Intelligent monitoring and online anomaly detection system for coal conveying system

Abstract

The invention relates to the field of industrial intelligent monitoring and fault detection, and discloses a coal conveying system intelligent monitoring and online anomaly detection system which comprises a data processing module, a feature extraction module, a dynamic tensor fusion module, an anomaly scoring module, a depth detection module and a decision control module. Carrying out pretreatment; the feature extraction module extracts features and converts the features into multi-dimensional feature vectors; the dynamic tensor fusion module fuses the multi-modal data into a unified tensor; the abnormal scoring module constructs a Markov decision process model and optimizes a scoring strategy; the depth detection module is used for extracting correlation characteristics among sensor modes and analyzing a time sequence dependency relationship of data; and the decision control module executes an intelligent control strategy according to the abnormal score and the depth detection result. According to the invention, the whole-process intelligent monitoring of the coal conveying system can be realized, the accuracy and real-time performance of anomaly detection are improved, the manual inspection cost is reduced, and the operation safety and stability of the system are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial intelligent monitoring and fault detection, and in particular to an intelligent monitoring and online anomaly detection system for a coal transportation system. Background Art

[0002] Coal conveying systems are a critical component of thermal power plants, steel mills, and the coal transportation industry. Their operational status directly impacts production efficiency and equipment safety. With the increasing automation of industry, traditional coal conveying system monitoring methods are gradually evolving toward intelligent systems. However, existing coal conveying system monitoring methods still have many limitations and cannot meet the modern industrial demand for high-precision, low-latency anomaly detection.

[0003] Currently, anomaly detection in coal conveying systems primarily relies on analyzing single sensor data, such as temperature, vibration, or coal flow images. This approach is susceptible to noise interference when handling complex operating conditions, resulting in insufficient anomaly detection accuracy. Furthermore, traditional methods typically use fixed thresholds to define anomaly criteria, making them difficult to adapt to the varying operating conditions and loads of coal conveying systems, resulting in significant false positives and missed detections.

[0004] On the other hand, while existing intelligent detection methods incorporate deep learning technology, most rely on static feature extraction, focusing only on data from a single time step and ignoring the dynamic nature of coal handling systems. Furthermore, deep learning models exhibit high computational complexity, making them difficult to implement in industrial field deployments due to limited computing resources and real-time requirements, making them difficult to meet the low-latency online detection requirements of coal handling systems.

[0005] When it comes to applying anomaly detection results, existing coal handling systems often rely on fixed rules for exception handling and are unable to intelligently adjust based on detection results. For example, when an anomaly is detected, only a simple alarm mechanism is triggered, lacking intelligent adjustments based on the severity of the anomaly. This results in the system being unable to effectively prevent equipment damage or production interruptions. Furthermore, the lack of an adaptive model optimization mechanism prevents the system from adjusting its detection strategy based on new data during long-term operation, leading to a gradual decline in detection effectiveness over time.

[0006] In summary, the current coal transportation system monitoring technology still has obvious deficiencies in data fusion, anomaly detection accuracy, real-time performance and intelligent decision-making capabilities. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention provides an intelligent monitoring and online anomaly detection system for coal transportation systems, which solves the problems of traditional anomaly detection methods in coal transportation systems, such as insufficient utilization of multimodal data, low detection accuracy, poor real-time performance, high false alarm rate and lack of intelligent control.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a coal transportation system intelligent monitoring and online anomaly detection system, comprising:

[0009] Data processing module, used to collect real-time data from the coal transportation system, perform pre-processing, and output standardized data format;

[0010] A feature extraction module, configured to extract feature values from the preprocessed data and generate feature vectors;

[0011] Dynamic tensor fusion module, used to fuse multimodal data and optimize feature representation;

[0012] An anomaly scoring module, configured to perform an anomaly scoring on the fusion feature and output an anomaly scoring value;

[0013] A deep detection module, configured to perform anomaly detection on the coal transportation system based on the anomaly score value and the feature vector;

[0014] A decision control module is used to perform system control operations based on the abnormality score value and the detection result.

[0015] Preferably, the data processing module includes:

[0016] A sensor collection unit for collecting real-time data from the coal conveying system, including temperature, vibration, infrared images, coal flow and humidity;

[0017] The data preprocessing unit is used to preprocess the collected data.

[0018] Preferably, the feature extraction module includes:

[0019] A temperature feature extraction unit is used to extract the mean and variance from the temperature data as temperature features;

[0020] A vibration feature extraction unit, configured to extract spectrum energy features from the vibration data as vibration features;

[0021] An image feature extraction unit, used for extracting a grayscale histogram from the infrared image as an image feature;

[0022] The coal flow feature extraction unit is used to extract the time mean and volatility from the coal flow signal as coal flow features.

[0023] Preferably, the dynamic tensor fusion module includes:

[0024] A tensor construction unit, used to construct the preprocessed multimodal data into a three-dimensional tensor;

[0025] The constrained tensor decomposition unit is used to decompose the three-dimensional tensor by non-negative tensor decomposition, and the objective function is:

[0026]

[0027] in, is a three-dimensional tensor, C(t) is the operating condition covariance matrix generated by coal flow and ambient humidity, W, U, V are decomposition factor matrices, and α is the dynamically adjusted regularization coefficient;

[0028] The dynamic feature optimization unit is used to combine the decomposition factor matrix with the real-time operating condition parameters to generate a fusion feature vector.

