An industrial production internet of things data anomaly detection method, medium and system

Abstract

The application provides an industrial production Internet of Things data anomaly detection method, medium and system, belongs to the technical field of industrial production Internet of Things, and establishes a multidimensional data set by collecting industrial equipment sensor data and preprocessing, constructs a dimension reduction characteristic data set by using principal component analysis and mutual information algorithm, compensates for data missing by using a virtual sensor algorithm to form an extended data set, converts the microscopic state of massive high-dimensional sensor data into a macroscopic thermodynamic parameter based on a statistical mechanics law to establish an analog statistical mechanics anomaly analysis model, calculates a statistical mechanics characteristic vector by simulating the equipment running state through a virtual particle system, identifies an abnormal mode and performs hierarchical marking by adopting a state evaluation model and a dynamic threshold function, and constructs a feedback optimization mechanism to continuously improve system performance, so that the technical problem that traditional algorithms cannot effectively detect anomalies due to the dimensional curse caused by the extremely high dimension of industrial equipment sensor data is solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of industrial production Internet of Things, and in particular, relates to an industrial production Internet of Things data anomaly detection method, medium and system. BACKGROUND

[0002] Device anomaly detection in the industrial Internet of Things environment is a key technology to ensure the safety of industrial production and improve the efficiency of device operation. Traditional anomaly detection methods mainly rely on statistical analysis, machine learning and pattern recognition techniques, and establish a normal mode baseline of device operation to identify abnormal behavior that deviates from the normal state. These methods have been widely used in industrial scenarios such as power system monitoring, petrochemical device safety protection and manufacturing equipment maintenance. However, traditional anomaly detection techniques have significant defects when dealing with high-dimensional sensor data. When the dimension of sensor data increases, data points become sparsely distributed in high-dimensional space, causing traditional distance metrics to fail. At the same time, the computational complexity of high-dimensional data increases exponentially with the number of dimensions, making it difficult for traditional algorithms to complete the anomaly detection task within a reasonable time. In the prior art, due to the characteristics of high dimension, complex correlation and large data volume of multi-dimensional sensor data generated by industrial equipment, traditional anomaly detection algorithms cannot effectively convert massive high-dimensional data information into processable low-dimensional feature representation, lack effective dimension compression and information aggregation mechanisms, resulting in huge consumption of computing resources and poor detection effect. That is, the prior art has the technical problem that the dimension of sensor data generated by industrial equipment is extremely high, and traditional algorithms are prone to dimension curse phenomenon in high-dimensional space, resulting in exponential growth of computational complexity and inability to achieve effective anomaly detection. SUMMARY

[0003] Therefore, the present application provides an industrial production Internet of Things data anomaly detection method, medium and system, which can solve the technical problem that the dimension of sensor data generated by industrial equipment is extremely high in the prior art, and traditional algorithms are prone to dimension curse phenomenon in high-dimensional space, resulting in exponential growth of computational complexity and inability to achieve effective anomaly detection.

[0004] The application is implemented as follows: the first aspect of the application provides an industrial production Internet of Things data anomaly detection method, including collecting sensor data generated by an industrial device and preprocessing to establish a multi-dimensional sensor data set, performing dimension reduction processing on the multi-dimensional sensor data set by principal component analysis, and using a mutual information algorithm to select features to establish a dimension-reduced feature data set, constructing a virtual sensor algorithm to generate virtual sensor data from the dimension-reduced feature data set to form an extended sensor data set, establishing an analog statistical mechanics anomaly analysis model based on statistical mechanics laws to map the extended sensor data set to a statistical mechanics space to calculate a statistical mechanics feature vector, using an industrial device state evaluation model to analyze the statistical mechanics feature vector, calculating an anomaly detection threshold by a dynamic threshold adjustment function to identify a device operation anomaly mode, establishing an anomaly mode classification system to classify and label the device operation anomaly mode, calculating an anomaly level by an anomaly severity evaluation function to generate an anomaly detection report, constructing a feedback optimization mechanism to adjust detection parameters according to the anomaly detection report, and optimizing the performance of the industrial device state evaluation model by a parameter self-adaptive updating algorithm to realize continuous improvement of the anomaly detection system.

[0005] The virtual sensor algorithm is specifically a calculation method for inferring missing or faulty sensor data based on existing sensor data, which infers the data value of a virtual sensor by establishing a correlation model between sensors and using the data of normal sensors to calculate the data value of the virtual sensor, thereby making up for the data missing problem in the sensor network.

[0006] The virtual sensor algorithm uses a virtual collection router model, which is a data path optimization model based on graph theory and network topology analysis, and identifies the weak links and data transmission bottlenecks in the network by constructing a connection relationship graph between sensor nodes and calculating the shortest path weight between nodes from the aspect of the sensor network topology structure.

[0007] When the virtual collection router model detects a sensor node failure or a data transmission interruption, it automatically reconstructs the data transmission path and generates virtual sensor node data according to the redundant path and backup node information of the network topology structure, and makes up for the missing data by using adjacent node data interpolation and time series data extrapolation algorithms.

[0008] The analog statistical mechanics anomaly analysis model is specifically a device state analysis model constructed based on statistical mechanics theory and thermodynamic equilibrium principle, which maps the multi-dimensional operating parameters of the industrial device into the microscopic particle state distribution and macroscopic thermodynamic parameters in the statistical mechanics space, and simulates the device operating state by establishing a virtual particle ensemble and a thermodynamic system.

[0009] The simulation statistical mechanics anomaly analysis model maps the random fluctuations of the device operating parameters as the Brownian motion and thermal fluctuations of virtual particles, maps the device energy consumption changes as the internal energy and enthalpy changes of the virtual thermodynamic system, and maps the device operating stability as the entropy value and free energy gradient of the virtual system.

[0010] The simulation statistical mechanics anomaly analysis model analyzes the regularity of the device operating data through the Boltzmann distribution law and Maxwell statistical distribution, calculates the partition function, correlation function and fluctuation dissipation relation of the virtual particle ensemble, and identifies the abnormal mode by detecting the entropy increase anomaly, phase transition phenomenon and statistical distribution deviation of the virtual thermodynamic system.

[0011] The statistical mechanics feature vector is a vector set representing the device operating state calculated by the simulation statistical mechanics anomaly analysis model, and includes the microscopic state distribution characteristics, macroscopic thermodynamic equilibrium state and statistical fluctuation law in the device operating process.

[0012] The device operating anomaly mode is a device behavior mode deviating from the normal operating state identified by the dynamic threshold adjustment function, and includes the abnormal occurrence time, abnormal type and abnormal degree information.

[0013] The anomaly severity assessment function is used to quantify the severity of the abnormal situation, and the input includes the abnormal deviation degree, duration, influence range and historical abnormal frequency, and the output is an abnormal grade value, which divides the device operating anomaly mode into four grades of slight, medium, severe and urgent.

[0014] The anomaly detection report is a comprehensive report document containing the abnormal grade value and detailed information of the device operating anomaly mode, and is used to guide the device maintenance and fault handling decision.

[0015] The parameter adaptive update algorithm is an optimization algorithm for automatically adjusting the model parameters according to the detection effect, receives the anomaly detection report as input, monitors the accuracy and recall rate of anomaly detection, and automatically adjusts the model weight and threshold parameters when the detection performance decreases.

[0016] The industrial equipment state evaluation model is a calculation model for comprehensively evaluating the operation state of the industrial equipment, takes a statistical mechanics feature vector as input, calculates the overall health state of the equipment through a composite evaluation algorithm, and provides a state benchmark for anomaly detection.

