from a continuously operating sensor network. This data undergoes the same feature abstraction processing as historical data. Feature abstraction processing includes multi-scale feature transformation and feature reduction steps. Multi-scale feature transformation extracts time-domain statistical features, decomposes frequency-domain energy features, and performs joint time-frequency domain feature analysis on the real-time data. Feature reduction filters key feature dimensions based on pre-calculated feature association strength weights, ultimately obtaining a real-time equipment condition feature summary. The real-time equipment condition feature summary is a feature vector with reduced dimensionality, representing the comprehensive state of the equipment at the current moment. The projection process involves calculating the similarity between the feature vectors of the real-time equipment condition feature summary and various typical state patterns in the industrial equipment condition monitoring knowledge graph. The industrial equipment condition monitoring knowledge graph stores various typical state patterns learned from historical data, such as "normal operation mode," "minor wear mode," and "bearing failure mode," with each typical state pattern represented by a feature vector center point. Feature vector similarity is calculated using cosine similarity, which focuses on the directional consistency of vectors. The cosine of the angle between the real-time feature vector and the feature vector of each typical state pattern is calculated; a cosine value closer to 1 indicates higher similarity. Simultaneously, the deviation of the real-time device state feature summary from the normal state baseline in the knowledge graph is calculated. The normal state baseline is a reference point for feature vectors statistically obtained from historical normal data, typically represented by the mean vector of the feature dimensions. The feature deviation is calculated using Mahalanobis distance, which considers the covariance relationship between feature dimensions and can more accurately reflect the multivariate deviation of the real-time state from the normal state.

[0048] In practical implementation, a tiered early warning decision is triggered based on the matching and calculation results. The tiered early warning decision process first obtains the matching calculation results between the real-time equipment status feature summary and the industrial equipment status monitoring knowledge graph. The matching calculation results include two key indicators: feature vector similarity and feature deviation. A weighted fusion method is used to synthesize the feature vector similarity and feature deviation into a comprehensive anomaly index. The weighted fusion method assigns a weight coefficient to each index. The weight coefficient is optimized and determined based on historical verification data. For example, the feature vector similarity weight is set to 0.6, and the feature deviation weight is set to 0.4. The fusion formula is a linear weighted sum: Comprehensive Anomaly Index = (1 - Feature Vector Similarity) × Weight Coefficient A + Feature Deviation × Weight Coefficient B, where (1 - Feature Vector Similarity) converts similarity into difference, so that the larger the comprehensive anomaly index value, the higher the probability of an abnormal status. Then, the comprehensive abnormality index is compared with multiple preset level thresholds step by step to classify the equipment status level. The preset level thresholds are usually set to three or more, such as threshold T1=0.3, threshold T2=0.6, and threshold T3=0.8. The comparison process divides the equipment status into different levels such as "normal", "attention", "warning" and "danger" according to the threshold range. When the comprehensive abnormality index is less than T1, it is a normal state; between T1 and T2, it is an attention state; between T2 and T3, it is a warning state; and greater than T3, it is a danger state.

[0049] In practice, the system queries the early warning rule base based on the equipment status level to obtain the corresponding early warning action configuration. The early warning rule base is a data table that stores the mapping relationship between status levels and early warning actions. The early warning action configuration includes information such as the specific operation instructions to be triggered, the notification message template, and the target recipient. For example, the early warning action corresponding to the "Attention" level is to log and send an email to the maintenance engineer; the early warning action corresponding to the "Warning" level is to send an SMS notification and generate a maintenance work order; and the early warning action corresponding to the "Danger" level is to trigger an audible and visual alarm and execute an automatic shutdown procedure. Finally, the early warning action configuration is executed to generate an early warning signal and distribute it to the monitoring terminal. The early warning signal generation module assembles a specific alarm message according to the early warning action configuration. The message content includes information such as the equipment number, abnormal time, status level, and specific abnormal characteristics. The distribution process pushes the early warning signal to the designated monitoring terminal through message middleware or enterprise service bus. The monitoring terminals include the central control room large screen, engineer mobile terminals, and enterprise manufacturing execution systems.

[0050] In some embodiments, feature vector similarity calculation can employ multiple similarity measurement algorithms for complementary verification. For example, cosine similarity can be calculated simultaneously with Euclidean distance similarity, where Euclidean distance similarity focuses on the absolute distance between vectors. The two similarity results are then combined for a comprehensive judgment. The determination of the normal state baseline can utilize a dynamic update mechanism, adaptively adjusting the feature vector reference point of the normal state baseline according to changes in the equipment's operating cycle, thus avoiding false alarms caused by baseline drift.

[0051] 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.

[0052] 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 smart monitoring system for the status of industrial equipment based on big data, characterized in that, The system includes: An industrial equipment operation data acquisition module is used to continuously capture raw monitoring data streams from a sensor network deployed on industrial equipment, and to perform data cleaning and quality verification on the raw monitoring data streams to form a standardized set of equipment operation data. The equipment status feature abstraction module is used to perform multi-scale feature transformation on the standardized equipment operation data set, generate a time-series correlated equipment status feature image, and perform feature reduction on the equipment status feature image to obtain an equipment status feature summary. The monitoring knowledge graph evolution module is used to parse the feature association rules hidden in the equipment status feature summary, construct the equipment status feature association network based on the feature association rules, and evolve the equipment status feature association network into an industrial equipment status monitoring knowledge graph through a dynamic update mechanism. The intelligent early warning decision module is used to project real-time acquired industrial equipment operation data onto the industrial equipment status monitoring knowledge graph for similarity matching and deviation calculation, and trigger hierarchical early warning decisions based on the matching and calculation results.