[0029] Preferably, the anomaly scoring module includes:

[0030] The state space modeling unit is used to define the fused feature vector F(t) output by the dynamic tensor fusion module as the state s(t), where d is the dimension of the feature vector;

[0031] The action space definition unit is used to generate anomaly scoring results a(t), where a(t)∈[0,1], 0 indicates normal state and 1 indicates severe anomaly;

[0032] A dynamic reward function unit is used to adjust the reward value according to the false alarm rate and the missed alarm rate. The reward function is:

[0033]

[0034] in:

[0035] Q(t) is the current coal flow, Q max is the maximum design coal flow rate, TPR(t) represents the true positive rate, FPR(t) represents the false positive rate, and β(t) is the dynamic adjustment coefficient;

[0036] The policy optimization unit uses the asynchronous advantage actor-critic algorithm to update the policy network parameters θ, and the optimization objective is:

[0037]

[0038] Among them, π θ (a(t)|s(t)) represents the strategic probability of taking action a(t) in state s(t), A(s(t),a(t)) is the advantage function, is the expected value operator.

[0039] Preferably, the depth detection module includes:

[0040] The spatiotemporal convolution unit is used to perform spatiotemporal separation convolution processing on the fused feature vector, including:

[0041] Spatial convolution layer, the convolution kernel dimension is Extract spatial correlation features between sensor modalities, where M is the number of sensor modalities and R is the potential feature dimension;

[0042] Temporal convolution layer, the convolution kernel dimension is Extract time-dependent features, where k is the length of the time window;

[0043] A quantization calculation unit, configured to perform integer quantization on the weights of the spatiotemporal convolutional layer;

[0044] The prediction output unit is used to generate the abnormal probability value y(t)∈[0,1], and the calculation formula is:

[0045] y(t)=σ(W o ·h t )

[0046] Among them, h t is the output feature of the temporal convolution layer, is the output layer weight, and σ(·) is the Sigmoid function.

[0047] Preferably, the decision control module includes:

[0048] The early warning trigger unit is used to perform the following operations based on the abnormal probability value y(t)∈[0,1] output by the deep detection module and the scoring result a(t)∈[0,1] of the abnormal scoring module:

[0049] When the abnormal probability value y(t) is within a predefined first threshold range, reducing the operating rate of the coal transportation system and sending an early warning signal;

[0050] an emergency shutdown unit, configured to trigger an emergency shutdown instruction when the abnormal probability value y(t) exceeds a predefined second threshold and the abnormal score value a(t) exceeds a predefined third threshold;

[0051] The online model update unit is used to dynamically update the model parameters based on the newly collected data. Specifically, the tensor decomposition weights are adjusted using the following formula:

[0052]

[0053] Where η is the learning rate, is the gradient of the loss function with respect to the weight matrix W, A tensor constructed for newly added data; the update process is performed asynchronously with the PLC control cycle.

[0054] The present invention also provides a method for intelligent monitoring and online anomaly detection of a coal transportation system, comprising the following steps:

[0055] Collecting real-time data of the coal transportation system and preprocessing the data;

[0056] Extract features from the preprocessed data to generate feature vectors;

[0057] Fuse multimodal data into tensors and optimize them;

[0058] Use reinforcement learning to dynamically adjust anomaly scores;

[0059] Use deep learning models to perform anomaly detection on the fused features;

[0060] Execute early warning or emergency shutdown control operations based on the detection results.

[0061] The present invention provides an intelligent monitoring and online anomaly detection system for coal transportation systems. It has the following beneficial effects:

[0062] 1. This invention utilizes a dynamic tensor fusion method to unify the modeling of multiple sensor data points within the coal handling system, including temperature, vibration, and infrared images. This overcomes the limitations of traditional single-source detection. Through non-negative tensor decomposition technology, the system extracts correlation features between modal data and optimizes the data fusion process by combining real-time operating parameters. Compared to traditional single-sensor anomaly detection methods, this solution significantly improves the accuracy and reliability of anomaly detection, enabling the system to maintain high detection accuracy even under complex operating conditions.

[0063] 2. This invention utilizes a reinforcement learning-driven anomaly scoring mechanism that dynamically adjusts the scoring strategy based on the real-time status of the coal handling system. The system employs Markov decision modeling and incorporates historical data to construct a reward function, ensuring that anomaly scoring reduces false positives while improving anomaly identification sensitivity. Compared to fixed threshold judgment methods, this scoring mechanism continuously learns and optimizes its discriminative capabilities, enabling the system to maintain accurate anomaly identification capabilities across diverse coal handling environments.

[0064] 3. This invention utilizes a lightweight spatiotemporal convolutional network, which significantly reduces computational resource consumption and improves the system's real-time performance while ensuring high detection accuracy. The system extracts correlation features between sensor data through spatial convolution and combines it with temporal convolution to capture the temporal changes in the system's operating state. Model quantization techniques are also employed to reduce the computational burden. Compared to traditional deep learning methods, this solution can still operate efficiently on edge devices, enabling the system to quickly respond to abnormal situations and meet industrial-grade real-time requirements.