[0017] The dynamic threshold adjustment function is a function for dynamically adjusting the anomaly detection threshold according to the equipment operation environment and historical data, receives the output result of the industrial equipment state evaluation model, and automatically adjusts the detection sensitivity to reduce false positives and false negatives.

[0018] The second aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores program instructions, the program instructions are run in the computer, and are used for executing the above-mentioned industrial production Internet of Things data anomaly detection method.

[0019] The third aspect of the present application provides an industrial production Internet of Things data anomaly detection system, which comprises the above-mentioned computer readable storage medium, the system is any one of a computer, a server and a single-chip microcomputer, the computer readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer readable storage medium.

[0020] The present application realizes a substantial compression of data dimensions by constructing a pseudo-statistical mechanics anomaly analysis model, converting the microstate distribution of massive high-dimensional sensor data into a few macroscopic thermodynamic parameters, and effectively solving the dimension curse problem of high-dimensional sensor data. The pseudo-statistical mechanics model draws on the successful experience of statistical physics in dealing with a large number of microscopic particle systems, regards each sensor data point as a virtual particle microstate, and converts the complex high-dimensional data correlation into a few macroscopic thermodynamic parameters such as temperature, entropy, internal energy, and free energy through statistical mechanics theory. This conversion process naturally has the ability to compress dimensions, can compress thousands of dimensions of sensor data into a few thermodynamic parameters, while retaining the core statistical characteristics and anomaly information of the data. By establishing a virtual particle system and a thermodynamic system to simulate the running state of the equipment, using the Boltzmann distribution law and Maxwell statistical distribution to analyze the regularity of the equipment running data, and detecting the entropy increase anomaly, phase change phenomenon and statistical distribution deviation of the virtual thermodynamic system, this dimension reduction method based on physical principles has a good theoretical basis and mathematical interpretability. In summary, the present application solves the technical problem of the high dimension of sensor data generated by industrial equipment in the background art, the dimension curse phenomenon of traditional algorithms in high-dimensional space, and the exponential growth of calculation complexity, which makes it impossible to realize effective anomaly detection. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a flowchart of the method of the present application.

[0022] Figure 2 is a schematic diagram of the industrial equipment state evaluation model network structure involved in the present application.

[0023] Figure 3 is a virtual sensor network topology structure diagram in Example 2.

[0024] Figure 4 is a statistical mechanics feature vector distribution comparison diagram in Example 2.

[0025] Figure 5 is an anomaly detection result time series analysis diagram in Example 2. DETAILED DESCRIPTION

[0026] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0027] As shown in Figure 1 is a flowchart of an industrial production Internet of Things data anomaly detection method provided by the first aspect of the present application, the method includes the following steps:

[0028] S01, collect sensor data generated by industrial equipment, preprocess the collected sensor data, including data cleaning and preliminary filtering of outliers, and establish a multi-dimensional sensor data set;

[0029] S02, principal component analysis is used to reduce the dimension of the multi-dimensional sensor data set, the main components are extracted through eigenvalue decomposition, and feature selection is performed using mutual information algorithm, and a reduced dimension feature data set is established;

[0030] S03, a virtual sensor algorithm is constructed to generate virtual sensor data for the reduced dimension feature data set, and the data missing is made up from the sensor network topology through a virtual collection router model, forming an extended sensor data set;

[0031] S04, an analog statistical mechanics anomaly analysis model is established based on the laws of statistical mechanics, the extended sensor data set is mapped to the statistical mechanics space, the regularity of the equipment operation data is analyzed through the microscopic state statistical distribution and the macroscopic thermodynamic parameters, and the statistical mechanics feature vector is calculated;

[0032] S05, an industrial equipment state evaluation model is used to analyze the mechanics feature vector, an abnormal detection threshold is calculated through a dynamic threshold adjustment function, and an equipment operation abnormal mode is identified;

[0033] S06, an abnormal mode classification system is established, the detected equipment operation abnormal mode is classified and labeled, the abnormal level is calculated through an abnormal severity evaluation function, and an abnormal detection report is generated;

[0034] S07, a feedback optimization mechanism is constructed, the detection parameters are adjusted according to the abnormal detection report, the performance of the industrial equipment state evaluation model is optimized through a parameter self-adaptive updating algorithm, and the continuous improvement of the abnormal detection system is realized.

[0035] The virtual sensor algorithm is a calculation method for inferring missing or faulty sensor data based on existing sensor data, which establishes an association model between sensors, uses the data of normal sensors to calculate the data value of virtual sensors, and thus makes up for the data missing problem in the sensor network.

[0036] The virtual collection router model is a data path optimization model based on graph theory and network topology analysis, which identifies weak links and data transmission bottlenecks in the network from the aspect of sensor network topology structure by constructing the connection relationship graph between sensor nodes and calculating the shortest path weight between nodes. When detecting sensor node failure or data transmission interruption, the virtual collection router model automatically reconstructs the data transmission path and generates virtual sensor node data according to the redundant path and backup node information of the network topology structure, compensates for missing data through adjacent node data interpolation and time series data extrapolation algorithm, and dynamically adjusts network load balancing and data transmission priority to ensure reliable transmission and integrity of critical sensor data.

[0037] The pseudo-statistical mechanics anomaly analysis model is an equipment state analysis model based on statistical mechanics theory and thermodynamic equilibrium principle, which maps multi-dimensional operation parameters of industrial equipment to microscopic particle state distribution and macroscopic thermodynamic parameters in statistical mechanics space, simulates equipment operation state by establishing virtual particle ensemble and thermodynamic system, maps random fluctuations of equipment operation parameters to Brownian motion and thermal fluctuations of virtual particles, maps changes of equipment energy consumption to internal energy and enthalpy changes of virtual thermodynamic system, and maps equipment operation stability to entropy and free energy gradient of virtual system. The pseudo-statistical mechanics anomaly analysis model analyzes the regularity of equipment operation data by Boltzmann distribution law and Maxwell statistical distribution, calculates the partition function, correlation function and fluctuation dissipation relationship of virtual particle ensemble, and identifies abnormal patterns by detecting entropy increase anomaly, phase change phenomenon and statistical distribution deviation of virtual thermodynamic system when equipment operation deviates from normal working condition, and establishes the mapping relationship between statistical mechanics parameters and equipment health status.

[0038] The statistical mechanics feature vector is a vector set representing the equipment operation state calculated by the pseudo-statistical mechanics anomaly analysis model, which contains microscopic state distribution characteristics, macroscopic thermodynamic equilibrium state and statistical fluctuation law in the equipment operation process.

[0039] The industrial equipment state evaluation model is a calculation model for comprehensive evaluation of industrial equipment operation state, which takes the statistical mechanics feature vector as input and calculates the overall health status of the equipment by a composite evaluation algorithm to provide a state benchmark for anomaly detection.

[0040] The dynamic threshold adjustment function is a function for dynamically adjusting the anomaly detection threshold according to the equipment operation environment and historical data, which receives the output results of the industrial equipment state evaluation model and automatically adjusts the detection sensitivity to reduce false positives and false negatives.

[0041] The device operation abnormal mode is a device behavior mode deviating from the normal operation state identified by the dynamic threshold adjustment function, containing abnormal occurrence time, abnormal type and abnormal degree information.

[0042] The abnormal severity assessment function is used to quantify the severity of the abnormal situation, and the input includes the abnormal deviation degree, duration, impact range and historical abnormal frequency. The output is an abnormal level value, which divides the device operation abnormal mode into four levels: slight, medium, severe and urgent.