2. The intelligent monitoring system for industrial equipment status based on big data as described in claim 1, characterized in that, The raw monitoring data stream includes vibration signals, temperature readings, pressure values, current and voltage waveforms, equipment control parameters, and ambient temperature and humidity data.

3. The intelligent monitoring system for industrial equipment status based on big data as described in claim 1, characterized in that, The process of cleaning and quality verification of the original monitoring data stream to form a standardized set of equipment operation data specifically includes: The original monitoring data stream is subjected to outlier removal and missing value imputation to obtain pre-processed monitoring data; The pre-processed monitoring data is subjected to dimensional unification and numerical normalization operations to obtain standardized monitoring data. Standardized monitoring data are aligned and integrated according to equipment number and timestamp to form a standardized set of equipment operation data.

4. The intelligent monitoring system for industrial equipment status based on big data as described in claim 1, characterized in that, The step of performing multi-scale feature transformation on the standardized equipment operation data set to generate a time-series correlated equipment state feature image specifically includes: The standardized equipment operation data set is subjected to time-domain statistical feature extraction, frequency-domain energy feature decomposition, and time-frequency domain joint feature analysis to obtain a multi-scale feature set. By sliding and stitching together various features from the multi-scale feature set according to time windows, a device status feature image is constructed.

5. The intelligent monitoring system for industrial equipment status based on big data as described in claim 1, characterized in that, The step of reducing the device state feature image to obtain the device state feature summary specifically includes: Calculate the correlation strength between each feature dimension in the device status feature image and the device health status; The feature dimensions are sorted according to their correlation strength, and feature dimensions with correlation strength exceeding a preset threshold are selected. The selected feature dimensions are recombined to generate a device status feature summary with reduced dimensionality.

6. The intelligent monitoring system for industrial equipment status based on big data as described in claim 1, characterized in that, The step of parsing the feature association rules implicit in the device status feature summary and constructing the device status feature association network based on the feature association rules specifically includes: The device status feature summary was analyzed using association rule mining methods to discover frequently co-occurring feature combination patterns. An initial device state feature association network is constructed using features as nodes and the co-occurrence strength between features as edges. Redundant edge pruning and core community discovery are performed on the initial device state feature association network to obtain a simplified device state feature association network.

7. The intelligent monitoring system for industrial equipment status based on big data as described in claim 1, characterized in that, The process of evolving the equipment state feature association network into an industrial equipment condition monitoring knowledge graph through a dynamic update mechanism specifically includes: New device status feature summaries are continuously injected into the device status feature association network; Adjust the weights of network nodes and edges based on the newly injected data, and identify newly added feature relationships; After iterative updates, an industrial equipment condition monitoring knowledge graph containing the evolutionary patterns of equipment states is formed.

8. The intelligent monitoring system for industrial equipment status based on big data as described in claim 1, characterized in that, The step of projecting real-time acquired industrial equipment operation data onto the industrial equipment condition monitoring knowledge graph for similarity matching and deviation calculation specifically includes: The real-time industrial equipment operation data is subjected to the same feature abstraction process as historical data to obtain a summary of real-time equipment status features. Calculate the similarity between the real-time equipment status feature summary and the feature vectors of each typical status pattern in the industrial equipment status monitoring knowledge graph; Calculate the degree of deviation of the real-time device state feature summary from the normal state baseline in the knowledge graph.

9. The intelligent monitoring system for industrial equipment status based on big data as described in claim 1, characterized in that, The specific provisions for triggering tiered early warning decisions based on matching and calculation results include: The matching calculation results between the real-time equipment status feature summary and the industrial equipment status monitoring knowledge graph are obtained, and the matching calculation results include feature vector similarity and feature deviation. A weighted fusion method is used to synthesize feature vector similarity and feature deviation into a comprehensive anomaly index; The comprehensive anomaly index is compared with multiple preset level thresholds step by step to classify the equipment status level; Query the warning rule library based on the device status level to obtain the corresponding warning action configuration; Configure early warning actions, generate early warning signals, and distribute them to monitoring terminals.

10. The intelligent monitoring system for industrial equipment status based on big data as described in claim 5, characterized in that, The calculation of the correlation strength between each feature dimension in the device status feature image and the device health status specifically includes: Collect historical equipment operation data and label the equipment health status to form a training dataset; For each feature dimension in the device state feature image, extract the numerical sequence of the feature dimension in the training dataset; Calculate the statistical correlation measure between the numerical sequence of each feature dimension and the device health status label sequence; The correlation strength value for each feature dimension is determined based on the absolute value of the statistical correlation measure. The association strength values ​​are normalized to obtain standardized association strength weights.

Images