[0065] 4. This invention provides an intelligent decision-making and control strategy that dynamically adjusts the coal handling system's operating status based on anomaly scores and in-depth detection results. The system automatically reduces the coal handling rate in the early stages of an anomaly to minimize the impact of the failure. When the anomaly score exceeds a set threshold, the system immediately triggers an emergency shutdown to prevent equipment damage or production accidents. Furthermore, the system incorporates an online model update mechanism that continuously optimizes the detection model as the coal handling system's operating conditions evolve, improving long-term operational stability and adaptability.

[0066] 5. Through the coordinated operation of data acquisition, feature extraction, tensor fusion, anomaly scoring, deep detection, and intelligent decision-making and control, the present invention forms a complete set of intelligent monitoring and anomaly detection solutions for coal transportation systems. The system can cover multiple key links in the coal transportation process, including data monitoring, anomaly identification, intelligent scoring, risk assessment, and automatic control, realizing the intelligent management of the entire coal transportation process. Compared with traditional solutions that rely on manual inspections or simple rule-based judgments, this system can enhance the intelligence level of the coal transportation process, reduce manual intervention, improve equipment safety and production efficiency, and provide a highly efficient and stable intelligent monitoring solution for the coal transportation industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 Schematic diagram of the system architecture of the present invention;

[0068] Figure 2 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0070] Please see the attached Figure 1 The present invention provides an intelligent monitoring and online anomaly detection system for a coal transportation system, which can monitor the operating status of the coal transportation system in real time and intelligently identify and regulate abnormal situations.

[0071] System architecture such as Figure 1 As shown, the system can include a data processing module, a feature extraction module, a dynamic tensor fusion module, an anomaly scoring module, a depth detection module, and a decision-making control module. These modules work together to achieve intelligent monitoring, anomaly identification, and control and regulation of the coal transportation system.

[0072] The following is a detailed description of each module in the method of the present invention.

[0073] The data processing module of this invention is designed to collect and preprocess raw data from various sensors to provide accurate input data for subsequent analysis. This module uses optimized data collection methods and signal processing techniques to ensure high-quality and real-time data, effectively supporting the system's intelligent anomaly detection capabilities. The following is a detailed implementation of this module.

[0074] The sensor acquisition unit collects multimodal data from the coal conveying system in real time, including temperature T(t), vibration V(t), infrared image I(t), coal flow Q(t), and humidity H(t).

[0075] The collected data will be passed to the data preprocessing unit, which is used to perform necessary preprocessing on the temperature, vibration and infrared image signals to eliminate noise, remove redundant information and extract meaningful features. The specific processing method is as follows:

[0076] Temperature signal processing: The temperature signal T(t) will be processed by calculating its mean μ T and variance To extract features, we can reduce the noise caused by sensor errors or external environmental interference. The specific mathematical calculation is as follows:

[0077]

[0078] Where n represents the total number of data points, T i The temperature signal value at each moment. Through these statistical features, the average value and fluctuation degree of the temperature signal can be obtained, providing basic data for subsequent analysis.

[0079] Vibration signal processing: The vibration signal V(t) is subjected to frequency domain analysis by fast Fourier transform (FFT) to extract its spectrum energy characteristics E V Spectral energy reflects the intensity and frequency distribution of the vibration signal and can effectively identify vibration patterns caused by equipment failure or abnormality. Its mathematical expression is as follows:

[0080]

[0081] Among them, F k (V) is the FFT transform result of signal V(t), and K is the number of frequency components calculated by the FFT. By calculating the spectrum energy, the characteristics of the vibration signal can be extracted to assist in determining the operating status of the equipment.

[0082] Infrared Image Processing: Calculate the grayscale histogram of the infrared image I(t) to obtain the image's grayscale distribution information. This information reflects the temperature distribution in the coal flow area and provides strong support for fault detection. Specifically, by calculating the image's grayscale histogram H(G), key visual features can be extracted. Principal Component Analysis (PCA) dimensionality reduction is then performed to reduce the data's dimensionality and redundant information, making subsequent analysis more efficient.

[0083] H(G)=[h1,h2,...,h 220 ]

[0084] Among them, h r Represents the rth grayscale value of the image.

[0085] After completing the above processing steps, the data preprocessing unit converts the temperature, vibration and image signals into standardized feature data, providing reliable input for subsequent feature extraction and tensor construction.

[0086] The feature extraction module of the present invention converts the raw signals output by the data processing module into multidimensional feature vectors for subsequent tensor construction and analysis. This module generates input data suitable for model training and anomaly detection by extracting deep features from temperature, vibration, and infrared image data. The feature extraction process involves the calculation and conversion of temperature, vibration, and image features. The following is a detailed implementation of this module.

[0087] The temperature feature extraction unit extracts statistical features from the temperature signal T(t). Specifically, by calculating the mean μ of the temperature signal T and variance As a statistical feature of temperature. Mean μ T reflects the average temperature level, while the variance It indicates the amplitude of temperature fluctuation.

[0088] The vibration feature extraction unit is responsible for extracting the spectrum energy features from the vibration signal V(t). The energy of the vibration spectrum reflects the intensity and frequency distribution of the signal and can effectively distinguish between normal and abnormal vibration modes. The spectrum energy E of the vibration signal V As a feature, it can help identify abnormal vibrations caused by equipment failure or external disturbances.