[0043] The abnormal detection report is a comprehensive report document containing the abnormal level value and the detailed information of the device operation abnormal mode, which is used to guide the device maintenance and fault handling decision.

[0044] The parameter adaptive update algorithm is an optimization algorithm that automatically adjusts the model parameters according to the detection effect, receives the abnormal detection report as input, monitors the accuracy and recall rate of abnormal detection, and automatically adjusts the model weight and threshold parameter when the detection performance decreases, ensuring the long-term stability and accuracy of the detection system.

[0045] The industrial equipment state evaluation model adopts a state fusion network based on attention mechanism for optimization, which dynamically allocates attention weights according to the importance and relevance of sensor data. The calculation of attention weights needs to be determined according to three key parameters: sensor data dimension, data acquisition frequency and sensor reliability coefficient. The sensor data dimension affects the number of attention head configuration, the data acquisition frequency determines the length of the time window setting, and the sensor reliability coefficient adjusts the weight allocation proportion of each sensor data.

[0046] The structure of the state fusion network is a multi-layer attention encoder architecture, including data embedding layer, multi-head self-attention layer, feedforward neural network layer and output mapping layer. The data embedding layer converts the mechanical feature vector into a unified vector representation. The multi-head self-attention layer captures the complex correlation between sensor data by parallel computing multiple attention heads. The feedforward neural network layer performs nonlinear transformation and feature extraction on the attention output. The output mapping layer maps high-dimensional features to device state evaluation results. The whole network uses residual connection and layer normalization technology to improve training stability and convergence speed.

[0047] The training data set establishment step of the state fusion network includes collecting equipment operation data in different industrial scenarios, covering sensor data records of normal operation state, various abnormal states and fault modes, performing quality evaluation and noise filtering on the collected raw data, removing obviously incorrect and severely missing data samples, classifying and labeling the data according to the equipment type and operation condition, establishing a supervised learning data set containing equipment state labels, dividing the data set into training set, validation set and test set according to time sequence, ensuring uniform distribution of data in different time periods, performing normalization processing and data enhancement operation on the training data, and generating more diversified training samples to improve the model generalization ability.

[0048] The state fusion network training step includes initializing network parameters and optimizer configuration, setting learning rate scheduling strategy and regularization parameter, using batch gradient descent algorithm for model training, calculating model output and loss function value through forward propagation, calculating parameter gradient and updating network weight using back propagation algorithm, periodically evaluating model performance on validation set during training, monitoring the change trend of training loss and validation accuracy, using early stopping strategy to avoid overfitting when validation performance no longer improves, using learning rate decay and gradient clipping techniques to improve training stability, and performing final performance evaluation on the test set after training to ensure that the model generalization ability on unseen data meets the expected requirements.

[0049] The state weight adjustment function is used to adjust the attention weight distribution of the state fusion network, based on sensor data variance, data correlation coefficient, abnormal detection historical accuracy, and current equipment load rate, to obtain an attention adjustment coefficient value, when the attention adjustment coefficient value belongs to different ranges, different weight distribution strategies are used to adjust the attention weight parameters of the state fusion network, including four adjustment ranges, the first interval value is calculated by square root of the product of the sensor data variance and the data correlation coefficient, the second interval value is calculated by the geometric mean of the abnormal detection historical accuracy and the current equipment load rate, the third interval value is calculated by the harmonic mean of the sensor data variance and the abnormal detection historical accuracy, when the attention adjustment coefficient value is less than the first interval value, a conservative weight distribution strategy is used to reduce the dynamic adjustment range of the attention weight, when the attention adjustment coefficient value is between the first interval value and the second interval value, a balanced weight distribution strategy is used to maintain the moderate adjustment strength of the attention weight, when the attention adjustment coefficient value is between the second interval value and the third interval value, an aggressive weight distribution strategy is used to enhance the response sensitivity of the attention weight, and when the attention adjustment coefficient value is greater than the third interval value, an aggressive weight distribution strategy is used to maximize the adaptive adjustment ability of the attention weight.

[0050] The specific implementation of the above steps is described in detail below.

[0051] The implementation of step S01 is to obtain raw data from various sensors of industrial equipment through a distributed data acquisition system. The sensor types include temperature, pressure, vibration, current, and rotational speed, among other physical quantity monitoring devices. The data acquisition process uses a time synchronization mechanism to ensure the time consistency of different sensor data. The acquisition frequency is set to a suitable value within the range of 1 Hz to 1000 Hz according to the type of equipment. In the preprocessing stage, data cleaning is performed first. A sliding window smoothing filter algorithm is used to remove high-frequency noise, and the window length is set to 5 to 15 sampling points. The preliminary filtering of outliers uses a combination of the 3σ criterion based on statistical distribution and the quartile distance method. When a data point deviates from the mean by more than 3 times the standard deviation or exceeds the quartile distance by 1.5 times, it is marked as an outlier. The establishment of the multi-dimensional sensor data set is achieved through data alignment and interpolation completion operations to ensure the integrity and consistency of the data in the time axis. The purpose of this step is to provide high-quality basic data for subsequent analysis, and to improve data quality and reduce the impact of noise on the accuracy of anomaly detection through preprocessing operations.

[0052] The implementation of step S02 is to perform dimensionality reduction processing and feature selection on the multi-dimensional sensor data set. Principal component analysis dimensionality reduction processing uses the singular value decomposition algorithm to calculate the eigenvalues and eigenvectors of the covariance matrix, and retains the principal components with a cumulative contribution rate of 85% to 95% as the reduced features. The eigenvalue decomposition process is implemented through the Jacobi iteration algorithm or the QR decomposition algorithm, and the iteration convergence threshold is set to When the mutual information algorithm is used for feature selection, the mutual information value between each feature dimension and the target variable is calculated, the maximum information coefficient method is used to quantify the nonlinear correlation, and the top 70% to 90% of the feature variables are retained. The dimensionality of the reduced feature data set is usually reduced from dozens of dimensions of the original data to 10 to 20 dimensions, which not only maintains the main information of the data but also reduces the computational complexity. The role of this step is to eliminate data redundancy and reduce the curse of dimensionality problem, improving the computational efficiency and model generalization ability of subsequent algorithms.

[0053] The implementation of step S03 is to construct a virtual sensor algorithm to generate data for missing or faulty sensors. The virtual sensor algorithm uses machine learning methods such as multiple linear regression, radial basis function networks, or support vector regression to establish a correlation model between sensors. The training process uses historical normal operation data, and the model parameters are optimized by least squares method or gradient descent algorithm, with the model training error threshold set to less than 5% root mean square error. The virtual acquisition router model is based on graph theory algorithm to construct the sensor network topology graph, with nodes representing sensor positions and edge weights representing the correlation strength between sensors. The shortest path calculation uses Dijkstra algorithm or Floyd-Warshall algorithm, with the path weight threshold set to a correlation coefficient greater than 0.6 between sensor pairs. When a sensor failure is detected, the model automatically selects a backup data source based on network redundancy paths, and generates virtual sensor data through linear interpolation, spline interpolation, or time series extrapolation algorithms. The extended sensor data set is formed by fusing the original data and virtual data, with data integrity reaching more than 95%. The purpose of this step is to solve the problem of data missing caused by sensor failure or communication interruption, and to ensure the continuity and integrity of the input data of the anomaly detection algorithm.

[0054] The implementation of step S04 is to establish a statistical mechanics anomaly analysis model to map sensor data into statistical mechanics space for analysis. The model treats device operating parameters as a virtual particle system, with random fluctuations in parameters corresponding to Brownian motion of particles, and changes in device energy consumption corresponding to changes in system internal energy. Statistical distribution analysis uses Boltzmann distribution and Maxwell distribution to describe the velocity and energy distribution of virtual particles, and the distribution parameters are fitted by maximum likelihood estimation method. Partition function calculation uses Monte Carlo method for numerical integration, with the sampling number set to to times to ensure calculation accuracy. Correlation function calculation obtains the time and space correlation of the system through autocorrelation and cross-correlation analysis, with the correlation length threshold set to 0.1 to 0.3. Fluctuation dissipation relation describes the relationship between system response and thermal fluctuations through Einstein relation. The statistical mechanics feature vector contains physical quantities such as system entropy, free energy gradient, partition function value, and correlation length, with the vector dimension usually being 8 to 15. The purpose of this step is to understand the device operation rules from a physical perspective, and to provide physical basis and theoretical support for anomaly detection through statistical mechanics theory.

[0055] The implementation of step S05 is to analyze the statistical mechanics feature vector and calculate the anomaly detection threshold using an industrial equipment state evaluation model. The state evaluation model uses a neural network architecture based on an attention mechanism, with the statistical mechanics feature vector as input and the equipment health state score as output. The dynamic threshold adjustment function dynamically adjusts the detection threshold according to factors such as equipment operating environment temperature, load conditions, and maintenance history. The environmental temperature correction coefficient is 0.8 to 1.2, and the load correction coefficient is 0.9 to 1.1. The anomaly detection threshold is determined by the statistical control chart method, using a combination of moving average control charts and exponentially weighted moving average control charts, with a control limit set to 3 times the standard deviation. The equipment operation anomaly pattern recognition uses pattern matching and clustering analysis algorithms, including K-means clustering, hierarchical clustering, or density clustering methods, with the number of clusters determined to be 3 to 8 categories based on historical anomaly types. The purpose of this step is to establish a quantitative evaluation standard for equipment state, adapt to different working conditions through a dynamic threshold mechanism, and improve the accuracy and robustness of anomaly detection.

[0056] The implementation of step S06 is to establish an anomaly pattern classification system to classify and grade the detected anomalies. The anomaly pattern classification uses a supervised learning method to train a classifier, with feature input including equipment state parameters during anomaly occurrence, anomaly duration, anomaly deviation degree, etc. The classification algorithm can choose support vector machines, random forests, or deep neural networks, with a classification accuracy threshold set to 90% or higher. The anomaly severity evaluation function considers the anomaly deviation degree weight coefficient 0.4, the duration weight coefficient 0.3, the impact range weight coefficient 0.2, and the historical frequency weight coefficient 0.1 for weighted calculation. The anomaly level classification uses a four-level system, with a score of 0 to 25 for minor anomalies, a score of 25 to 50 for moderate anomalies, a score of 50 to 75 for severe anomalies, and a score of 75 to 100 for urgent anomalies. The anomaly detection report generation uses a templated format, including information such as anomaly time, location, type, level, impact range, and treatment recommendations. The role of this step is to structure the analysis and level classification of anomalies, providing quantitative basis and processing priority guidance for maintenance decisions.

[0057] The implementation of step S07 is to construct a feedback optimization mechanism to adjust the model parameters according to the detection effect. The parameter adaptive updating algorithm monitors the performance indicators of anomaly detection such as accuracy, recall rate, and false positive rate, and the accuracy threshold is set to 85%, the recall rate threshold is set to 80%, and the false positive rate threshold is set to 15%. When the detection performance is lower than the threshold, the algorithm automatically adjusts the model weight parameters, detection threshold parameters, and feature selection parameters. The weight update adopts gradient descent algorithm or genetic algorithm for optimization, and the learning rate is set to 0.001 to 0.01, and the iteration number is set to 100 to 1000 times. The threshold parameter adjustment adopts grid search or Bayesian optimization method to find the optimal parameter combination, and the search space is set to a reasonable range according to historical experience. The feature selection parameters are evaluated by cross-validation method to evaluate the detection effect of different feature combinations, and the feature subset with the best performance is selected. The purpose of this step is to realize the adaptive optimization and continuous improvement of the detection system, and to maintain the stable performance of the system under different working conditions and time conditions through the feedback mechanism.

[0058] The detailed structure of the industrial equipment state evaluation model adopts a state fusion network architecture based on attention mechanism, including data embedding layer, multi-head self-attention layer, feedforward neural network layer and output mapping layer. The data embedding layer converts the input statistical mechanics feature vector into a unified high-dimensional vector representation through linear transformation and position encoding, and the embedding dimension is set to 128 to 512 dimensions. The multi-head self-attention layer adopts a parallel computing architecture, and the number of attention heads is determined according to the input feature dimension, which is 4 to 16. Each attention head independently calculates the query, key and value matrices, and captures the complex correlation between features through the scaled dot-product attention mechanism. The feedforward neural network layer adopts a two-layer fully connected network structure, the number of neurons in the first layer is 2 to 4 times the embedding dimension, and the activation function adopts ReLU or GELU function, and the number of neurons in the second layer is the same as the embedding dimension. The output mapping layer maps the high-dimensional features to the equipment state evaluation results through linear transformation, and the output dimension is determined according to the number of state categories. The whole network adopts residual connection and layer normalization technology to improve the training stability, and the Dropout probability is set to 0.1 to 0.3 to prevent overfitting.

[0059] The detailed steps of training data set establishment include data collection, quality evaluation, classification annotation, data division and preprocessing operation. The data collection stage collects equipment operation data from different industrial scenes, covering typical equipment in steel, chemical, power, mechanical manufacturing and other industries, with a data time span of not less than 6 months, containing complete records of normal operation, various abnormalities and fault modes. Quality evaluation is carried out through data integrity check, consistency verification and noise level evaluation, and data samples with missing rate exceeding 20%, time stamp error and obvious abnormality are excluded. Classification annotation is manually annotated by domain experts according to equipment maintenance records and fault reports, and a supervised learning data set containing normal, slight abnormal, medium abnormal and severe abnormal four state labels is established. Data division divides the data set into training set 60%, validation set 20% and test set 20% according to the time sequence principle, ensuring that the data distribution of different time periods uniformly covers various working conditions. The preprocessing operation includes data normalization, standardization and data enhancement, normalization adopts maximum and minimum value scaling or Z-score standardization method, and data enhancement generates more training samples through time warping, noise injection and sliding window sampling.

[0060] The industrial equipment state evaluation model is suitable for solving the technical problems of the present application because it can effectively fuse multi-dimensional sensor data and accurately evaluate the equipment running state. Compared with the traditional abnormal detection method based on single threshold judgment, the model can dynamically allocate the importance weight of each sensor data through the attention mechanism, and can adaptively identify the key fault features. Compared with the existing abnormal detection technology based on statistical analysis, the model does not depend on the preset statistical distribution assumption, and automatically learns the complex mode of the equipment state through the deep learning method. The advantage of the model is that it can handle the correlation relationship of high-dimensional and nonlinear sensor data, and through the multi-head attention mechanism, it can simultaneously focus on the local features and global trends of the time series, and has stronger feature expression ability and generalization performance than the traditional support vector machine or random forest method.

[0061] The virtual sensor algorithm is suitable for solving the problem of sensor failure and data missing, and its technical advantage is that it can reconstruct the missing data through multi-sensor information fusion. Compared with the traditional data interpolation method, the virtual sensor algorithm establishes a prediction model based on the physical correlation between sensors, which can more accurately estimate the values of missing sensors. Compared with the existing method based on time series prediction, the algorithm considers the spatial correlation between sensors, and establishes a more accurate sensor correlation model through multivariate regression or neural network method, which can maintain the continuity and reliability of the data in the case of sensor failure.