[0089] To ensure consistency and comparability of vibration signatures, the vibration signature normalization unit normalizes the extracted spectral energy to produce standardized vibration signatures. This normalization process eliminates the effects of factors such as sensor location and sensitivity, ensuring comparability of vibration signatures across different time steps and sampling points.

[0090] The image feature extraction unit is responsible for extracting grayscale information from the infrared image data I(t) and converting it into useful features. First, the image grayscale information is extracted by calculating the grayscale histogram H(G). The grayscale histogram of an image reflects the distribution of grayscale values within the image and can be used to represent the thermal distribution of the coal flow area.

[0091] To reduce the dimensionality of image data and remove redundant information, the image dimensionality reduction unit uses principal component analysis (PCA) to reduce the dimensionality of image features. PCA calculates the principal components of the data, mapping the original features into a lower-dimensional space while retaining the most important feature information. This process significantly reduces computational complexity while preserving the core information of the image, helping to improve the efficiency and accuracy of subsequent analysis.

[0092] Through this processing, the feature extraction module extracts representative features from temperature, vibration, and image signals and converts them into multidimensional feature vectors suitable for input into the next tensor building module. These features become the basis for subsequent data analysis, anomaly detection, and intelligent decision-making.

[0093] The dynamic tensor fusion module of this invention fuses and optimizes multimodal data to construct an efficient and accurate feature representation for subsequent anomaly detection and decision control. This module utilizes tensor decomposition technology combined with real-time operating constraints to dynamically fuse data from different sensors. The following describes a specific implementation of this module.

[0094] The tensor construction unit converts the pre-processed multimodal data (temperature, vibration, infrared image, etc.) into a three-dimensional tensor. Specifically, assuming the input data is X(t), its dimension is N×M×K, where:

[0095] N represents the time step, which indicates the change of data over time;

[0096] M represents the modality number, which is the type of data collected by different sensors (such as temperature, vibration, infrared image, etc.);

[0097] K is the feature dimension of each modality, indicating the number of features extracted from each modality data.

[0098] The tensor construction process can be expressed as follows:

[0099]

[0100] This tensor represents data at different time steps, different modes and feature dimensions, and can provide a basis for subsequent tensor decomposition and feature fusion.

[0101] The constrained tensor decomposition unit performs non-negative tensor decomposition (NTF) on the constructed three-dimensional tensor. Decompose into three factor matrices W, U, V, and optimize them according to the working condition constraints. The specific decomposition form is:

[0102]

[0103] in:

[0104] is the feature matrix in the time dimension;

[0105] is the feature matrix in the modal dimension;

[0106] is the feature matrix on the feature dimension;

[0107] R is the rank of the decomposition, which controls the accuracy and complexity of the decomposition;

[0108] Represents a tensor product operation.

[0109] The non-negative tensor decomposition method decomposes the original data tensor into three matrices, so that a low-rank approximation of the original data can be achieved, thereby reducing the computational complexity and achieving efficient data representation.

[0110] In order to optimize the stability and adaptability of tensor decomposition, the objective function optimization unit adopts a constrained optimization strategy, introduces the dynamic working condition constraint matrix C(t), and constructs the objective function as follows:

[0111]

[0112] in:

[0113] Represents the reconstruction error of tensor decomposition. Minimizing this term can ensure that the decomposed tensor can approximate the original data to the greatest extent;

[0114] C(t) is the real-time working condition constraint matrix, which is defined as follows:

[0115]

[0116] in:

[0117] Var(Q) and Var(H) represent the variance of coal flow and humidity, respectively;

[0118] Cov(Q,H) represents the covariance relationship between coal flow and humidity.

[0119] The second term of the objective function α·tr(W TC(t)W) controls the structure of the matrix W to adapt to the dynamic changes of the real-time working conditions during the decomposition process. α is a regularization parameter used to adjust the degree of influence of working condition constraints on the optimization objective.

[0120] The optimization problem is solved using the alternating least squares (ALS) algorithm, updating W, U, and V in sequence until convergence. The update rules are as follows:

[0121] When U and V are fixed, update W:

[0122]

[0123] When W and V are fixed, update U:

[0124]

[0125] When W and U are fixed, update V:

[0126]

[0127] in:

[0128] Represents the matrix expansion of the tensor in the i-th dimension;

[0129] ⊙ represents the Khatri-Rao product, which ensures that the low-rank structure is maintained during the matrix update process.

[0130] After the update is completed, the dynamic feature optimization unit fuses the decomposed factor matrices W, U, and V, and combines them with real-time working condition information to generate the final feature vector F(t):

[0131] F(t)=Flatten([W,U,V])⊙[Q(t),H(t)]

[0132] in:

[0133] Flatten([W,U,V]) means flattening the tensor factor into a one-dimensional feature vector;

[0134] ⊙ represents the element-by-element multiplication operation, which is used to make dynamic adjustments based on the working condition information.

[0135] Finally, the output features F(t) serve as the input of the anomaly scoring, deep detection and decision control modules.