[0062] The technical advantage of the virtual acquisition router model is to optimize the data transmission path and fault recovery mechanism from the perspective of network topology. Compared with the traditional fixed routing transmission mode, this model dynamically calculates the optimal transmission path through graph theory algorithm, and can automatically switch to the backup path when the network node fails. Compared with existing network redundancy technologies, the virtual acquisition router model not only realizes path backup, but also dynamically adjusts the data transmission priority according to the network load and transmission quality, ensuring the real-time and integrity of critical sensor data.

[0063] The technical advantage of the pseudo-statistical mechanics anomaly analysis model is to understand the device operation law from the perspective of physics, while realizing efficient data dimension reduction processing. This model converts the microscopic fluctuation information of massive sensor data points into a few macroscopic thermodynamic parameters through statistical mechanics principles, which are used to describe the overall operation state of the device, realizing effective mapping from high-dimensional complex data to low-dimensional physical characteristics. Compared with traditional detection methods based on empirical thresholds, this model not only provides a theoretical basis, but also compresses hundreds of sensor measurements into a dozen statistical mechanics feature vectors through statistical aggregation effect, greatly reducing the computational complexity of data processing. Compared with existing mathematical dimension reduction methods such as principal component analysis, the dimension reduction process of the pseudo-statistical mechanics model has clear physical meaning, describes the device state through statistical laws such as Boltzmann distribution and Maxwell distribution, and converts the complex multi-dimensional sensor correlation into interpretable physical quantities such as entropy, partition function, and correlation length, which has better detection ability and interpretability in the case of small sample and unknown abnormal type.

[0064] The key technical ideas of the application include four aspects of virtual sensor data reconstruction, pseudo-statistical mechanics state modeling, attention mechanism state fusion and adaptive parameter optimization. The advantage of the virtual sensor data reconstruction technology over the traditional interpolation method is that the physical correlation between sensors is considered, the missing data is reconstructed by multivariate statistical model and machine learning method, and the physical consistency and logical rationality of the data are ensured. The pseudo-statistical mechanics state modeling technology maps the device running state to the statistical physics space, which has a solid theoretical basis compared with the experience threshold method, and identifies abnormalities through entropy change and phase change, providing an interpretable physical mechanism. The attention mechanism state fusion technology dynamically allocates the importance weight of different sensor data, which can adaptively identify key fault features compared with the fixed weight method, improving the accuracy and robustness of anomaly detection. The adaptive parameter optimization technology automatically adjusts the model parameters according to the detection effect, which can adapt to different working conditions and time changes compared with the static parameter setting, maintaining the long-term stability of the detection system. The synergistic effect of the four technical ideas forms a complete anomaly detection solution, the virtual sensor technology ensures the data integrity, the statistical mechanics modeling provides theoretical support, the attention mechanism realizes intelligent feature fusion, and the adaptive optimization ensures the system stability, which has higher detection accuracy, stronger adaptability and better engineering practicability compared with the existing technology.

[0065] It should be noted that the application solves the technical problem of abnormal detection system reliability decline caused by data missing and sensor failure in sensor network. In the industrial Internet of Things environment, due to factors such as equipment aging, environmental interference, communication failure, etc., sensors often have data missing or complete failure, and traditional anomaly detection methods have strong dependence on complete data, which will lead to a significant decline in detection accuracy or even system failure once data is missing. The application constructs a virtual sensor algorithm and a virtual acquisition router model, establishes an association model between sensors, calculates the data value of the virtual sensor using the data of the normal sensor, and optimizes the data transmission path based on graph theory and network topology analysis. When a sensor node failure is detected, the data transmission path is automatically reconstructed and the virtual sensor node data is generated, the missing data is made up by interpolation of adjacent node data and extrapolation of time series data algorithm, and the abnormal detection system can still maintain normal operation under the condition of partial sensor failure.

[0066] It should be noted that the present application also solves the technical problem of high false positive rate and false negative rate caused by abnormal detection threshold fixation. The conventional abnormal detection method usually uses a fixed threshold for abnormal judgment, but the running state of the industrial equipment will dynamically change with factors such as load change, environmental condition change and equipment aging, and the fixed threshold cannot adapt to such dynamic change, which is easy to produce false alarm when the equipment is lightly loaded, and is easy to appear false negative when the equipment is heavily loaded. The present application establishes a dynamic threshold adjustment function and a parameter self-adaptive updating algorithm, dynamically adjusts the abnormal detection threshold according to the equipment running environment and historical data, automatically adjusts the detection sensitivity according to the output result of the industrial equipment state evaluation model, and at the same time establishes a feedback optimization mechanism to monitor the accuracy and recall rate of abnormal detection, automatically adjusts the model weight and threshold parameter when the detection performance decreases, realizes the self-adaptive optimization of the detection system, and effectively reduces the false positive rate and false negative rate.

[0067] The second aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores program instructions, the program instructions are used to execute the industrial production Internet of Things data anomaly detection method when running in the computer.

[0068] The third aspect of the present application provides an industrial production Internet of Things data anomaly detection system, which contains the computer readable storage medium, the system is any one of computer, server and single chip microcomputer, the computer readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer readable storage medium.

[0069] Specifically, the principle of this invention is as follows: This invention can solve the dimensionality curse problem in anomaly detection of high-dimensional sensor data. Its core principle lies in borrowing successful methods from statistical mechanics for processing massive microscopic particle systems, transforming the microscopic state of high-dimensional sensor data into macroscopic thermodynamic parameters for dimensionality reduction analysis. In statistical mechanics theory, complex systems containing microscopic particles at the Avogadro constant level can be fully described by a few macroscopic thermodynamic parameters such as temperature, pressure, volume, and entropy, which is essentially an efficient dimensionality compression process. This invention applies this principle to industrial sensor data processing, treating each sensor data point as the microscopic state of a virtual particle. Multidimensional sensor data constitute a virtual particle ensemble. By establishing a virtual thermodynamic system, the random fluctuations of sensor data are mapped to the Brownian motion and thermal fluctuations of virtual particles, the changes in equipment energy consumption are mapped to the internal energy and enthalpy change of the virtual thermodynamic system, and the operational stability of the equipment is mapped to the entropy and free energy gradient of the virtual system. This mapping process utilizes the partition function, correlation function, and fluctuation dissipation relationship of statistical mechanics to calculate macroscopic thermodynamic parameters, compressing the original hundreds or thousands of dimensions of sensor data into a dozen or so statistical mechanics feature vectors, achieving an exponential reduction in dimensionality. Since thermodynamic parameters can reflect the equilibrium and non-equilibrium characteristics of a system, when equipment deviates from normal operating conditions, the virtual thermodynamic system exhibits observable macroscopic changes such as entropy anomalies, phase transitions, and statistical distribution shifts. These changes can serve as effective indicators for anomaly detection. Principal component analysis and mutual information algorithms further optimize data input quality, the virtual sensor algorithm ensures data integrity, and the state fusion network dynamically adjusts weight allocation through an attention mechanism. The entire technical solution forms a complete closed loop from data preprocessing, dimensionality compression, anomaly analysis to feedback optimization, significantly reducing computational complexity while maintaining the accuracy and reliability of anomaly detection.

[0070] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0071] The specific implementation of step S01 involves acquiring sensor data from industrial equipment through a distributed data acquisition system and performing preprocessing. The sliding window smoothing filter algorithm is specifically represented as follows:

[0072] ;

[0073] In the formula, For the first The filtered output value of each sampling point; This is the original sensor data sequence; This is half the window length, and its value ranges from 2 to 7; is the total length of the sliding window. The preliminary outlier filtering adopts the 3σ criterion, and the judgment condition is expressed as follows:

[0074] ;

[0075] wherein, is the data point to be detected; is the mean of the data sequence; is the standard deviation of the data sequence; when the condition is met, the is marked as an outlier.

[0076] The specific implementation of step S02 is to perform dimension reduction processing on the multi-dimensional sensor data set. The eigenvalue decomposition of the covariance matrix of principal component analysis is expressed as follows:

[0077] ;

[0078] ;

[0079] wherein, is the dimensional covariance matrix; is the dimensional standardized data matrix, is the sample number, is the feature dimension; is the transpose matrix of the matrix , and the dimension is ; is the th feature vector, and the dimension is ; is the corresponding th eigenvalue. The principal component selection is determined according to the cumulative contribution rate:

[0080] ;

[0081] wherein, is the number of retained principal components. The calculation of the mutual information feature selection is expressed as follows:

[0082] ;

[0083] wherein, is the mutual information value of the feature and the target variable ; is the th feature variable; is the specific value of the feature variable ; is the specific value of the target variable ; is the and The joint probability density; and They are respectively and The marginal probability density.

[0084] The specific implementation of step S03 involves constructing a virtual sensor algorithm to generate missing sensor data. The multiple linear regression model is represented as follows:

[0085] ;

[0086] In the formula, For virtual sensor estimates; For the first Normal sensor data; For the intercept term; These are the regression coefficients; The number of sensors participating in the regression; The error term is estimated, ranging from 0.01 to 0.05. The path weight calculation for the virtual acquisition router model is expressed as follows:

[0087] ;

[0088] In the formula, For sensor nodes and Path weights between them; For sensors and Correlation coefficient of the data; This represents the physical distance between nodes, in meters (m).

[0089] The specific implementation of step S04 involves establishing a pseudo-statistical mechanics anomaly analysis model, which consists of the following set of equations. The Boltzmann distribution of the equipment operating parameters is expressed as follows:

[0090] ;

[0091] In the formula, Equipment operating status The probability distribution; For state The corresponding effective energy is obtained by mapping the sensor data after normalization, and the unit is joules. It is the partition function; Boltzmann constant, with values ​​ranging from 1 to 10. Joules per Kelvin; The effective temperature parameter, in Kelvin, is obtained by fitting equipment operating data. The partition function is calculated as follows:

[0092] ;

[0093] where, is the total number of states, determined by the degree of discretization of the sensor data, typically in the range of 100 to 1000. The system entropy is calculated as follows:

[0094] ;

[0095] where, is the system entropy, in Joules per Kelvin. The average energy is calculated as follows:

[0096] ;

[0097] where, is the average energy, in Joules. The energy fluctuation is calculated as follows:

[0098] ;

[0099] where, is the standard deviation of the energy fluctuation, in Joules; is the expectation value of the energy squared, calculated as ; is the square of the average energy. The heat capacity is calculated as follows:

[0100] ;

[0101] where, is the heat capacity at constant volume, in Joules per Kelvin. The correlation length is calculated as follows:

[0102] ;

[0103] where, is the spatial correlation length, in m; is the spatial position vector of state , in m; is the spatial position vector of state , in m; is the Euclidean distance between states and ; and are the probability distributions of states and , respectively. The correlation time is calculated as follows:

[0104] ;

[0105] where, is the system relevant time, in s; is the time variable, in s; is the normalized autocorrelation function; is is the system instantaneous energy at time t; is the system energy at initial time t0; is the time correlation average of energy at different times. The construction of the statistical mechanics eigenvector is shown as follows:

[0106] ;

[0107] wherein, is the 6-dimensional statistical mechanics eigenvector.

[0108] The specific implementation of step S05 is to analyze the statistical mechanics eigenvector by using the industrial equipment state evaluation model. The dynamic threshold adjustment function is shown as follows:

[0109] ;

[0110] wherein, is the dynamically adjusted detection threshold; is the basic threshold; is the ambient temperature correction coefficient, ranging from 0.8 to 1.2; is the load correction coefficient, ranging from 0.9 to 1.1; is the maintenance history correction coefficient, ranging from 0.85 to 1.15.

[0111] The specific implementation of step S06 is to establish an abnormal pattern classification system. The abnormal severity evaluation function is shown as follows:

[0112] ;

[0113] wherein, is the abnormal severity score; is the abnormal deviation degree; is the abnormal duration; is the impact range; is the historical abnormal frequency; , , , is the corresponding weight coefficient.

[0114] The specific implementation of step S07 is the same as the foregoing, and will not be described in detail here.

[0115] The attention adjustment coefficient value calculation of the state weight adjustment function is shown as follows:

[0116] ;

[0117] wherein, is the attention adjustment coefficient value; is the sensor data variance; is the data correlation coefficient, ranging from 0 to 1; is the anomaly detection historical accuracy, ranging from 0 to 1; is the current device load rate, ranging from 0 to 1; and is the weighting coefficient, usually ranging from 0 to 1 , ; is the adjustment error term, ranging from -0.05 to 0.05. The first interval value calculation is represented as follows:

[0118] ;

[0119] wherein, is the first interval value. The second interval value calculation is represented as follows:

[0120] ;

[0121] wherein, is the second interval value. The third interval value calculation is represented as follows:

[0122] ;

[0123] wherein, is the third interval value, calculated using the harmonic mean method.

[0124] The principles and effects of each formula are explained as follows. The sliding window smoothing filter formula Based on the principle of finite impulse response filter in digital signal processing, the formula suppresses high-frequency noise by weighted average of adjacent data points. Compared with the traditional single-point detection method, this formula can effectively reduce the influence of random noise on anomaly detection accuracy, improving the stability and reliability of data preprocessing. Criterion judgment formula Based on the statistical characteristics of normal distribution, the formula uses the standard deviation multiple of data deviation from the mean to identify abnormal points. Compared with the fixed threshold method, this formula can adaptively adjust the judgment standard according to the data distribution characteristics, reducing the misjudgment caused by data distribution differences.

[0125] Principal component analysis eigenvalue decomposition formula Based on the matrix eigenvalue decomposition theory in linear algebra, the formula realizes dimension reduction by finding the projection direction with the maximum data variance. Compared with the traditional feature selection method, this formula can retain the main information of the data while reducing the dimension, improving the computational efficiency of the subsequent algorithm. Mutual information calculation formula Based on the information measurement principle in information theory, the feature importance is evaluated by calculating the statistical dependence between variables. Compared with the linear correlation coefficient method, this formula can capture nonlinear correlation relationships and improve the accuracy of feature selection.

[0126] Virtual sensor multiple regression formula Based on the least squares estimation theory, the missing data is predicted by establishing a linear relationship model between sensors. Compared with the simple interpolation method, this formula considers the synergistic effect of multiple sensors, improving the accuracy and physical reasonableness of data reconstruction. Path weight calculation formula Considering the correlation and physical distance of sensor data, the connection strength between nodes is quantified by distance-normalized correlation coefficients. Compared with fixed topology, this formula can dynamically optimize data transmission paths and improve the adaptive ability of the network.