[0136] The anomaly scoring module of the present invention uses a reinforcement learning approach to adaptively score the features extracted by the dynamic tensor fusion module to optimize the accuracy and robustness of anomaly detection. This module uses a Markov decision process (MDP) model and trains the scoring strategy using a reinforcement learning algorithm to achieve a quantitative assessment of anomaly states. The following describes a specific implementation of this module.

[0137] The state space modeling unit establishes a state representation for anomaly scoring. The input state s(t) is composed of the feature vector F(t) extracted by the dynamic tensor fusion module. Here, s(t) represents the state of the system at the current moment and contains characteristic information of multiple physical quantities. This state space provides the input basis for system anomaly detection.

[0138] The action space modeling unit defines the scoring action a(t), whose value range is:

[0139] a(t)∈[0,1]

[0140] Here, a(t) represents the system's abnormality score, with larger values indicating a higher likelihood of the system being abnormal. The score range ensures interpretability and can be used for subsequent decision-making and control.

[0141] The state transition modeling unit constructs the state transition probability based on historical data. Assume that the state of the system at time t is s(t), and after executing action a(t), the system state will transition to s(t+1). The state transition probability is modeled based on empirical data:

[0142] P(s(t+1)|s(t),a(t))

[0143] Among them, P(s(t+1)|s(t),a(t)) represents the probability distribution of the system state transitioning from s(t) to the new state s(t+1) after being scored a(t).

[0144] The reward function optimization unit constructs the optimization objective of anomaly scoring and defines the reward function R(t) to measure the quality of the scoring strategy. The reward function expression is as follows:

[0145]

[0146] in:

[0147] TPR(t) - stands for true positive rate, which is the rate of correctly identified anomalies;

[0148] FPR(t) represents the false positive rate, that is, the rate of misjudging normal states as abnormal;

[0149] Q(t) represents the coal flow, Q max is the maximum value of coal flow;

[0150] β(t) is a weight parameter that controls the balance between false positives and false negatives. Its value is as follows:

[0151]

[0152] The reward function optimizes the true positive rate of anomaly scoring while reducing the false alarm rate, and is dynamically adjusted based on working condition information.

[0153] The strategy optimization unit uses the "Advantage Actor-Critic (A3C)" reinforcement learning algorithm to train the scoring strategy. The A3C algorithm optimizes the strategy π θ (a(t)|s(t)), optimized by gradient update. The policy gradient calculation formula is as follows:

[0154]

[0155] in:

[0156] π θ (a(t)|s(t)) represents the scoring strategy based on parameter θ;

[0157] A(s(t),a(t)) represents the advantage function, which is calculated as follows:

[0158] A(s,a)=Q(s,a)-V(s)

[0159] in:

[0160] Q(s,a) represents the state-action value function;

[0161] V(s) represents the state value function.

[0162] The strategy optimization process adopts an asynchronous update method to ensure that the model can learn efficiently under real-time requirements.

[0163] After training, the scoring output unit calculates the final anomaly score based on the current state s(t):

[0164] a * (t) = argmax a π θ (a|s(t))

[0165] Among them, a * (t) represents the optimal anomaly scoring result, which is used for subsequent decision control.

[0166] Through the implementation of this abnormality scoring module, the system can perform real-time scoring of current working conditions based on historical data and reinforcement learning methods, and provide optimization support for decision-making and control.

[0167] The deep detection module of this invention uses a lightweight spatiotemporal convolutional network (Light-STCN) to deeply analyze the feature data output by the anomaly scoring module to identify abnormal patterns in the coal handling system. This module combines spatial convolution with causal temporal convolution to extract deep features from spatiotemporal data through efficient computation. The following describes a specific implementation of this module.

[0168] The input feature construction unit receives the feature vector F(t) output by the anomaly scoring module as input data and organizes it into a time series matrix X for subsequent convolution calculations. The matrix is defined as follows:

[0169]

[0170] in:

[0171] T represents the length of the time window;

[0172] This matrix stores the feature information of multiple time steps and provides input for subsequent spatiotemporal convolution calculations.

[0173] The spatiotemporal convolution unit is used to perform spatiotemporal separation convolution processing on the fused feature vector, including spatial convolution layer and temporal convolution layer to extract inter-modal correlation features and temporal dependency features respectively.

[0174] The spatial convolution layer is used to model the correlation between different sensor modalities and uses a one-dimensional convolution operation for feature extraction. The spatial convolution calculation formula is as follows:

[0175] h s =W s *X

[0176] in:

[0177] h s is the feature matrix after spatial convolution;

[0178] is the spatial convolution kernel weight, responsible for extracting potential features between sensor modalities;

[0179] R is the number of convolution channels, which represents the extracted potential feature dimension;

[0180] * represents a one-dimensional convolution operation.

[0181] This spatial convolution operation can effectively model the spatial dependencies between different sensor signals and provide feature enhancement for subsequent temporal dependency modeling.

[0182] The temporal convolution layer is used to model the temporal dependency of features and uses causal temporal convolution to extract time series features. The temporal convolution calculation formula is as follows:

[0183] h t=W t *h s

[0184] in:

[0185] h t is the feature matrix after time convolution;

[0186] is the temporal convolution kernel weight, k is the time window length, and controls the receptive field range;

[0187] Causal convolution ensures that the output depends only on the current and past time steps, avoiding future information leakage.