[0127] Boltzmann distribution formula Originating from the canonical ensemble theory in statistical physics, the device operating state is mapped to the energy distribution of a thermodynamic system, and the state probability is described by an exponential decay function. Compared with the empirical probability model, this formula provides a solid physical theoretical basis and improves the explainability of anomaly detection. Partition function formula As a core concept of statistical mechanics, the probability distribution is normalized by statistical summation of all possible states. Compared with direct probability calculation, this formula ensures the mathematical rigor and physical consistency of the probability distribution. System entropy formula Based on the second law of thermodynamics, the disorder degree of the system is quantified by the weighted logarithmic sum of the probability distribution. Compared with the traditional variance index, this formula can more accurately reflect the complexity and stability of the system state. Average energy formula The expected energy value of the system is calculated by probability-weighted summation, providing a basic thermodynamic parameter for device state evaluation. Energy fluctuation formula Based on the variance definition, the statistical fluctuation of energy is calculated, reflecting the stability of the system. Compared with direct variance calculation, this formula has a clear physical meaning in the framework of statistical mechanics. Heat capacity formula Based on the fluctuation dissipation theorem, the heat capacity is calculated by the relationship between energy fluctuation and temperature, providing a quantitative index for system response characteristics.

[0128] Dynamic threshold adjustment formula The adaptive adjustment of the threshold is realized by the product form of multiple factor correction, considering factors such as environmental conditions, operating load, and maintenance history. Compared with the fixed threshold method, this formula can adapt to changes in different working conditions, significantly reducing the false positive rate and false negative rate. Abnormal severity evaluation formula The multiple abnormal feature indicators are fused into a single score value in the form of weighted linear combination, and the importance of each factor is balanced by the weight coefficient determined by expert experience. Compared with single index evaluation, this formula can comprehensively quantify the severity of abnormality and provide a scientific basis for maintenance decision-making.

[0129] State weight adjustment function The linear combination of weighted geometric mean is adopted, and the system state is comprehensively evaluated by weighted summation of two geometric mean terms. Compared with simple linear combination, the geometric mean term in this formula can better handle the relationship between parameters of different dimensions. Interval value calculation formula 、 The mathematical form of geometric mean is adopted respectively. Geometric mean is suitable for representing the comprehensive level of two positive numbers. The harmonic mean formula is more suitable for processing the average of ratio data. Compared with simple arithmetic mean, these formulas can more accurately reflect the internal relationship between different parameters, improving the scientificity and effectiveness of the weight allocation strategy.

[0130] In order to better understand and implement the present application, the following provides an embodiment 2 of a specific application scenario of the present application: a technical team deploys an industrial production Internet of Things data anomaly detection system in a production line, and configures 276 sensors for monitoring equipment operating state. The system operating environment temperature range is 15℃ to 45℃, the equipment load rate fluctuates between 60% to 95%, and the sensor data acquisition frequency is set to 50Hz to 500Hz.

[0131] The technical team first establishes a distributed data acquisition system, and in the heating furnace area, 42 temperature sensors, 18 pressure sensors and 12 concentration sensors are deployed, with measurement ranges of 800℃ to 1200℃, 0.5MPa to 2.5MPa and 200ppm to 800ppm respectively. The rough rolling mill group is configured with 36 vibration sensors, 24 current sensors and 15 speed sensors, with vibration acceleration measurement range of to m / s², current range of 100A to 2000A, and speed range of 50rpm to 800rpm. The finishing rolling mill group is installed with 48 temperature sensors, 32 pressure sensors, 28 vibration sensors and 20 displacement sensors, with displacement measurement accuracy reaching 10μm level. The coiling machine area is deployed with 21 tension sensors and 18 temperature sensors, with tension measurement range of 500N to N.

[0132] The sensor data preprocessing stage adopts a sliding window smoothing filter algorithm, and the window length is set to 9 sampling points, effectively removing high-frequency noise with a frequency higher than 25 Hz. In the preliminary filtering process of abnormal values, it is found that the mean value of the temperature data in the heating furnace area is 1050℃, and the standard deviation is 45℃, identifying 218 abnormal data points and eliminating them. The mean value of the rough rolling mill vibration data is 0.8 m / s², and the standard deviation is 0.15 m / s², filtering out 134 outliers. The established multi-dimensional sensor data set contains 276 feature dimensions, with a time span of 72 consecutive hours, and the total amount of data reaches sampling points.

[0133] The principal component analysis dimension reduction processing stage calculates the eigenvalue distribution of the covariance matrix, and the cumulative contribution rate of the first 15 principal components reaches 91.3%, so 15 principal components are retained as the reduced features. The mutual information feature selection algorithm calculates the correlation between each feature and the device abnormal state, and selects the top 68 feature variables in the mutual information value, and the final reduced feature data set dimension is 68, with a dimension reduction ratio of 75.4% compared to the original 276-dimensional data.

[0134] In the construction process of the virtual sensor algorithm, the technical team found that the No. 3 vibration sensor of the rough rolling mill unit had intermittent failures, with a data missing rate of 12.8%. Through a multiple linear regression model, the correlation between this sensor and the adjacent 8 sensors was established, with regression coefficients of 0.73, -0.42, 0.56, 0.31, -0.28, 0.65, -0.39, 0.44, and an intercept of 0.067, with an estimation error controlled within 3.2%. The virtual acquisition router model establishes a node connection graph based on the sensor network topology, as shown in Figure 3 , identifying 6 key path nodes and 12 redundant transmission paths. When the No. 3 sensor fails, the system automatically switches to a backup data source, and generates virtual sensor data through interpolation of adjacent sensor data, with data integrity improved to 98.7%.

[0135] The anomalous analysis model of statistical mechanics maps the device operating parameters to the statistical physics space, and the effective temperature parameter fitting value is 850K, corresponding to the comprehensive operating temperature level of the device. The partition function calculation adopts the Monte Carlo method for numerical integration, with a sampling number of times, and the calculation accuracy reaches level. The system entropy value stabilizes at J / K near the normal operating state, with an average energy of J, and an energy fluctuation standard deviation of J. The calculated constant volume heat capacity is J / K, the spatial correlation length is 2.3 m, and the time correlation length is 42 s. The constructed statistical mechanics feature vector contains 6 components, effectively compressing the complex information of high-dimensional sensor data.

[0136] The industrial equipment state evaluation model adopts a state fusion network based on an attention mechanism. The network structure includes a 128-dimensional data embedding layer, 8 self-attention layers, a 256-dimensional feedforward neural network layer, and an output mapping layer. The training data set includes 65,000 samples of normal operation data, 8,500 samples of slight abnormal data, 3,200 samples of medium abnormal data, and 1,800 samples of severe abnormal data, as shown in Table 1.

[0137] Table 1: Statistical table of equipment abnormality type distribution

[0138]

[0139] The dynamic threshold adjustment function adjusts in real time according to environmental conditions. The environmental temperature correction coefficient varies from 0.85 to 1.15 during the observation period, the load correction coefficient varies from 0.92 to 1.08, and the maintenance history correction coefficient is stable around 0.95. The basic threshold is set to 0.75, and the detection threshold after dynamic adjustment fluctuates between 0.68 and 0.82, effectively adapting to changes in different working conditions.

[0140] The abnormality detection operation results show that the system detected 127 equipment operation abnormalities during the 72-hour monitoring period, including 72 slight abnormalities, 38 medium abnormalities, 15 severe abnormalities, and 2 emergency abnormalities. As shown in Figure 4 , the statistical mechanics feature vector shows obvious distribution differences under different abnormal states. The system entropy value is concentrated in the low value interval under normal state, and the entropy value increases significantly under abnormal state, reflecting the increase of system disorder degree. The abnormal severity evaluation function calculation results show that the average score of slight abnormality is 18.5, the average score of medium abnormality is 42.3, the average score of severe abnormality is 68.7, and the average score of emergency abnormality is 89.2, with reasonable score distribution and obvious discrimination.