[0188] The calculation of the temporal convolution layer can capture the short-term dynamic changes of features and improve the perception of anomaly detection on temporal features.

[0189] The nonlinear transformation unit performs ReLU activation on the convolution output to enhance the network's nonlinear expression capabilities. The ReLU activation operation introduces nonlinear transformations, enhancing the separability of abnormal patterns and reducing the gradient vanishing problem.

[0190] The feature normalization unit performs batch normalization (BatchNorm) on the convolution output to stabilize the training process and accelerate convergence. This normalization operation can reduce internal covariate shift and improve the generalization ability of deep networks.

[0191] The fully connected prediction unit uses a fully connected layer to calculate the anomaly detection probability and uses Sigmoid normalization to normalize the output value. The anomaly detection probability is calculated as follows:

[0192] y(t)=σ(W o ·h t )

[0193] in:

[0194] y(t) is the anomaly detection result, ranging from [0,1];

[0195] is the output layer weight;

[0196] σ(·) is the Sigmoid function.

[0197] This output represents the abnormal probability of the system at the current time step and is used for the abnormality handling strategy of the subsequent decision control module.

[0198] The model quantization unit is used to optimize computing efficiency and reduce computing resource consumption by quantizing the neural network weights. The quantization calculation formula is as follows:

[0199]

[0200] in:

[0201] W quant is the quantized weight;

[0202] W float is the original floating point weight;

[0203] μ and σ are the mean and standard deviation of the weights, respectively.

[0204] Model quantization reduces storage and computation costs, allowing the network to run efficiently on resource-constrained industrial equipment.

[0205] Through the implementation of this deep detection module, the system can efficiently extract spatiotemporal features and provide accurate anomaly detection results to support intelligent monitoring and safety control of the coal transportation system.

[0206] The decision-making control module of the present invention dynamically adjusts the operating status of the coal conveying system based on the anomaly probability output by the depth detection module and the anomaly score output by the anomaly scoring module to ensure system stability and safety. This module establishes an intelligent coal conveying control decision-making process by setting thresholds, triggering warnings, adjusting coal conveying rates, and implementing emergency shutdown strategies. The following describes a specific implementation of this module.

[0207] The input data acquisition unit receives the anomaly probability y(t) from the deep detection module and the scoring result a(t) from the anomaly scoring module, and establishes a control decision input set based on historical data. The input variables are defined as follows:

[0208] D(t)={y(t),a(t),Q(t),H(t)}

[0209] in:

[0210] y(t) is the probability of system abnormality at the current moment;

[0211] a(t) is the reinforcement learning anomaly scoring result;

[0212] Q(t) is the coal flow data;

[0213] H(t) is the humidity data.

[0214] This input data set provides all the information required for decision-making control and provides the basis for the execution of subsequent control strategies.

[0215] The warning trigger unit sets the warning threshold of the abnormal probability and decides whether to issue a warning signal based on the current status. The warning trigger rules are as follows:

[0216] 0.7 <y(t)≤0.9

[0217] When the abnormal probability y(t) is between 0.7 and 0.9, the system triggers an early warning signal and adjusts the coal transportation rate to reduce operational risks.

[0218] The coal transport rate adjustment unit dynamically adjusts the coal transport rate v(t) based on the abnormality score and warning signal. The coal transport rate adjustment rules are as follows:

[0219] v′(t)=v(t)·(1-γ)

[0220] in:

[0221] v′(t) is the adjusted coal transport rate;

[0222] v(t) is the original coal transport rate;

[0223] γ is the adjustment coefficient, and its value is calculated based on the anomaly score a(t) as follows:

[0224] γ=λa(t)

[0225] in:

[0226] λ is the adjustment proportional coefficient, which ensures that the adjustment range of the coal conveying rate is within a safe range.

[0227] The coal conveying rate adjustment unit reduces the load on the system at the initial stage of abnormality by reducing the coal conveying rate, thereby reducing the risk of further deterioration.

[0228] The emergency shutdown unit is used to handle serious abnormal situations. When the system abnormality score and abnormality probability exceed the safety threshold, the shutdown command is executed immediately. The emergency shutdown rules are as follows:

[0229] y(t)>0.9

[0230] and

[0231] a(t)>0.8

[0232] When these conditions are met, a shutdown signal is triggered for the coal handling system. The system control logic transmits the shutdown signal to the PLC control unit, which executes the equipment shutdown command to ensure the safety of the coal handling system.

[0233] The online model update unit optimizes the anomaly detection model online based on the latest operating data. The trigger condition for model update is to perform a parameter update after accumulating N data, calculate the gradient and adjust the model weight. The update rules are as follows:

[0234]

[0235] in:

[0236] W new is the updated model weight;

[0237] W old is the current model weight;

[0238] η is the learning rate;

[0239] Calculate the gradient value for the latest data.

[0240] This updating process can ensure that the model can adapt to different operating conditions in the long term and improve the accuracy of anomaly detection.

[0241] Through the implementation of this decision-making control module, the system can execute corresponding control strategies under different abnormal situations, ensure the safe and stable operation of the coal transportation process, and optimize the abnormal handling response capability.