[0141] In the calculation process of the attention adjustment coefficient value of the state weight adjustment function, the sensor data variance is 0.34, the data correlation coefficient is 0.67, the abnormality detection historical accuracy is 0.91, and the current equipment load rate is 0.78. The calculated attention adjustment coefficient value is 0.72, located between the second interval value 0.83 and the third interval value 0.59, the system adopts an active weight allocation strategy, enhancing the response sensitivity of attention weight. Figure 5 As shown in Figure 5 , compared with the traditional fixed threshold detection method, the dynamic threshold adjustment mechanism of the invention significantly reduces the false alarm phenomenon and improves the accuracy of abnormality detection.

[0142] The feedback optimization mechanism continuously adjusts the model parameters according to the detection effect, and the system accuracy rate is maintained at 92.4% during monitoring, the recall rate reaches 88.7%, and the false positive rate is controlled within 6.8%. The parameter adaptive update algorithm adjusts the network weight 15 times, optimizes the detection threshold parameter 8 times, and updates the feature selection parameter 5 times, ensuring the continuous stability of the detection system performance.

[0143] The system establishes a sensor connection graph containing 276 nodes and 456 edges, identifies 18 key hub nodes and 32 redundant path branches. The network diameter is 7 hops, the average path length is 3.2 hops, and the clustering coefficient is 0.43, which reflects good small-world network characteristics. When a sensor fails, the network can complete path reconstruction in an average of 1.8 seconds, and the data transmission delay increase is controlled within 15%.

[0144] Statistical analysis of the detection results shows that the accuracy rate of the heating furnace area anomaly detection is 94.2%, and the main abnormal types are temperature overrun and uneven combustion. The accuracy rate of the rough rolling mill group anomaly detection is 91.8%, and the main fault modes are vibration anomaly and current fluctuation. The accuracy rate of the finishing rolling mill group detection is 93.5%, and the typical abnormal characteristics are pressure anomaly and displacement overrun. The coiling machine area detection accuracy rate is 90.7%, and tension fluctuation and temperature anomaly are more common. The performance indicators of each area anomaly detection are shown in Table 2.

[0145] Table 2 Performance statistics table of each device area anomaly detection

[0146]

[0147] The present application realizes significant technical progress compared with traditional anomaly detection methods. Traditional methods mainly rely on single sensor data and fixed threshold judgment, which are easily affected by environmental interference and equipment working condition changes, resulting in high false alarm rate and poor adaptability. The present application solves the problem of data loss caused by sensor failure through a virtual sensor algorithm, and the data reconstruction method based on multi-sensor information fusion has higher physical rationality and estimation accuracy than simple interpolation. The anomalous analysis model based on statistical mechanics maps the device operating state to the statistical physics space, identifies abnormal patterns using the principle of entropy increase and phase transition, and has a solid theoretical basis and stronger interpretability compared with the empirical threshold method. The attention mechanism state fusion network can dynamically allocate importance weights to different sensor data, adaptively identify key fault features, and significantly improve the accuracy and robustness of anomaly detection compared with the fixed weight method. The dynamic threshold adjustment mechanism optimizes the detection parameters in real time according to environmental conditions and equipment state, which can better adapt to working condition changes and effectively reduce the false alarm rate and the false negative rate compared with the static threshold setting. The parameter adaptive updating algorithm continuously optimizes the model performance through detection effect feedback, ensuring the stability and accuracy of the system in long-term operation, which is more intelligent and efficient than the traditional manual parameter adjustment method.

[0148] It should be noted that the variables involved in the present application are explained in detail as shown in Table 3.

[0149] Table 3 Variable explanation table

[0150]

[0151] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An industrial production Internet of Things data anomaly detection method, characterized in that, The method comprises the following steps: collecting and preprocessing sensor data generated by industrial equipment to establish a multidimensional sensor data set, performing dimension reduction processing on the multidimensional sensor data set by principal component analysis, and establishing a reduced dimension feature data set by using a mutual information algorithm for feature selection, constructing a virtual sensor algorithm to generate virtual sensor data from the reduced dimension feature data set to form an extended sensor data set, establishing an analog statistical mechanics anomaly analysis model based on statistical mechanics laws to map the extended sensor data set to a statistical mechanics space to calculate a statistical mechanics feature vector, analyzing the statistical mechanics feature vector by using an industrial equipment state evaluation model to identify an equipment operation anomaly mode by calculating an anomaly detection threshold value through a dynamic threshold adjustment function, classifying and labeling the equipment operation anomaly mode by establishing an anomaly mode classification system, generating an anomaly detection report by calculating an anomaly severity evaluation function, adjusting detection parameters according to the anomaly detection report to optimize the performance of the industrial equipment state evaluation model through a parameter self-adaptive updating algorithm to realize continuous improvement of the anomaly detection system; the analog statistical mechanics anomaly analysis model is a device state analysis model constructed based on statistical mechanics theory and thermodynamic equilibrium principle, which maps the multidimensional operation parameters of the industrial equipment into the microscopic particle state distribution and macroscopic thermodynamic parameters in the statistical mechanics space, and simulates the equipment operation state by establishing a virtual particle ensemble and a thermodynamic system; the analog statistical mechanics anomaly analysis model maps the random fluctuations of the equipment operation parameters into the Brownian motion and thermal fluctuations of the virtual particles, maps the equipment energy consumption changes into the internal energy and enthalpy changes of the virtual thermodynamic system, and maps the equipment operation stability into the entropy value and free energy gradient of the virtual system; the equipment operation data are analyzed regularly by using the Boltzmann distribution law and the Maxwell statistical distribution, the partition function, the correlation function and the fluctuation dissipation relationship of the virtual particle ensemble are calculated, and the anomaly mode is identified by detecting the entropy increase anomaly, the phase change phenomenon and the statistical distribution deviation of the virtual thermodynamic system. 2.The industrial production IoT data anomaly detection method of claim 1, wherein, The virtual sensor algorithm is a calculation method for inferring missing or faulty sensor data based on existing sensor data, which calculates the data value of the virtual sensor by establishing the correlation model between the sensors and using the data of the normal sensors, thereby making up for the data missing problem in the sensor network. 3.The industrial production IoT data anomaly detection method of claim 2, wherein, The virtual sensor algorithm uses a virtual acquisition router model, which is a data path optimization model based on graph theory and network topology analysis, and identifies the weak links and data transmission bottlenecks in the network by constructing the connection relationship graph between the sensor nodes and calculating the shortest path weight between the nodes.

4. The industrial production IoT data anomaly detection method of claim 3, wherein, When the virtual acquisition router model detects a sensor node failure or a data transmission interruption, it automatically reconstructs the data transmission path and generates virtual sensor node data according to the redundant path and backup node information of the network topology structure, and makes up for the missing data by using the interpolation of adjacent node data and the extrapolation algorithm of time series data.

5. The industrial production IoT data anomaly detection method of claim 4, wherein, The statistical mechanics feature vector, in particular, is a vector set representing the equipment operation state calculated by the pseudo-statistical mechanics anomaly analysis model, and contains micro state distribution characteristics, macro thermodynamic equilibrium state and statistical fluctuation law in the equipment operation process.

6. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program instructions, and the program instructions are used to execute the industrial production Internet of Things data anomaly detection method in any one of claims 1-5 when running in the computer.

7. An industrial production IoT data anomaly detection system, characterized in that, The system is any one of a computer, a server, and a single-chip microcomputer, the computer readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer readable storage medium.

Images