[0242] In summary, the intelligent coal transportation system monitoring and online anomaly detection system of the present invention integrates six core modules: data processing, feature extraction, dynamic tensor fusion, anomaly scoring, deep detection, and decision control. This system achieves real-time monitoring and intelligent anomaly detection for the coal transportation system. Through multimodal data fusion, reinforcement learning-based anomaly scoring optimization, lightweight deep learning inference, and adaptive control strategies, the system improves the stability and safety of the coal transportation system. The following is a description of the overall system workflow.

[0243] First, the data processing module collects multimodal data, including temperature, vibration, infrared images, coal flow, and humidity, from multiple sensors within the coal handling system and performs signal preprocessing. The system uses standardized methods to calculate the mean and variance of the temperature data to remove environmental interference. It also performs a fast Fourier transform on the vibration data to extract spectral energy features. It also calculates the grayscale histogram of the infrared image and uses principal component analysis to reduce dimensionality and extract key visual information. All processed data forms a standardized feature vector, which serves as input for subsequent data analysis.

[0244] The feature extraction module then calculates features of the standardized signal output by the data processing module to form a multidimensional feature vector. This module extracts statistical characteristics of temperature and spectral energy of vibration, and performs dimensionality reduction on infrared image features to reduce redundant data while maintaining information integrity. The feature vector serves as the core input data for the system, providing input for subsequent tensor construction and deep learning model calculations.

[0245] The dynamic tensor fusion module then constructs a three-dimensional tensor based on the multimodal data and performs feature decomposition using non-negative tensor decomposition. The system applies low-rank approximations to the time, modal, and feature dimensions to reduce computational complexity and optimizes the tensor decomposition results using real-time operating constraints. This module further incorporates coal flow and moisture data to dynamically optimize the extracted feature information, ensuring the system's adaptability to diverse operating conditions. The fused feature vector serves as the final input for the anomaly scoring and deep detection modules.

[0246] The anomaly scoring module uses reinforcement learning to score the feature data output by the dynamic tensor fusion module. This module employs a Markov decision process model and learns state transition probabilities based on historical data. The system constructs a reward function based on the true positive rate and false positive rate to optimize the anomaly scoring strategy. The reinforcement learning algorithm calculates the current anomaly score based on the scoring strategy and adjusts the scoring strategy to adapt to different operating conditions. The anomaly score serves as the system's assessment of the degree of anomaly and is used by the decision-making control module.

[0247] The deep detection module uses a lightweight spatiotemporal convolutional network to perform in-depth analysis of the feature data output by the anomaly scoring module. This module first extracts correlation features between different sensor modalities through spatial convolution and time-dependent features through temporal convolution. The system further enhances the expressiveness of these features through nonlinear activation and normalization, and calculates anomaly probabilities using fully connected layers. Finally, this module outputs the anomaly detection probability for the current coal handling system and provides it to the decision-making and control module to support subsequent control decisions.

[0248] The decision-making control module combines the outputs of the anomaly scoring module and the in-depth detection module to implement intelligent control strategies. The system first determines whether the anomaly probability exceeds a set threshold. If it falls within the warning range, the coal delivery rate is adjusted to mitigate risk. If the anomaly score exceeds the emergency shutdown threshold, the system immediately triggers a shutdown command and sends an emergency shutdown signal to the on-site control system. Furthermore, the system features an online model update mechanism that automatically optimizes the anomaly detection model whenever a certain amount of new data is collected, improving detection accuracy and adaptability.

[0249] Overall, the system's workflow encompasses six phases: data acquisition, feature extraction, tensor fusion, scoring calculation, deep detection, and intelligent control. This ensures accurate anomaly detection and real-time decision-making and control for the coal handling system. This system enhances monitoring capabilities during the coal handling process, enabling rapid identification and resolution of anomalies, and enabling intelligent management and optimized control of the operating status of coal handling equipment.

[0250] Please see the attached Figure 2 The present invention also provides a method for intelligent monitoring and online anomaly detection of a coal transportation system, comprising the following steps:

[0251] S1. The system automatically collects multimodal data such as temperature, vibration, and infrared images during coal transportation, and performs standardization, noise reduction, and feature extraction on the data to ensure data quality;

[0252] S2. Extract key features from the preprocessed data, including temperature fluctuations, vibration spectrum energy, image grayscale distribution, etc., and convert them into feature vectors that can be used for model analysis;

[0253] S3. Fuse data from different sensors into a unified tensor representation and optimize it in combination with real-time operating parameters to improve the system's adaptability to complex environments;

[0254] S4. Use reinforcement learning algorithms to calculate anomaly scores and dynamically adjust scoring strategies so that the system can optimize anomaly judgment accuracy based on historical data and current status;

[0255] S5. Analyze the fused feature data through a lightweight spatiotemporal convolutional network, identify abnormal patterns, and calculate abnormality probabilities to determine the operating status of the equipment.

[0256] S6. Execute the corresponding control strategy based on the detection results. When the abnormality is relatively low, reduce the coal conveying rate and issue an early warning. When a serious abnormality occurs, immediately trigger an emergency shutdown to ensure the safe operation of the system.

[0257] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Intelligent monitoring and online anomaly detection system for coal transportation system, characterized by: include: Data processing module, used to collect real-time data from the coal transportation system, perform pre-processing, and output standardized data format; A feature extraction module, configured to extract feature values from the preprocessed data and generate feature vectors; Dynamic tensor fusion module, used to fuse multimodal data and optimize feature representation; An anomaly scoring module, configured to perform an anomaly scoring on the fusion feature and output an anomaly scoring value; A deep detection module, configured to perform anomaly detection on the coal transportation system based on the anomaly score value and the feature vector; A decision control module is used to perform system control operations based on the abnormality score value and the detection result.

2. The intelligent monitoring and online anomaly detection system for coal transportation system according to claim 1 is characterized in that: The data processing module includes: A sensor collection unit for collecting real-time data from the coal conveying system, including temperature, vibration, infrared images, coal flow and humidity; The data preprocessing unit is used to preprocess the collected data.

3. The intelligent monitoring and online anomaly detection system for coal transportation system according to claim 1 is characterized in that: The feature extraction module includes: A temperature feature extraction unit is used to extract the mean and variance from the temperature data as temperature features; A vibration feature extraction unit, configured to extract spectrum energy features from the vibration data as vibration features; An image feature extraction unit, used for extracting a grayscale histogram from the infrared image as an image feature; The coal flow feature extraction unit is used to extract the time mean and volatility from the coal flow signal as coal flow features.

4. The intelligent monitoring and online anomaly detection system for coal transportation system according to claim 1 is characterized in that: The dynamic tensor fusion module includes: A tensor construction unit, used to construct the preprocessed multimodal data into a three-dimensional tensor; The constrained tensor decomposition unit is used to decompose the three-dimensional tensor by non-negative tensor decomposition, and the objective function is: in, is a three-dimensional tensor, C(t) is the operating condition covariance matrix generated by coal flow and ambient humidity, W, U, V are decomposition factor matrices, and α is the dynamically adjusted regularization coefficient; The dynamic feature optimization unit is used to combine the decomposition factor matrix with the real-time operating condition parameters to generate a fusion feature vector.

5. The intelligent monitoring and online abnormality detection system for coal transportation system according to claim 1 is characterized in that: The anomaly scoring module includes: The state space modeling unit is used to define the fused feature vector F(t) output by the dynamic tensor fusion module as the state s(t), where d is the dimension of the feature vector; The action space definition unit is used to generate anomaly scoring results a(t), where a(t)∈[0,1], 0 indicates normal state and 1 indicates severe anomaly; A dynamic reward function unit is used to adjust the reward value according to the false alarm rate and the missed alarm rate. The reward function is: Among them, Q(t) is the current coal flow, Q max is the maximum design coal flow rate, TPR(t) represents the true positive rate, FPR(t) represents the false positive rate, and β(t) is the dynamic adjustment coefficient; The policy optimization unit uses the asynchronous advantage actor-critic algorithm to update the policy network parameters θ, and the optimization objective is: Among them, π θ (a(t)|s(t)) represents the strategic probability of taking action a(t) in state s(t), A(s(t),a(t)) is the advantage function, is the expected value operator.

6. The intelligent monitoring and online anomaly detection system for coal transportation system according to claim 1 is characterized in that: The depth detection module includes: The spatiotemporal convolution unit is used to perform spatiotemporal separation convolution processing on the fused feature vector, including: Spatial convolution layer, the convolution kernel dimension is Extract spatial correlation features between sensor modalities, where M is the number of sensor modalities and R is the potential feature dimension; Temporal convolution layer, the convolution kernel dimension is Extract time-dependent features, where k is the length of the time window; A quantization calculation unit, configured to perform integer quantization on the weights of the spatiotemporal convolutional layer; The prediction output unit is used to generate the abnormal probability value y(t)∈[0,1], and the calculation formula is: y(t)=σ(W o ·h t ) Among them, h t is the output feature of the temporal convolution layer, is the output layer weight, and σ(·) is the Sigmoid function.

7. The intelligent monitoring and online anomaly detection system for coal transportation system according to claim 1 is characterized in that: The decision control module includes: The early warning trigger unit is used to perform the following operations based on the abnormal probability value y(t)∈[0,1] output by the deep detection module and the scoring result a(t)∈[0,1] of the abnormal scoring module: When the abnormal probability value y(t) is within a predefined first threshold range, reducing the operating rate of the coal transportation system and sending an early warning signal; an emergency shutdown unit, configured to trigger an emergency shutdown instruction when the abnormal probability value y(t) exceeds a predefined second threshold and the abnormal score value a(t) exceeds a predefined third threshold; The online model update unit is used to dynamically update the model parameters based on the newly collected data. Specifically, the tensor decomposition weights are adjusted using the following formula: Where η is the learning rate, is the gradient of the loss function with respect to the weight matrix W, A tensor constructed for newly added data; the update process is performed asynchronously with the PLC control cycle.

8. A method for intelligent monitoring and online anomaly detection of a coal transportation system, based on the intelligent monitoring and online anomaly detection system of a coal transportation system according to any one of claims 1 to 7, characterized in that: The following steps are involved: Collecting real-time data of the coal transportation system and preprocessing the data; Extract features from the preprocessed data to generate feature vectors; Fuse multimodal data into tensors and optimize them; Use reinforcement learning to dynamically adjust anomaly scores; Use deep learning models to perform anomaly detection on the fused features; Execute early warning or emergency shutdown control operations based on the detection results.

Images